[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vr / w / wg] [i / ic] [r9k / s4s / vip / qa] [cm / hm / lgbt / y] [3 / aco / adv / an / asp / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / qst / sci / soc / sp / tg / toy / trv / tv / vp / wsg / wsr / x] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/mlp/ - Pony


Thread archived.
You cannot reply anymore.



File: PP P.png (1.1 MB, 2119x1500)
1.1 MB
1.1 MB PNG
>What is this?
https://clyp.it/231umvkx [A whaaAAt!?]
This project is the first part of the "Pony Preservation Project" dealing with the voice.
It's dedicated to save our beloved pony's voices by creating a neural network based Text To Speech for our favorite ponies.
Videos such as https://youtu.be/GuJKTodX1FA or https://youtu.be/DWK_iYBl8cA have proven that we now have the technology to generate convincing voices using machine learning algorithms "trained" on nothing but clean audio clips.
With roughly 10 seasons (8 soon to be 9 seasons and 5 movies) worth of voice lines available, we have more than enough material to apply this tech for our deviant needs.

Any anon is free to join, and many are already contributing. Just read the guide to learn how you can help bring on the wAIfu revolution. Whatever your technical level, you can help.
Document: https://anonlink.com/1uFcm
Spreadsheet: https://anonlink.com/1uFcn

>Active Tasks
-Create a dataset for speech synthesis (https://youtu.be/KmpXyBbOObM)
-Test some AI program with the current (unfinished) dataset
-Research AI (read papers and find open source projects)
-Find a good way to host a full archive of the project resources (ideally p2p)
-Track down any left behind audio
-Further audio cleaning (main focus right now)
-Investigate SAMPA/phonetic tagging

>Latest Developments
-https://clyp.it/xp4q1bru [Yay!]
-A group of anons (well mostly one Anon in fact) have completed full clip sets for over 8 seasons worth of audio
-Multiple anons have created converters to help import subtitles and dialogues into Audacity
-New noise reduction technique using multiple dubs

>Voice samples (So far)
-https://clyp.it/px11j2wn
-https://u.smutty.horse/lqctkqxclef.7z

>Clipper Anon's Master File:
https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

>Cool, where is the discord/forum/whatever unifying place for this project!?
You're looking at it.

Last Thread: >>34080783
>>
File: Very Big Anchor.jpg (103 KB, 601x850)
103 KB
103 KB JPG
Anchor
>>
Thread's a bit slow this time. Gotta keep it afloat somehow.

btw does anyone have any new TwAIlight clips? I'd like to hear how far we've come so far.
>>
>>34189929
No new ones quite yet, I believe efforts have been thrown into audio clean up.
>>
File: will you marry me.jpg (382 KB, 722x600)
382 KB
382 KB JPG
>"Anon, I have a surprise for you."
>"We're gonna be parents! How do two beautiful satyr children sound?"
>"Oh it'll be just so wonderful"

One day damnit, one day.
>>
File: dimorphism.png (587 KB, 853x1024)
587 KB
587 KB PNG
>>34190849
>satyr
we must agree to disagree there, and this is not the thread for that discussion
>filename
what i wouldn't give to have moonhoers ask me that
>>
File: 1489129846310.jpg (135 KB, 1153x525)
135 KB
135 KB JPG
>>34190855
it's actually nicer to have conversations about satyr's outside of /satyr/ since like only 3 characters exist there and they're all bats.
>>
File: 1141878.jpg (74 KB, 821x717)
74 KB
74 KB JPG
>>34190849
>>34190855
>tfw waifu has so few spoken lines that tweaking a perfected voice model of a similar voice still won't work
>tfw the only way I'll ever hear this is if equestria gets the portal working
It never stops hurting
>>
>>34190876
You could try and use accurate fan voices?
>>
>>34189332
I’ve re-clipped s1e21 - 26, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

With that, all of season one has now been run through the cleaning process. I’m currently running the season two dubs through iZotope, and will upload them when done. Will start re-clipping season two tomorrow.

Also found this - https://youtu.be/3Xqar7OgiIA

A fun combination of two AIs: inspirebot, which generates “inspirational” quotes, and notjordanpeterson, the Jordan Peterson speech synthesiser from the previous thread.
>>
Would it be possible to take the same effort that that made so many strides here and put it into preserving the show's animated assets? All of the animated projects in the fandom always died because of disagreements on the final product but what if we were working to create an open pool of resources that all future projects could use. Even vectoring out every location and background used in the show would go a long way in kickstarting projects. The best part is that there's already a shitton of stuff out there in the mlp vector club and DHX leaks that could serve as starting point while we nail down a structure to organize everything in.
>>
>>34192853
Yes, but create a separate project/thread for it.
>>
>>34192853
Would be a good idea for sure, as the other anon said make a separate thread for it. There’s already a lot of flash puppets available, but those are only really for the first 4 seasons or so
>>
>>34189332
>>34191127
Season 2 cleaned audio, English - https://anonfile.com/Z3J1Qa3dnd/S2_zip
Processed stereo tracks, English and German - https://anonfile.com/o7O1Ve34nb/S2_zip
>>
>>34190876
if it helps anon . I was planning on being the middle man for this thread and there voice actors
For example i have meet Tara Strong when my cousin married her cousin . You sould be surprised how nice they are how big of a "trolly they are.
>>
>>34191127
Great!
Also, interesting, funny and scray video you got here.
It also proves that if you add noise (like music or so) on top of generated voice, the result is even more convincing.

>>34192853
Great idea.
Maybe a PPPA : Pony Preservation Project : Assets?
But it's a huge work, probably harder than the voice.
Because, we do what we want here. But for the assets, you will need to agree on the formats, color, programs and so much things.
But please do an other thread an put a link here.
>>
>>34190876
>tfw waifu has so few spoken lines that tweaking a perfected voice model of a similar voice still won't work

This is just short sighted, as there will likely be models that can generate voice models as opposed to train them in the way we do now. In the same way that we can generate faces.

I haven't seen this discussed here before but I suspect the actual accuracy of the voice model matters a lot less than how appealing it is, in the same way that green text stories aren't written to be show accurate but rather to appeal to the reader with enough accuracy to suspend disbelief.

>tl;dr I suspect over multiple generations ponies will be made to sound more precious than any human could portray them so as long as a model is accurate enough to suspend disbelief it doesn't matter if there's not enough data to make a 100% accurate model.
>>
File: bump thread.gif (482 KB, 300x252)
482 KB
482 KB GIF
>>
So this happened: https://twitter.com/woot_master/status/1165364320575864832
>>
>>34195411
literally who?
>>
>>34195411
but didn't we already have this anyway?
>>
>>34195417
Wootmaster? Guy who drew Tracy Cage, plays the drums at concerts, makes edgy rap, and is /ourguy/.

>>34195419
Right, for a second I thought he got it from this thread. Apparently not.
I don't know if the quality is any better than what we have.
>>
>>34195411
>>34195417
>>34195419
>literally who?
Don't know, don't care. Doesn't matter who he is, but he does seem to care about preserving ponies, may be worth directing him here.

>but didn't we already have this anyway?
Yes. I've just downloaded his s1e1 and it's just the raw audio with the music removed, all the sound effects are still in there. It's therefore no better than what we already have. In fact, the file is about 50 Mb, and the stuff I'm working with is about 100Mb, so if anything I'd expect it to be slightly worse.
>>
>>34195437
>I don't know if the quality is any better than what we have.
It's just center channels from itunes, the quality is shit.
>>
>>34195437
>>34195440
I've heard of him but never really seen any of his stuff. Might be worth someone letting him know we've actually clipped the audio into specific lines of dialogue. I feel like that would be useful if he's using it in music.
Although at the same time, I'm cautious to let too many people know about this project in case someone decides they don't like the idea of it and goes running to a VA or tries to get it shut down somehow. You can never be sure with Twitter.
>>
>>34195452
>You can never be sure with Twitter.
I know, but if >>34195437 is right, I think it would be a reasonably safe bet. He also doesn't seem to have that many followers, so not sure how much of a fuss he could realistically cause if he tried.
>>
I haven't seen anything of woot in fucking years. Im also not too keen on getting too many people outside of here involved at this stage, but iirc he's pretty chil, dming hi. About this might be good.
>>
well... is suppose that's it. dutch version is up for s9 finale now
>>
>>34197558
That doesn't kill the project though, right?
>>
>>34198162
no. it's just a bittersweet time for /mlp/.
>>
>>34197558
No
I refuse it to finish !
Isn't there a way to.. to...
Ho shit, it will be hard to watch this last episode when it will be up.

>>34198162
Nope.
Neither the other I have for me.
And I expect Santapone and other to keep it up at least this year.
But let's face it, it will probably slows down even more.
We will see what will happen when we will reach it.
Next gen may be good, or aweful. Fan may do more content because of C&D free, or may lose interest, I don't know.
As far as I know mlp is really the first cartoon to have such an impact a fandom, so I don't know what will happen in the end.
>>
>>34193752
We can take a similar approach to what we're doing here. Some anons work on extracting assets, other anons work on programs/AI to automate the process and generate new assets using data generated by the first anons.

I'm game to work on that, but not before >>34190876 gets his waifu. Until then, I'll be working on speech generation.
>>
>>34198184
>>34198764
Thank you. Every pillar of stability is good now, especially when it's such a great thing.
>>
>>34189332
>>34191127
I’ve re-clipped s2e1 - 18, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

The goal for tomorrow - finish the rest of season two and run the season three material through iZotope.

I realised today that I uploaded the wrong files for the German containing stereo tracks, it was supposed to be the material prior to running through iZotope, I uploaded the output instead. Correct links below, and apologies if you already downloaded the old stuff.

S1 - https://anonfile.com/Mba5b542n1/S1_zip
S2 - https://anonfile.com/a1V0b440nd/S2_zip

In case anyone is wondering, we want to keep the stuff prior to processing in case a new better method for cleaning is found, keeping the processed stuff won't offer that possibility.

>>34193752
I think you may have missed this reply from myself at the end of the previous thread >>34184490.
>>
>>34190876
and yer waifu is...
>>
>>34201037
You're doing well Clipper Anon. We're getting closer every day.
>>
Fuck, just got the final spoilered, guess that means this is my final day here.
>>
>>34202290
Why? And why is that relevant to this thread?
>>
>>34202339
>Why
What's the use of doing anything pony when it's all over?
>Why
Because I was contributing a lot to the thread and thought I should say goodbye
>>
>>34202381
>Contributing to a thread about pony preservation
>Leaving once the show's over
The whole point of this project is to keep the pony going long after the show's demise.
>>
>>34202439
I want to remain optimistic and hold onto the belief that any anon who was actually "contributing a lot to the thread" would at least have the wherewithal to stick with the thread after the show ended, especially knowing that a) of course the ending was going to get leaked and b) Pony Preservation is about continuing to work on this even after the show ends.
>>
File: image.png (608 KB, 1800x1348)
608 KB
608 KB PNG
>>34202381
>>
>>34202381

Fuck, I forgot my trip. No wonder you guys didn't recognise me
>>
>>34202381
I don’t know if I’d say it’s “all” over just because the show is. You’ll probably feel like this for a bit but once the dust settles maybe you’ll come back. If not, I guess thanks for helping out, always a shame to see people go.
>>
>>34202565
you are that discord infrahuman, aren't you?(confirmation here https://desuarchive.org/mlp/thread/34019408/#34025996)
also known as United Union, aSS, AuroraSagebloom, Cutie Mark Crybaby, red board Blueberry Cuddlecakes (Element of Hugs), Lauren, Lauren Faust (really), Mitch, MusicComposer, Biscuit, Glados, Fluttershy <3, Ponyraper666, err4tic, Tripfag, Secrets&Lies, Newpony, Leothore and isaac on mlp, iloveponies and Transfeminist on co, and Pup in r9k
https://desuarchive.org/_/search/boards/a.aco.an.c.co.d.fit.gif.his.int.k.m.mlp.q.qa.r9k.tg.trash.vr.wsg/tripcode/!IHwejl%2FaNY/page/2/
you have contributed nothing to ANY of these thread, you are just a waste on oxygen, fuck off and stay fucked
also for the enjoyment of the thread denizens here is your first post with the trip, you 24+ year old failure.
https://desuarchive.org/mlp/thread/37922/#39056
>>
>>34202806
>>34202582
cease wasting posts, you lobotomitic mongoloid
>>
File: Rules of nature.png (88 KB, 653x863)
88 KB
88 KB PNG
Safety bump, boards emotional right now and prone to sudden thread flood.
>>
>>34203585
Damn, you are right, there is quite some talk now.
It will be hard to doge spoilers till I watch the last episode...
>>
File: update.png (139 KB, 1123x1179)
139 KB
139 KB PNG
>>34189332
I SAID WHEN IT'S DONE
My Little Pony Animated Shorts:
https://mega.nz/#!7MFnHKQT!QU2MLO3uoUKcNd7eJHUwPHxjdPZ5g3L1sW6IuSfcW_o

I hope the secret to digital black magic has been discovered while I was away, because what I've been hearing has not been filling me with hope.
>>
>>34204843
Whohou someone else than me whom complete the Follow Up sheet! Good to see your tool being used, and kudo for the delivery.

But if I may, you forgot to include >>34201037
so, here is the last one with (hopefuly) everyone's job up to date
>>
File: pony tracy.jpg (207 KB, 1175x1100)
207 KB
207 KB JPG
Hey guys, Wootmaster here. I sincerely don't want to step on you guys toes or anything with regards to pony vocals. I'm just doing this for musicians who might need music-free vocal samples to work with. I haven't taken anything from this project and if I do I will ask permission and give credit.

Now I do need clean vocals-only rips of some EQG stuff and the Movie if possible, because my sources have some of the EQG things in 5.1 but not all of them. I might be jonesing for that but otherwise I'm not trying to cause drama.

Keep up the good work horsemen.
>>
>>34206125
Was there any drama? I don't think there was. You're a cool guy. You keep on doing your own thing.
>>
>>34206125
Other people told me to hate you so fuck you.
/)
>>
>>34206125
Check the end of resources in the doc. I uploaded basically all the 5.1 that exists for fim/eqg there. Please feel free to mirror the content, the more people with copies, the better.
>>
bump, board moving fast. even if the show is ending we can keep the ride going forever with the power of AI
>>
>>34208037
>The ride has only just begun.
>>
>>34206125
Thx.

I can't speak for everyone obviously, only me.
But as far as I'm concerned, the two main goals of this project are :
- Avoid Asdrone radars as much as possible (so no ad or social media until finished and saved and backed up multiple time),
- Share the pony pony autism power with anyone.

So, if you need fast results, do as >>34206774
said, and extract the voices yourself like you have already done for mlp.
If you can bare to wait like three months to one year, wait for us to finish and publish the dataset and programs.
(both are not incompatible I think).

Have a good time with your pony project, whatever the form it takes.
>>
Imagine the insanity we would get if anons could make their own episodes.
>>
>>34208464
Not going to lie fixing this pile of garbage is high on my list.
>>
>>34208464
I think that if you are on this thread, it's more or less your long term goal, isn't it?
>>
File: hurr.png (2.21 MB, 849x1641)
2.21 MB
2.21 MB PNG
>>34206125.
The only thing you can do is fuck right off of this board, you treacherous gutter rat.
You fucking shat on people who warned this shit finale would happen and shilled out to Hasbro.
>>
File: 546378954.png (161 KB, 245x348)
161 KB
161 KB PNG
>>34190876
I know that feel. Never give up, Anon. I'm sure in 5-10 years we'll have the technology.
>>
File: 1540988188242.jpg (304 KB, 928x754)
304 KB
304 KB JPG
>>34208037
I hope AI will one day be able to capture the magic that came with the show and it's fan content.

I don't even need it to bring the characters from the show to life, An AI that could simulate a pony with all the care and wholesome values that comes with that would be a beutiful thing.

A world in which a teenager can stumble upon an app that creates a pony which will try and be your friend and companion is a world worth fighting for.
>>
>>34209308
>A world in which a teenager can stumble upon an app that creates a pony which will try and be your friend and companion is a world worth fighting for.
A-fucking-men.
>>
>>34208037
>>34208065
like I said before if this works I'm going to recreate Exchange, teach myself how to storyboard and animate, maybe even teach myself to compose if i don't bring in my music buddy, and release it to the internets as my gift to the world
>>
>>34209394
And we would love you for that, Anon.
>>
>>34209429
appreciated, but i wouldn't do it for your love, i'd do it out of my love for pone. it's why i love this place so much: it encourages content to be more important than the creator, rather than the other way around. it's how things were long ago and i feel society/the internets have lost sight of that
>>
>>34189332
>>34201037
I’ve re-clipped s2e18 - 26, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

Cleaned season three tracks:
English - https://anonfile.com/4cubf740n1/S3_zip
Raw stereos - https://anonfile.com/E4Aff144n1/S3_zip

I did a little extra processing on s2e24 with Audacity’s noise reduction tool to reduce the constant train sounds, which worked pretty well. I’ve uploaded the audio for that as well in case anyone wants it.

https://anonfile.com/g9Caf244n2/mlp.s02e24_dialog_extra_flac

I’ll start working on re-clipping season three tomorrow.
>>
Bamp
>>
>>34209394
For what it’s worth I could probably compose something vaguely similar to the music in the show
>>
>>34189328
You ARE using the 5-channel audio versions, right?
>>
>>34211743
Yes.
>>
>>34189328
Why does the spreadsheet suggest conda/jupyter when you have a pip requirements.txt in the GitHub repo?

pip install -r requirements.txt

Jupyter isn't even a package manager.
>>
>>34211965
Never mind, you can use conda too. Don't use it often myself.

conda install --yes --file requirements.txt

"Needs a newline at the end or the final requirement isn't installed." - stackoverflow
>>
>>34190876
There was a recent visual "talking head" GAN thingy that used a meta-learning stage, so that you would only need 5 or so frames of the target face in the second stage to generate a talking head of your choice.

https://towardsdatascience.com/meta-learning-of-adversarial-generative-models-fb5e77ade719

I expect such a meta-learning approach could also be applied to audio TTS somehow, so it might be possible to generate your "small dataset" waifu someday.

Who is your waifu, anyway?
>>
>>34189328
Is it me or did twiggles stop updating the document and spreadsheet?
>>
>>34212103
https://www.youtube.com/watch?v=wLDGvPRrzCA
Middle pone
That clip and one sentence in the movie is all I've got to work with.
>>
>>34212112
It's been months. For the moment it's alright. At some point if we can't establish contact we might want to copy it over to another drive that someone can keep updated.
>>
File: qt_yo.jpg (232 KB, 732x1024)
232 KB
232 KB JPG
>>34212310
Good taste in pony.
Fuck that bit / reference was gold and still makes me laugh.
>>
>>34212310
Kek, I remember this clip.
"Tank for the memory" if I remember well?
>>
>page 9

no
>>
File: PVPP Rarity Profile.png (1.33 MB, 1500x2749)
1.33 MB
1.33 MB PNG
Anon that made these voice profiles here

I'm currently watching the show start to finish, but after that, I'm gonna continue making these references for other prominent characters with the voices fresh in my mind. Celestia, Luna, Starlight, Trixie, Sunset, and Discord are what I'm thinking.
>>
>>34216515
That'll be be fun.
>>
Hi... new.
Reading through but want to absolutely clarify - complete AI voice recreation, in the same manner of NotJordanPeterson; LyreBird?
Text input and handful of outputs, plus regeneration for different inferences and inflections of voice?
Because I'm looking at doing a VR pony game.
>>
>>34217375
anon let me give you some tips if you want to do a vr game
1 dont advertise that much hasbro is know for C&D fanprodcets that get to whell know
2 if you want to advertise DO NOT post your name , DO NOT say anywere close to where you live , Do NOT release it before you think its ready .
please anon we are depending on you .
>>
File: 123636.png (1.44 MB, 680x434)
1.44 MB
1.44 MB PNG
>>34217411
>Be retard
>Release pony game on servers subject to US copyright
>>
>>34217375
>VR pony game
Nice.
Fellow VR-Dev-fag here, what is yours going to be ?
>>
>>34217726
Aiming for VR My Little Investigations, but with my own shit.
Look, examine, deduce, OBJECTION
>>
>>34216515
i'd love to get one of these on Luna
>>
>>34212310
>"waifuing" a literal nobody meme character
Entirely your fault, that's not how waifus work. You just like her color scheme and the occasional r34 fan art. She has no personality or character trait to grow fond of.
>>
>>34217375
Yep, that’s the idea. You can hear some of the samples we’ve generated so far in the OP, although those were a little while ago and using not very much of the data
>>
Update: I've aligned (phoneme-level) about 16.6k of the 17.6k utterances between seasons 1 and 3. Notably, I'm having issues with Luna. I have ideas for getting the rest of S1-S3, but I'm putting that on pause since the alignments I have are good enough for now. I got all of Twilight's, at least.

I'm having download issues with Mega, which is why I haven't done seasons 4+. If I get bored, I'll try downloading the rest of the seasons from an AWS instance. If some nice anon wants to download them from Clipper's mega folder and seed a torrent or put them onto another file hosting provider, that would be much appreciated.

Up next is testing my phoneme-level tools with all of the S1-S3 data. Once that's done, I'll dump what I have to github and dockerhub, then work on attaching prosody information to the utterances.
>>
>>34218272
https://anonfile.com/f016m846ne/alignments_zip
>>
>>34218143
>implying its a choice for you to make, like you're at a restaurant
Your waifu finds you, anon.
>>
>>34218272
Awesome!
When the tool will be ready, please do say so that way, we could help.
>>
>>34218272
>I'm having download issues with Mega, which is why I haven't done seasons 4+. If I get bored, I'll try downloading the rest of the seasons from an AWS instance. If some nice anon wants to download them from Clipper's mega folder and seed a torrent or put them onto another file hosting provider, that would be much appreciated.
Mega can be annoying, yeah. It's throttling me a bit right now but I can mirror S4+ if that helps. Check here http://pubshare.ponemusic.net/Clipper%20Anon%27s%20Master%20File/ in a few hours, I'll started a few downloads I'll copy them when they're done.
>>
>>34218272
>>34220757
Just dropping in with a quick note, don’t bother downloading anything other than season one and two just yet, as most of it will be changing soon as I go through the cleaner audio.
>>
>>34220772
Oh that's okay I'll just download them again, I have a lot of spare bandwidth. Just ping the thread when you're done with it.
>>
File: 1566934782457.jpg (233 KB, 1760x1760)
233 KB
233 KB JPG
Currently working on tacatron+waveglow from nvidia notebook.
Almost finished script for preparing dataset for it (will update github repo)
Also training noisy+clean version of deepvoice 22k steps.
Last week i wasn't active, because i hadn't access to pc.

Looking forward to your cleaning process, anons!
Not only because of removing noise, but for double checking audio and quotes. I still think there're some errors in them.


> leaks
God, why. We can't have just nice and cozy end of the show together, can we?
Really was sad for steping into some spoilers.
>>
>>34220808
Yeah, it’s such a shame that the show’s ending had to be revealed like this. Oh well, all we can do is move onwards towards our AI waifu overlords
>>
File: nervousLaughter2.gif (281 KB, 510x502)
281 KB
281 KB GIF
Trying real hard to keep my mouse away from those black bars, but the only thing I hate more than knowing too much is not knowing.
>>
>>34220878
Twilight can never resist black bars
>>
I'm not sure I understood this project right, but, this means that you're getting the original lines of the show and cutting or separating them?, because I wanted to do that, like, get all of the lines of Twilight from every episode
>>
>>34221002
We’ve already done that. The idea of this project is to use AI to basically clone the voices of the ponies for our own use.
Like a very advanced version of text to speech if I were to simplify it
>>
>>34189332
>>34209796
I’ve re-clipped s3e1 - 6, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw
>>
>>34221285
You are almost as fast as Voyager1!
>>
>>34222288
>all that green

We’re getting there bois
>>
this is an interesting project if i wasn't a literal peasant id throw some bux at it
>>
>>34223210
There really isn't a need for money rn.
Not unless you wanted to pay Clipper anon for his huge contribution
>>
>>34223583
if i could right now i totally would but im in a not so great position economically speaking. if this project is still going by the time i find work and get my shit together id glady throw some mons towards whomstever
>>
>>34218272
>>34220772
https://anonfile.com/E7rdo444nb/MLP_Movie_7z
https://anonfile.com/A4u0o740n1/EQG_7z
https://anonfile.com/85M1o94en8/S4_7z
https://anonfile.com/TbMdo649n6/S9_7z
https://anonfile.com/X5c6p44dn6/S5_7z
https://anonfile.com/O8e3p746nd/S6_7z
https://anonfile.com/12i9p042na/S7_7z
https://anonfile.com/1dl3pa4bn8/S8_7z

I went and mirrored the stuff that's there now. So here's that I guess.
>>
Can we get some demos? I don't mean to sit on the sidelines and act all uppity, but I'm genuinely curious to see how far this whole thing has come as of late.
>>
>>34220257
Will do. Once the basics are in for generating speech, I'll see what I can do to make it easier for other anons to help out.
>>34220757
>>34223803
Thanks a ton, both of you. I'm in the middle of processing S6/S7 and downloading the rest from >>34223803.
>>34220772
I'm setting things up to be easy to re-run. The audio from the current set is good for testing things out, even if it's going to be redone.

Also, some errata in the transcripts. >>34220808, maybe you can check for these to see where exactly they occur. When I found these, I just fixed them in my local copies and didn't think to take specific notes.
* Braeburn's dialogue seems to be missing from s1e21
* Diamond Tiara's name is misspelled as "Diamond Tiarra"
* Granny Smith's name is misspelled as "Grany Smith"
* One of the files is missing the corresponding .txt transcript
* One of the .txt files has an extra (or missing?) underscore at the end of the filename
* I think Zecora's name was shortened to "z" for one file
* In s6e19/00_05_45, there's an extra "." before the ".flac"
>>
>>34223998
Check the OP. There's a zip linked with most of the examples so far.
>>
>>34222300
And it's only phase 2.
But we (mostly Clipper and IAnon and Synthbot in fact) are still going after 4 month, and indeed, not stopping!
>>
>>34220772
>>34220808
More in the transcript filenames:
* In S5/s5e23/00_20_11, "Ma Hooffield" is misspelled as "Ma Hooffiled".
* In S8/s8e5/00_15_49, "Granny Smith" is misspelled as "Ranny Smith".
* In S4/s4e10, "Fleetfoot" is shortened to "ff".
* In S4/s4e24/00_05_48, "Shining Armor" is shortened to "sh".
* MLP Movie/00_31_39 is missing the character name ("Capper").
* A lot of the audio files have extra periods before the file extensions, which the corresponding txt files do not.

All of these can be caught and fixed pretty easily with a custom script, assuming you know which character names to accept. I'm currently using this list of character names: https://pastebin.com/S2ebrzSS.
>>
>>34224147
>Braeburn’s dialogue
Already fixed in the cleaned update.

>Diamond Tiarra, Grany Smith, “z” for Zecora
I’m re-running the checking script on all episodes as I go, so any occurrences of those in the name field should get corrected. However, it won’t recognise anything in the transcript field, so let me know if you find anything else like that.

>Missing .txt transcript
I’m regenerating all the text file transcripts as I go, so that should get fixed whenever it comes up.

>Extra underscore
Question marks are disallowed in filenames, so we swap them for underscores, which is done automatically by Audacity when exporting the audio. If you think any of it is wrong, show me the line and I’ll take a look.

>>34225006
>Typos
I’ll fix those when I get to them, and will continue to correct any others I come across as I go. Please do continue to point out any others you find, especially in the cleaned episodes.

>Extra/missing periods
That’s a product of the script I’m using to generate the transcript text files, it seems to strip periods that occur at the very end of the file name. The text within the file always remains unchanged, and so will be correct to the audio file it’s associated with. If I remember correctly, the script can be modified to include all periods in the file name, so if it really is an issue that must be changed, I can probably make that happen.
>>
Bump.
>>
>>34225842
i approve of this bump
>>
>>34189332
>>34221285
I’ve re-clipped s3e7 - 11, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

Tomorrow, I will finish the last two episodes in season three, and then run season four through iZotope.

At the rate I'm going, I should be able to finish season four by the end of this week, and then season five a few days later. At that point, I will be out of material to work with for cleaning, so to anyone who has 5.1 dub audio of anything other than FiM seasons 1 - 5, if you could upload that somewhere for me soon that would be great.
>>
>>34227314
The EQG dubs have been uploaded already. If you haven't already done them, could do those next. You're absolutely on fire man. I've been meaning to get around to processing s5, but have been rather busy. Maybe this weekend I will see about it, but if you get there before me go ahead. Just give a heads up if you do.
>>
>>34227314
A question I do have is are you preserving the original clean audio clips? While it's probably fine to use the izo output for everything, I'd imagine it's likely best to use as unprocessed audio as possible.
>>
>>34227343
I must have missed those, could you show me the link?

>>34227360
I haven't been preserving the original individual clips, as I'm running a bit low on storage space. Anyone who's downloaded stuff from the master file previously should be able to re-upload if needed. Failing that, the label files can be used to re-obtain original clean clips pretty easily.

I'm preserving pretty much everything else while cleaning:
>The original English, with the music removed, from the torrent in Google Doc.
>The original dubs, mostly German, with the music removed.
>The raw stereos of the aligned English and dub, have uploaded previously.
>The iZotope processed stereos of the aligned English and dub.
>The iZotope processed English, have uploaded previously.

If at any point anyone wants any of the above, I should be able to get a link up within a day or so. Here's everything I've uploaded so far:

Raw stereos:
https://anonfile.com/Mba5b542n1/S1_zip
https://anonfile.com/a1V0b440nd/S2_zip
https://anonfile.com/E4Aff144n1/S3_zip

English:
https://anonfile.com/A7h6943bnd/S1_zip
https://anonfile.com/Z3J1Qa3dnd/S2_zip
https://anonfile.com/g9Caf244n2/mlp.s02e24_dialog_extra_flac
https://anonfile.com/4cubf740n1/S3_zip
>>
>>34227314
>>34227403
not that anon but according to this https://desuarchive.org/mlp/thread/34080783/#34180744
these are all the dubs he has, which reminds me, what brave soul is going to upload these to rome?
https://desuarchive.org/mlp/thread/34080783/#34153097 BGE+ToCH(MMN)+RoF+RR+FF
https://desuarchive.org/mlp/thread/34080783/#34156954 FG
https://desuarchive.org/mlp/thread/34080783/#34158474 s6+LoE
https://desuarchive.org/mlp/thread/34080783/#34158637 s7+8
https://desuarchive.org/mlp/thread/34080783/#34163789 s4
https://desuarchive.org/mlp/thread/34080783/#34171994 s2
https://desuarchive.org/mlp/thread/34080783/#34175672 s3
https://desuarchive.org/mlp/thread/34080783/#34177076 s5
https://desuarchive.org/mlp/thread/34080783/#34180718 s1
or should i do it, although i warn that i will be slow as molasses
also if he sees this post, he should dl the first eqg https://www.netflix.com/es-en/title/70276351
>>
>>34222288
add to that list
Rainbow Roadtrip
Sunset's Backstage Pass
Pinkie Pie Presents Her New Show 'Hello Pinkie Pie'! https://www.youtube.com/watch?v=CLa8CAQKH-4
bluray release extras(baking with pinkie, etc)
also clipper, i think that rome has a collection of all languages of the happy birthday song, so there should not be much problem with cleaning that one up
>>
>>34227444
>Lots of files, several Gb each.
I suppose those are all of the dubs in all available languages? I only need one dub for each, so could I request an upload like that for each EQG thing from whoever has it? Would be much more efficient on my end.

>>34227494
>Happy birthday song
Good to know, will probably get round to that sometime later.
>>
>>34227494
wait, that youtube video might not have andrea's voice so nevermind it
>>
>>34227515
I'll see about starting an upload of a single dub for the eqg stuff later today. Any particular languages you'd prefer?
>>
>>34227444
>or should i do it, although i warn that i will be slow as molasses
nevermind that noise, i thought the files where big, but not that big, i'm not made of hard drives, anonthatuploadedtheminthecourseof8day sorry to ask this of you but could you upload them to rome?
and if i am to be an absolute chode about it, could you uploaded them in folder so that anons can download the language they want for whatever episode without having to download 15 parts(at least for seasons 2(69.5GB) and 4(66.5GB), for the curious 36.68GB s1, 30.11GB s3, 27.96GB s5, 10.83GB s6, 8.2GB s7, 8.26GB s8, 1.02GB BGE, 1.42GB RR, 12.75GB FG, 12.6GB LoE, 1.04GB FF, 1.02GB RoF, 1.52GB ToCH, 289.41GB in total+whatever EQG ends up using when its downloaded, nearly 300GB, pure madness, i do not even think that rome has the space for that, nor the want)
>>
>>34227728
I'd prefer that they all match up without a need for re-syncing, but realistically I can't really expect it to go that way every time. Since the Finnish material apparently matched up perfectly with season six, I suppose that would be the best bet to make. If for whatever reason Finnish is unavailable, it doesn't really matter what dub is used so long as it's in 5.1 with the music removed, so I guess German to be consistent with what I have used already. Thanks.
>>
>>34227494
Damn, the animation is so cute!
I want more!
But with the real voice indeed...
This one's too young for me.
>>
>>34227784
It looks like Forgotten Friendship, Rollercoaster of Friendship, and Tales of Canterlot High should be already up with just a couple dubs. Eng 5.1, Eng 2.0, Fin 2.0, and Ger 2.0. Rainbow Rocks is already up with only Eng 5.1 and Spa 5.1. I will work on uploading Legend of the Everfree and Friendship Games with just the Eng 5.1 and Fin 5.1 Dubs. See >>34227444 (or resources in the doc) for existing links.
Anything else, let me know and I'll see what I can do.
>>
>>34227444
I downloaded all the fim/eqg that came up in the Netflix search. The page you linked comes up 404 for me.
>>
>>34227444
>>34227494
>>34227741
If you have any episodes/movies/etc in any language then please upload them to me or give me a link so I can download and add them.
I'm sadly missing a lot of non-english stuff.

Upload/give me anything pony you have in general.

>>34227741
I have more than enough space. If not I'll just kick off some of the horse porn again.

Sure, I'll take it.
>>
>>34228829
The anonfile links in the archive are all I got. Generally whats contained is all the 5.1 audio tracks that were available on netflix. More languages were available in 2.0 which I did not get. I did not get video tracks either, but should sync to any english netflix rip.
>>
>>34228870
could you make a list of of the 2.0 audio?
maybe even get them for archival purposes, at least until a 5.1 source is found for them.
also you forgot to get the audios for the first eqg(https://www.netflix.com/en/title/70276351)
>>
>>34216515
>Trixie
Based uberanon.
>>
>>34216515
It would be interesting to see one on Spike.
>>
>>34229245
I ripped the audio using the trial of the software. That trial is now expired Though there's likely a way around that. More importantly I should mention that the ripper very much likes to crash. Combined with the fact the you have to manually select each and every audio track for each episode. Then when it crashes all selections are lost. There were times I spent close to half an hour ticking all the boxes and the thing crashed on me and I had to restart. Was kind of a pain. So basically, I was going for somewhere between getting only what was needed for the project and a wider preservation effort without having to get everything. If we need to do a second round of audio ripping for whatever reason, I might do it then. But otherwise I don't know that it will be high on my to do list. I'll see about getting a list of languages later.

Your Netflix link still 404's for me. Pic related.
>>
>>34227784
>>34228600
Here's some dubs.
https://anonfile.com/Z465u147ne/EQG_Legend_of_the_Everfree_English_5.1_flac
https://anonfile.com/bb77u042n5/EQG_Legend_of_the_Everfree_Finnish_5.1_flac
https://anonfile.com/e07eud4en5/EQG_Friendship_Games_English_5.1_flac
https://anonfile.com/n876uc45n4/EQG_Friendship_Games_Finnish_5.1_flac
>>
>>34229774
try this link https://www.netflix.com/us-en/title/70276351
also what is the program?
i ask so that the anons here can search for a workaround
right now i remember something that might work,by creating another user on the computer and using the left click button on the executable it might give you an option of opening the program as another user, i could be very wrong though.
>>
>>34229809
Still a no go for me on that link. If you're interested in ripping audio yourself, here is the ripper:
https://www.flixgrab.com/
Make sure to get the one from this site as apparently some else copied the program and are putting it out as they're own. The only reason this matters is because the copied one does not work.
>>
>>34229839
how about this one?
https://www.netflix.com/es-en/title/70276351
and if that one does not work, click the mlp related "More TV Shows & Movies" until you see horse twilight looking at her human version in the mirror portal.
also if the user thing did not work for you see if one of the thing listed here works for you
https://www.tricksforums.net/how-to-extend-or-reset-trial-period-of/
>>
>>34224147
Phoneme-level alignments for S1-S8 and the movie: https://anonfile.com/U3y8v74cnf/mfa-alignments_zip.
>>
>>34228600
>>34229784
Thanks, I’ll get on those.

>Anything else
Everything that has a 5.1 dub version available. I’ll get round to them eventually.
>>
Bump
>>
>>34189332
>>34227314
I’ve re-clipped s3e12 - 13, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

I’m currently running season four through iZotope, will upload and post links when done.
>>
>>34232018
Doing gods work!
>>
>>34232018
Uploaded the cleaned season four English tracks, will do the raw stereos tomorrow - https://anonfile.com/d1acz34anb/S4_zip

>>34232396
I aim to please.
>>
>>34230335
Great job!

>>34232018
I will update the image tomorrow. Great job too! as usual
>>
>>34230335
Great job!

>>34232018
I will update the image tomorrow. Great job too! as usual
>>
>>34233344
remember to add these to image
Rainbow Roadtrip
Sunset's Backstage Pass
DVD/Bluray Extras(Baking with Pinkie, etc)
>>
bump
>>
>>34233904
Oof
>>
When you get to "Rarity Investigates!" (5x15), keep in mind that the film noir scenes were intentionally downmixed to mono, so it may be impossible to extract clean dialogue from them. (This was the reason RainShadow gave for not being able to extract clean BGM.)
>>
>>34234717
I didn't know they put so much effort in details like this.
>>
>>34189332
>>34232018
I’ve re-clipped s4e1 - 6, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

Raw stereos for season 4 - https://anonfile.com/taT0244fne/S4_zip

>>34234717
I remember that one, the music was still there the first time around. There are a few other scenes where music still remains, presumably due to the different mixing method you mentioned. We’ll see what iZotope can do about that, but I’m not overly hopeful. Not much can be done for clips like that unfortunately, as the clearing method struggles with sound effects that are rapidly oscillating or of comparable or greater volume to the dialogue.

I’ve added an example of such a clip to the front page of the master file, it’s the Vault-Tec style intro video explaining the tornado power concept from s2e22, Hurricane Fluttershy. The before and after should give you a decent idea of what to expect for segments like the film noir part. It’s certainly a significant improvement, but not really good enough to feed to the AI.
>>
>>34235863
>>
>>34236934
fuck my bad i didn't mean to attach that to the anchor
>>
can someone do cheerilee?
>>
>>34236943
I don’t think she will have much audio to work with but hopefully someday
>>
>>34236730
the dvd/bluray extras can also go for the mlp movie(baking with pinkie and others)
are you sure that we have good audio for Rainbow Roadtrip and Sunset backstage?
we do have the audio for the birthday thing in rome in a fuckton of languages, so you can mark the high quality soundtrack box on that
>>
>>34237473
The issue with the Happy Birthday short is that it is only available in stereo when ideally we would like 5.1 to be able to use all the techniques.
>>
How many of you anons have set up working models? Also, is there an anon in charge of creating the e2e model and or running it?
>>
>>34238192
Training deepvoice3 third day, also working on nvidia tacatron+waveglow
>>
Have any of you anons also had any experience in voice acting? It could provide much help to me, or some of the other anons ITT, if you could talk in depth about voice acting; and what are the key factors that distinguish, the good, and bad voice actors.
>>
>>34229774
>>34229809
>>34229839
>>34230157
one word - license
if something is no longer available in one country you can't play it
it's that hard to notice?
>>
>>34238305
An anon creating image-descriptions of what each mane character sounds like, and the common mistakes that voice actors will make. I think its in the OP somewhere
>>
>>34238440
Here they are https://desuarchive.org/mlp/thread/34019408/#34019800
>>
How's it going lads?
>>
babump
>>
Any new voice clips?
>>
>>34239331
Mos active anon works on cleaning as much audio as possible, so nothing really new.
But at least one anon is working on ML and NN, so we may have some new things soon.
>>
Is there a torrent for the EQG source files? I'd like to download them but anonfile keeps crapping out on me mid-download for some reason.
>>
>>34240342
Guess I can make one
magnet:?xt=urn:btih:87e53a217a637cdd1be59ae2093889ea26194a07&dn=EQG.7z&tr=udp%3a%2f%2fexodus.desync.com%3a6969&tr=udp%3a%2f%2ftracker.leechers-paradise.org%3a6969&tr=udp%3a%2f%2ftracker.uw0.xyz%3a6969%2fannounce&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce&tr=udp%3a%2f%2ftracker.coppersurfer.tk%3a6969&tr=udp%3a%2f%2ftracker.kamigami.org%3a2710%2fannounce
>>
>>34189332
>>34235863
I’ve re-clipped s4e7 - 16, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

I’ll finish the rest of season four tomorrow.

>>34239108
>>34239331
Progress here - >>34236730
Best clips so far - https://clyp.it/px11j2wn
Archive of all clips - https://u.smutty.horse/lqctkqxclef.7z

Cleaning what’s available will probably take another week or so at the rate I’m going, so likely won’t be any new voice clips for a short while.

>>34240342
Seconding this. Anonfile has been really temperamental recently, an alternative option would be appreciated.
>>
>tfw trying to torrent the dataset is slower than having sex recursively, with multiple copies of Granny Smith.
>>
>>34241135
There's Zippyshare of course. Downsides are a one month link life (from last download) and a 500mb limit per file.
Nofile is most similar to anonfile. However in my experiance large files have a rather short life on there.
Mega is nice as it's fast, indefinite link life and 50gb per free account. Down side is the 5gb per 8hr limit.
I'll look into some other hosts.
>>
>>34241920
The solution doesn't have to be overly elegant, at least not for me. All I really need at the moment is a 5.1 dub for everything that's available, which will be a one-time thing that most others won't need. Rome should be able to deal with anything else that needs to be archived, and I'll maintain the master file for as long as is necessary.
>>
>>34241920
I can still mirror about 500GB here if that helps: https://pubshare.ponemusic.net
I just can't download from Mega very fast, they throttle me like mad.
>>
>>34241920
/tg/ has mostly gone over to yandex.ru and allsync.com now that mega only gives you 15 gigs free after the trial period ends.
>>
>>34241135
>>34241920
>>34241960
If you need any unlimited webspace where you can upload your stuff for this project I can just give you a sftp/ftp/ftps/webdav/rsync/webupload/etc.
>>
>>34242353
The more mirrors, the better. Perhaps we should organize a directory like we had before when clipping episodes. If you want to give a place for a mirror, I can start uploading what there is. I kind of have a slower connection. Can do maybe 60gb-80gb per day if everything goes well. Maybe mirror on other hosts as well. Perhaps on anonfile/bayfile/megaupload or whatever.
>>
>>34241920
>Down side is the 5gb per 8hr limit
try this
https://github.com/tonikelope/megabasterd/releases
>>
>>34216515
Is there a place where I can see the rest of these profiles?
>>
>>34242541
I've used it. It's not the best. What I would recommend if you want to download from mega is to use jdownloader 2. Just make sure to get the adware free version from thier forum. This works well with large files because if you started a download a download, mega won't cut you off mid download. Increase the amount of simultaneous downloads, set up your queue, and start them together. Doesn't work great with lots of small files though. Best to use the official client for that.
>>
>>34242958
here brah >>34238507
>>
>>34238217
You're doing all this on a NOTEBOOK?
>>
>>34243068
Thanks
>>
>>34239331
Any requests for lines?
>>
>>34243624
not him but how about the little mermaid IMAGINE meme?
>>
>>34243704
I dont know what are you asking for.
>>
>>34243624
"I'm Twilight Sparkle. And this...is Jackass."
>>
File: 1567325121429.gif (1.35 MB, 505x506)
1.35 MB
1.35 MB GIF
>>34243723
https://drive.google.com/open?id=1SVmhBF2UHm9LEi2l5nFf6hHAmY2N20DR
Uploaded.

Noise+Clean result is worse than only clean. Even after 300k of steps.
I am waiting you guys to finish cleaning, and for now will try to set up tacatron and waveglow.
>>
>>34243816
Do you suppose it is worse because of the current smaller cleaned dataset?
>>
>>34243816
Still cute
>>
>>34243823
I think its because of noised material. Their sound material dont match with quotes good
>>
>>34243950
Realized I misread that. I thought you were referring to just the noised lines that were cleaned. Makes more sense now.
>>
>>34243816
Is it me or are the name mixed up for the files?
Anyway, great job at setting up a new NN!
>>
>>34244175
Yea, they're mixed. I am lazy to fix it.
Thank you, anon
>>
>>34244178
It's okay, not a big deal.
Just disturbing at first.
>>
>>34189332
>>34241135
I’ve re-clipped s4e17 - 26, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

>>34227343
>Season five
Have you been able to process any of it yet? If not, I could run it all through iZotope tomorrow. Also, is there a link for the season six Finnish material?

>>34243816
Was this using the old material or the newer stuff? Would be interested to know if being stricter with the noise tags is helping with quality output.
>>
>>34245083
I just got through sorting the languages for season 5 yesterday. I will try to align today. Season 6 Finnish material has been posted as well as izo processed. Check the doc or the thread for season 6 Finnish audio.
>>
>>34245252
>iZo processed
Is it one of these?

https://anonfile.com/A2Xaf330na/Season_6.7z_001
https://anonfile.com/79Xbf732n5/Season_6.7z_002
https://anonfile.com/DbWff130nb/Season_6.7z_003

They're all quite big files, so I don't want to waste time downloading stuff I don't need.
>>
>>34245271
Those are the English and Finnish RAW 5.1 Netflix rips. I had 7zip split into 4gb file to make them easier to upload. The season 6 aligned and iZo processed I posted in the previous thread >>34184453. Links below:

https://anonfile.com/8dz5Be32n4/Season_6_Processed_7z
https://anonfile.com/c6afCb37n6/Season_6_Aligned_7z
>>
>>34245322
Alright, thanks. I'll try downloading that tomorrow, hopefully Anonfile has stopped being slow now.
>>
How many people do you guys have working on this? Are there different people for each pony? Also do you have a backup site if this one goes down for an extended period?

thanks for not making a discord or any social media
>>
File: 1567365904380.jpg (170 KB, 1108x1079)
170 KB
170 KB JPG
Well, apparently Google colab has not enough of GPU ram for training tacatron with default settings.
I will dig more about it, but also i've started to think about using aws/azure.
I dont like that they give only couple of month, tho.
>>
>>34245543
Have you tried Colab with the TPU enabled? It should have stupid amounts of RAM (good enough for almost everything except big Transformers)
>>
>>34245543
You don't have a rig set up for this kind of thing?
>>
File: 1567366265361.png (310 KB, 750x844)
310 KB
310 KB PNG
>>34245549
I tried, tacatron have failed on the step of showing GPU information, like it didnt found one.
I have had similar problem with only CPU environment.
Should i try fixing this issue? Probably TPU will help, NVIDIA are advicing to use tensor cores.
>>
>>34245571
>I tried, tacatron have failed on the step of showing GPU information, like it didnt found one.
Yeah, you have to make a small change when initializing Tensorflow to tell it to use the TPU instead of GPU.
>>
>>34245559
What do you mean by "rig set up" ?
>>
>>34245576
By rig I meant like home machine/computer with a 10 series gpu.
>>
File: 1567366583920.jpg (73 KB, 540x540)
73 KB
73 KB JPG
>>34245574
Ok, i will search more about this. Thank you!

>>34245581
Nah, of course not. Who am i? The real ML-researcher or Billionaire?
I have access to my old university cluster, but it is worse than google colab or other free(almost) services.
>>
>>34245594
How did you pass your data into the model for training? Also, in general, do any of you guys know any py libs for operations (or maybe even just file conversions) with .flac files?
>>
File: 1567367936948.gif (3.47 MB, 431x306)
3.47 MB
3.47 MB GIF
>>34245679
I wrote converter for creating dataset from audio, for deepvoice and tacatron.
https://github.com/Twibot-ai/audio_proj_utils
(tacatron creation is not pushed yet)

You can also look on google colab notebooks
https://drive.google.com/open?id=1XBqXf3yeBKyQUTAfA0LuhGzeugSaNSIY
https://drive.google.com/open?id=17TMYl3anIwegk-sBU6jHngpqo_3sjtW8

> operations
i am using AudioSegment. It uses ffmpeg libraries.
>>
>>34245456
At least 5? There's been other Anons who stopped by and did something, but I'd say there's about that many who've consistently been here.

The work so far should be applicable to all ponies. Twilight is who we plan on getting going first because she has the most lines and so the most likely for success.

We'd like to host relevant files on several sites. If you mean a bunker for /mlp/ it hasn't been discussed. Seems unlikely for anything to happen, but if it does I'm sure people will put the word out on whatever pony sites they know of.
>>
>>34245456
>>34245739
5 seems a good number, even if it changes over time.
Mostly, 2 ML/NN Anon ("Twibot" and "SyntheticTwiggles") and "Clipper".
Back in the beginning, there was like 2 more clipers anon, and two Pyfags (I was one of them and made the audacity label converter/normilizer thingy). Then at least one anon that helped discuss with noise cleaning, and two that tried to figure out what to do and how to normalize phoneme (both of wich with a program).
Some anon have done a great job, like the one writing the "voice description" images of the main 6, or two drawfags.
I absolutely don't know if some anons have multiple caps (as implied by the "Anon" thing, you know?) but I can say that there surely is a nice group of dedicated Anons working, supporting or lurking for this thread.
Sorry if I do miss some.
>>
File: 1545723487571.png (342 KB, 779x1024)
342 KB
342 KB PNG
There, take a cute pone as a bump.
>>
>>34246913
Thank you for the cute pone bump
>>
Bump for science
>>
File: 1480733341124.gif (1.08 MB, 320x240)
1.08 MB
1.08 MB GIF
>>34216515
>>34243068
that's super impressive, autistic work. with people like you here we might just do this.
>>
File: tenor.gif (1.2 MB, 498x289)
1.2 MB
1.2 MB GIF
>>34189328
>when the pony preservation project, sex robots, and synthetic orifices + fur tails etc are all being developed
>>
>>34249300
Waifus are upon us lads
>>
File: yesyesYES.gif (471 KB, 267x200)
471 KB
471 KB GIF
>>34249334
Look how good this shit looks, now imagne a robot inside of it, and it can talk to you and its got onaholes or whatever fused to it and it self heats and lubricates and shit

https://twitter.com/i/status/1168204804889485313
>>
>>34249389
you know that's not real, that's a 3d render imposed onto a video
>>
File: computer model.png (306 KB, 563x830)
306 KB
306 KB PNG
Holy fucking christ my guys, look at this shit. What comes before a building? A blueprint, this is like a fucking blueprint for the sexbots
>>
>>34249396
yeah just saw that, all the same though, if they can make sexbots that are humans they can do it for ponies or whatever furry you want like renamon retsuko etc
>>
File: holy doolie.png (370 KB, 395x837)
370 KB
370 KB PNG
>>34249401
>>
File: ass.png (476 KB, 585x891)
476 KB
476 KB PNG
>>34249412
look at her ass
>>
File: thicc.png (259 KB, 582x899)
259 KB
259 KB PNG
>>34249416
>>
>>34249401
>>34249412
>>34249416
>>34249423
Sadly I think this particular dude only do sculpting, not the whole rigging and shit that would allow to actually make something with it (either in VR or printed as a robot's hull pieces).
Rocks for figurine printing, but that's it.
>>
File: twi.png (272 KB, 579x837)
272 KB
272 KB PNG
>>34249428
It dont matter bro, the point is if he can do it we can do it, someone can do it. Its like looking into the future. Just imagine it.
>>
>>34249408
>if they can make sexbots that are humans they can do it for ponies or whatever furry you want like renamon retsuko etc
I'd imagine if sex bots become a common thing in the future making a quadruped sexbot would be a ton harder and less standardized than a normal one. Will pony sexbots be in demand enough for a manufacturer to consider making one? Probably not.
>>
File: sexbot interior.png (1.68 MB, 1117x891)
1.68 MB
1.68 MB PNG
>>34249428
>>34249435
>>
File: 235.png (216 KB, 576x687)
216 KB
216 KB PNG
>>34249442
Bro you already know people are gonna want furrybots, and theres definitely enough demand for it to be profitable for them. People are gonna want ponybots, you know even the furries are gonna wanna try pinkie out for a bit. Plus we can always cough up the cash for custom orders, I'm sure the chinese koreans and especially the japanese will have 0 problem making them, those people are freaky and dont give a fuck lol. Giant loli statues outside and shit
>>
File: 1507760026504.png (36 KB, 358x349)
36 KB
36 KB PNG
>>34249456
A good sexbot would be insanely complicated and would take a long time to blueprint. It's not something you could just order "custom" beyond specifying to the manufacturer what size boobs you want or what color should the bot be.
You don't want your ponybot to just be a recolored ugly horse, right? You want it to look like your waifu. And that's going to need demand to be concepted and brought to life.
I also have a feeling ordering a sexbot would be close to impossible due to shipping restraints, costs, and just generally being illegal. How will they send you a pony sexbot from korea or japan?
>>
>>34249490
Dude do some research online about how sexbots are coming along. Its not that hard to make a pony or furry one. The hardest part of all is making the AI so its interactive and shit, other than that is pretty simple.
A furry one is just a person with fur and tail and ears and shit, a pony one is harder but they can create a base default model that they slap the colors and shit onto, so pinkie rainbow whoever would all have the exact same proportions but would have different hair colors and shit
>>
>>34249490
>>34249531
and it wouldnt be illegal either, you can ship anything you want as long as its not on the list of banned shit (google it), cia leave me alone
>>
I'm trying to download the iZo processed audio for season six from >>34245322, but Anonfile is still doing that thing where it randomly fails mid download for no obvious reason, similar to the problems >>34240342 was having. I've tried messing with my firewall/VPN/router/browser settings to no avail, which leads me to believe that the problem may be with Anonfile itself.

Could someone please re-upload the iZo processed season six audio somewhere else for me?
>>
>>34249676
It seems to be on their end, I get 503's after they interrupt it (and of course they don't support resuming).
>>
File: 1564854068460.png (168 KB, 328x284)
168 KB
168 KB PNG
>>34249531
and how would they dodge the issue of copyright?
>>
>>34249676
That's unfortunate that Anonfile is having issues. Any particular host you'd prefer me to reupload to?
>>
>>34249999
R34, porn and co are weirdly sorta immune to copyright claims.
Not for legal reasons of course, but because PR departments generally get hysterical about *any* light being thrown on that.
Unless it's already "main target audience" knowledge or really bathing in bucks, most companies won't let their lawyers poke at it even with a ten foot pole.
>>
>>34249531
Honestly, the only downside is the cost. Even if you're making a "dumb" sex-model, flesh-like platinum cure silicone isn't cheap yo. Most of the cost of a Bad Dragon toy goes into materials, after all

Add on proper servos, warmers, and whatever additional features you want, and that stuff starts building up fast.
>>
>>34250019
Given the choice, I'd prefer Mega, since I know that it works and haven't had a significant issues with it so far. Thanks.
>>
>>34249999
Also, checked
>>
File: 1567394798350.png (767 KB, 1496x844)
767 KB
767 KB PNG
>>34249999
See >>34250045

>>34250060
Yeah it'll probably be a few grand, it is what it is. Get a job and save up, if you have an actual salary earning job where your making tens of thousands a year it shouldn't be hard unless you blow it all on stupid shit
>>
File: Delay.png (2 KB, 178x52)
2 KB
2 KB PNG
>>34250067
On the way.
>>
>>34245679
I'm using SoundFile and LibROSA
>>
File: 1566450377786.png (326 KB, 500x481)
326 KB
326 KB PNG
>>34249999
>>34250045
>>34250156
Also even if the lawyers start gunning for the sexbots, there's gonna be ways to customize them like you would a car. Changing the voice is as easy as messing with the settings, and as for the colors that make the bot up (all have same base model but with essentially different paint jobs and hair styles), I bet you could buy customization kits. Plus these companies could always avoid using their real names for the characters, or make extremely subtle but legally safe changes to them, so its not technically copyright infringement

You gotta believe in the dream
>>
>>34250164
>Almost six hours.
I guess Mega hates you as much as Anonfile seems to hate me. Strange how these things happen.
>>
>>34250174
The beauty is chinaman or ivan don't give a shit anoit copyright law so even if shit is banned from production they will make them for you anyway
>>
File: this2.jpg (3 KB, 125x125)
3 KB
3 KB JPG
>>34250268
>>
>>34250268
As ivan, i would say that we have russian hasbro, and they could use copyright law, if it will be something big.
>>
>>34249442
You have it backwards.
Human like robots would be much harder to create due to a number of factors.
>Uncanny valley efffect
>A biped is relatively shit at keeping their balance compared to most other animals, resulting in more expensive hardware / designs for the robot keeping itself upright and in motion.
>Thots would be throwing everything they goddamn have at the idea of something that could even potentially replace a woman.
An animal robot though?
>Therapy animal.
>Replacement / Alternative for assistance dogs.
>Easier to make them look cute.
>Thots wont target it as aggressively since it wont be marketed for sex by normies.
These will 100% be on the market first before any others, Boston dynamic's spot mini being the first iteration of such.
All it will take is some mad lad making a DIY video for adding an onahole to it and it'll take off from there.
>>
File: 1460685310940.jpg (290 KB, 1920x1080)
290 KB
290 KB JPG
>>34250537
good points. it's going to be a lot easier to make a ponybot not look uncanny
>>
>>34250537
>Boston dynamic's spot mini being the first iteration of such. All it will take is some mad lad making a DIY video for adding an onahole to it and it'll take off from there
>Putting a BD in a BD
>>
File: was gucci.gif (342 KB, 500x377)
342 KB
342 KB GIF
>>34250537
women get so triggered at the idea of sexbots because they dont want competetion, just like prostitution and the AOC being any less than 18 (some want it higher than that)
All of the above non ironically scares them, they know they are parasites and they are terrified of us having some way to get pussy without having to cater to them

They wont be able to stop the sexbots though, its way too profitable
>>
>>34250164
>>34250219
Here you go:
https://mega.nz/#!Yg4EAQ6C!clcIm77JBUExLRC-c-2LpBrfw01UOROYXfdp4l_8ODU
https://mega.nz/#!o8g13IxQ!rBaU1dKySgv4Z3mzyOylWoWR38uoXA7SWpnAhg-ccBk
>>
I'm not a ponyfag, I just heard about this project on /v/ and thought it was interesting. Is the plan to make entirely AI generated shows or just use AI to replicate the voices of the actors?
>>
>>34251497
The goal right now is to make AI VAs as first, a proof of concept and after refining, a resource for other fan works and content creators to use freely.

Experimenting with AI to do other parts of a content creation process like animation are things for future anons or inspired people to figure out after we're done here.
>>
File: 1556790948110.jpg (12 KB, 480x360)
12 KB
12 KB JPG
>>34251497
lmao someone actually came here
sorry anons, it's my fault. I posted it in a dead thread, thought nothing would come of it
>>
>>34251508
I'd like to help out but I don't really have the time to do so or know much about ML, I wish you all the best of luck and hope this works for you all.
>>34251521
You can't just post something as interesting as that and not expect people to come. I won't tell anyone, I read earlier in the thread how hasbro treats fan works.
>>
>Page 10
>>
>>34245708
>>34250172
Damn, I need to look at this again.
Unfortunately, my drive is full, I must buy a new one (or clean this one... Naa better buy a new one)
>>
bompo bampo
>>
The same ponychan guy who hacked dhx in 2017 posted 922-926 in English. The archives are huge, around 10GB for one episode, maybe they have lossless sound.
>>
>>34253664
822-826, and already used, i think
>>
>>34253691
No, season 9. The episodes that were leaked by Dutch.
>>
>>34253664
link to the ponychan post?
>>
>>34253715
https://www.ponychan.net/pony/res/36830779.html
>>
>>34251390
Thanks man, I'll start working on those now.

>>34253664
>>34253720
I'll make a note of these for later, hopefully they'll be useful.
>>
>>34253664
Can confirm they have separated audio streams. No audio recovery needed. I have 9x22, somebody else can get the rest
>>
>>34254156
How did you get it? Mega limited me after 5GB, not after the "usual" 10.
>>
File: 1507508560114.jpg (125 KB, 707x1131)
125 KB
125 KB JPG
>>34254156
>Can confirm they have separated audio streams.
oh shit
>>
>>34254283
Try these to skirt the limitations
https://megadownloader.en.softonic.com/download

https://megatools.megous.com/
>>
>>34254156
Oh shit, really? Does this mean we have access to just the raw dialogue track?
>>
>>34253664
>>34254156
Am downloading now. Will post audio when done.
>>
>>34254156
Dam this is good.
But does this mean that your work on cleaning is useless?
>>
>>34254283
I've started downloading, mirroring them here: https://pubshare.ponemusic.net/F/
>>
>>34254418
It's just the last 5 episodes of the series that have separated tracks.
>>
>>34254440
Oh, ok.

But could we ask hackintosh to download the rest?
>>
>>34254419
Not that I'm complaining but I dont think whoever hosts that site will be happy. Nor from sudden high traffic and nor from the possible legal stuff.
>>
>>34254480
It's cool, I have permission. As long as the links don't get too public, traffic in this threqd is fine.
>>
File: 1125077.png (242 KB, 373x554)
242 KB
242 KB PNG
This audio must be lossless considering it's bringing Audacity to its knees. Is there a way to quickly separate audio tracks from video that isn't audacity?
>>
>>34254506
ffmpeg -i pony.mov -vn out.flac
>>
>>34254511
handhold me
>>
>>34254511
does this result in seperate files, or just one?
>>
>>34254524
He just did. The command is "ffmpeg -i [name-of-video-file] vn out.flac" It takes the file, tosses the video, and writes the audio out to "out.flac"
>>
File: streams.png (140 KB, 1406x784)
140 KB
140 KB PNG
>>34254524
Go download ffmpeg, doesn't matter if you're on Windows/Mac/Linux.
Copy the ffmpeg.exe and the episode file in a folder, and open some kind of Terminal or Powershell in that folder.

Now you want to find which audio stream you're interested in. There's several from #0:1 to #0:13, see picrelated. You can just do them all one by one until you find the one you want

Say you think audio stream 0:1 is the one you want, you type this command in the terminal:

ffmpeg -i 922.mov -vn -map 0:1 922-03.flac

Or something like .\ffmpeg.exe -i 922.mov -vn -map 0:1 922-01.flac if you're on windows.
Replace the 0:1 with 0:1 and 922-01.flac with 922-02.flac, etc until you're done.

>>34254559
Separate files in mono, one per track.
>>
>>34254524
Here's the tutorial I used earlier - https://www.youtube.com/watch?v=MPV7JXTWPWI
>>
>>34254440
I've asked if there's any way he'd be able to swipe the rest of the show for us
https://www.ponychan.net/pony/res/36830779.html#36830817
>>
>>34254562
if i just do >>34254511, will that be acceptable, or should i extract them one by one?
>>
>>34254592
It'll extract the audio, but I'm not sure if that's going to keep each channel separate, or downmix everything to stereo.
>>
>>34254597
iirc flac supports 5.1 audio
>>
>>34254597
>>34254609
Can confirm from my work earlier that it will keep all the channels separate.
>>
>>34254609
>Invalid audio stream. Exactly one FLAC audio stream is required.
I don't know the ffmpeg incantation for 5.1 then.
Anyways, I've extracted all the audio streams from 922 separately since I have nothing better to do.

magnet:?xt=urn:btih:3d7761c954e5e5aaaa115ebb66070347127fdec3&dn=922.zip&tr=udp%3a%2f%2ftracker.leechers-paradise.org%3a6969&tr=udp%3a%2f%2ftracker.uw0.xyz%3a6969%2fannounce&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce&tr=udp%3a%2f%2ftracker.coppersurfer.tk%3a6969&tr=udp%3a%2f%2ftracker.kamigami.org%3a2710%2fannounce&tr=udp%3a%2f%2fexodus.desync.com%3a6969
>>
File: send help.png (33 KB, 666x540)
33 KB
33 KB PNG
>>34254662
its autodetecting to mono. Should I be worried?
>>
>>34254680
If you do them separately, I know for sure you have nothing to worry about, because just you end up with 12 clean sepated mono channels.
Maybe some other anon can tell you if there's a better command
>>
>>34254680
I don't know, but probably not. All I ever did was type the command line that was shown in the video. I don't actually know how any of it works, but it all turned out fine each time.
>>
>>34254685
>>34254694
You know what: I'm kinda glad someone else did my work for me. >>34254679
My heart was in the right place, and that's that matters at the end of the day.
Now to get this stinker off my disc
>>
>>34254694
>>34254680
>>34254609
OK I know why it doesn't work, that's not 5.1 it's 10.2!
12 channel audio is a thing, appparently
>>
>>34254713
Hey if you have 924 and 925, I'd be really glad if you could help! I've blown off my Mega limit.
>>
>>34254729
Use an external downloader >>34254346

Also protip: wait for the subtitle file to be released and the audio wizards to merge the audio before attempting to clip the audio
>>
>>34254494
Have you managed to get 24,25? Just finished DLing the others.
Damn they are large. 37Gb for a single episode. Much larger than the leaked movie master.
>>
>>34254749
Nope, I though the anons over at Ponychan were going to reupload it but apparently not
>>
>>34254749
>Damn they are large. 37Gb for a single episode. Much larger than the leaked movie master.
Actually not that far! The movie is 100 minutes and 128GB, the episodes are 25 ish minutes and 35 ish GB
The leaker is the same and the format is the same, I wonder if he's had access to everything all this time..
>>
>>34254756
Reupload? Why?
>>
>>34254781
Some guy gave in and paid for Mega Pro apparently. I have like 15GB download limits on mega right now...
>>
>>34254479
I already did, but he posted the leak 2 days ago so who knows if hell even see it
>>
File: full.gif (667 KB, 573x521)
667 KB
667 KB GIF
yay active thread
>>
Clean voices for 922
https://www47.zippyshare.com/v/N5l1uR7R/file.html
>>
>>34254749
I got them finally, they're up in the /F/ folder in the pubshare.
>>
>>34189332
>>34245083
I’ve re-clipped s6e1 - 6, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

>>34255101
Thanks. Are you going to do the rest of the leaked episodes?
>>
>>34255174
>Thanks. Are you going to do the rest of the leaked episodes?
Yep, I'll extract it from the rest in few hours.
>>
>>34255116
Oh and in case anyone needs them my ffmpeg's done extracting the audio tracks from 922 thru 926, so I've added that to the folder.
Keep up the great work y'all
>>
>>34255557
So you extracted everything already? Should I upload voice tracks from next episodes like this >>34255101
or not?
>>
>>34255975
I just uploaded the extracted audio tracks. Not trying to step on your toes, the more the better.
>>
>>34255557
if you did things correctly, ffmpeg should take no more than 5 seconds to extract the audio of each episode.
if it takes any more than that, you're using a different output codec
each mono track has a 1152 kb/s bitrate.
the resulting stereo track should be exactly double that, 2304 kb/s

use these two commands. The second just zips the audio up into the flac container, can't pipe the source streams directly into .flac

ffmpeg -i INPUT.mov -filter_complex "[0:a:8][0:a:9]amerge=inputs=2" -c:a pcm_s24le -vn OUT.wav
ffmpeg -i OUT.wav OUT_FLAC.flac

9x22 .flac should be 148262kB total
>>
Just popping in to wonder, are you guys gonna use the official script leaks or are you going with your personal autisms? Or are you gonna diff/cross-reference them somehow?

I feel like listening might be more accurate in case the VAs took some creative liberties at any point at all, and didn't follow the script to the letter; but on the other hand IIRC that was prone to typos and inconsistencies which might not be realistically possible to totally weed out over the dozens and dozens of hours of transcribed footage, which the script wouldn't be affected by. I'm sure that's already been discussed though (I see you're on to extracting the raw audio from the leaks) so I'm just wondering.
>>
Hey Twibot, have you tried https://github.com/CorentinJ/Real-Time-Voice-Cloning ?
>>
>>34257338
I don't believe anybody is using the scripts, just going with straight up transcribing the real audio.
>>
File: 1555448463380.gif (91 KB, 336x392)
91 KB
91 KB GIF
Hey all, I'm new to this thread but not this field. I may take a crack at this and post results I get. Will update by Friday.
>>
File: 1567573480407.png (199 KB, 978x1024)
199 KB
199 KB PNG
>>34257353
I did not. I am still trying to run tacatron and also playing around with dataset.
I will try to find time and check this out.
Thank you, nonny!

>>34258242
Hey, this is nice to see. I guess i need to share what we already did:
Somewhat working jupiter notebook with deepvoice3:
https://drive.google.com/open?id=1CDLuciVPkZXw0N2-Jgm0-Ye5l_f7hdVn

Demo with using this models and links on checkpoints:
https://drive.google.com/open?id=1XBqXf3yeBKyQUTAfA0LuhGzeugSaNSIY

Also small tool to create dataset from audio for deepvoice3:
https://github.com/Twibot-ai/audio_proj_utils

Currently i am trying to create new notebook with tacatron and waveglow training.
The main problem i've had encounter that we have only around ~1 hour of clean audio. When i used any noised data the results were much worse then with clean.
So it would be awesome if you we collaborate (here of course) and try to find best training params for deepvoice or find and setup other model.
>>
>8
>>
File: 1557816559670.png (209 KB, 876x1024)
209 KB
209 KB PNG
>>34258242
welcome
>>
>>34258827
>9
>>
File: 1489514785482.jpg (30 KB, 357x392)
30 KB
30 KB JPG
>>34258316
Thanks, I'll play around with this and see what I get.

In the mean time have we considered our own private hosted GitHub instance or even a GitHub Org within public GitHub if we don't want to go through the trouble of self hosting?
>>
>>34259525
not the anon you where responding to, but gitlab would be good to
>>
>>34259525
Why private tho?
I am ok with my fake github page.

Shouldnt it be available for all?
>>
>>34189332
>>34255174
I’ve re-clipped s6e7 - 12, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

>>34255975
If you have just the voice tracks, you could still upload those separately as you did for 922. Would be easier for me to download as the pubshare thing from >>34254419 seems to be everything all together.

>>34257338
I use a combination of the subtitles, wiki transcripts, and my ears. I find that errors in the transcripts are quite rare, in fact most typos that exist in the dataset were most likely caused by me being clumsy. I’m correcting every error I find as I go through with the cleaning, but I’m sure a few will still slip through the cracks.
>>
>>34260153
To echo the sentiments of >>34217411, this too will eventually be seen by somebody on Hasbro's legal team. The C&D hammer will come down hard when this gets far off the ground and a git repository can be easily cloned to anybody's computer and redistributed to any git-based hosting platform.

We can also just leave it public so anybody can preserve that info. I don't really care about the specifics, I'm moreso concerned that data will go missing otherwise. Any git-based code repository would work well. Other options are >>34259623, BitBucket, etc. I can also put down cash to stand something up as well.

I brought it up since it's a priority and I wanted to get a discussion going. I personally like to checkpoint my work as much as possible and wanted to know what the current effort for code repositories were.
>>
>>34260434
Good job, as always!

>>34260513
It may be a good idea to back up the code, but -sorry for asking- isn't all the work done so far very dataset related?
I mean, I don't want to sound rude, downloading and making the program work is kind of an achievement when you see the mess it can be, but aren't the files configured for this specific dataset? (and "useless" for other dataset?)
I'm probably wrong, I don't know shit for now, but I'm willing to learn one day or an other.
>>
>>34260434
Here ya go
923
https://www104.zippyshare.com/v/bAS5W369/file.html
924
https://www104.zippyshare.com/v/bn6lCA5X/file.html
925
https://www104.zippyshare.com/v/MPZM80NE/file.html
926
https://www104.zippyshare.com/v/Z5e1xixa/file.html
>>
>>34261015
Cheers.
>>
>>34260434
>>34261015
Beat me to it. Here's a megaupload mirror as well because it's already done.

https://megaupload.is/K4mbS945nb/922-09_flac
https://megaupload.is/I3m9S24bn0/923-09_flac
https://megaupload.is/L4m9S749nb/924-09_flac
https://megaupload.is/Jem8Sf41n7/925-09_flac
https://megaupload.is/T1mcSb47n1/926-09_flac

Taken from the 9th track of each. Seemed identical to track 10. Sounded cleaner than track 3.
>>
>>34261125
>Taken from the 9th track of each. Seemed identical to track 10.
You fucked up m8, it's a stereo mix. So things that are placed on the right are missing in your upload.
>>
>>34261125
>>34261182
Should be stereo now:
https://megaupload.is/33x7S74bn8/922_flac
https://megaupload.is/xax0Sd4cn1/923_flac
https://megaupload.is/22x2Se49n6/924_flac
https://megaupload.is/03xcS541na/925_flac
https://megaupload.is/AdxdS34bn9/926_flac
>>
I don’t really know if it would have any use to this project, but have any anons considered going through and ripping all the gasps/laughs/sighs/grunts etc?
There’s one obvious use people would have for those sounds but it could also be useful if the AI voice is unable to replicate those kinds of things.
>>
Another mirror in case it's needed. These are aac and cut to the length of the episodes.

https://drive.google.com/file/d/1qQ9am-cdqmLtnB1rRHhVq8MzD5U44kuS/view
>>
Requested this in the fedorashy thread.
https://vocaroo.com/i/s0f4JOCM7nkw
>>
File: inconceivable.jpg (52 KB, 639x269)
52 KB
52 KB JPG
>>34262781
wat
was this made with one of the models?
>>
>>34263116
>Requested this in the fedorashy thread.
>>
>>34262012
When I clipped an episode, I saved some noises like this, but mostly sound effects.
Some are really funny, other are lewd when out of context. But I don't know if the AI can learn to add them from time to time...
>>
>>34263967
Maybe I’ll try and grab some from a few episodes too. They might still have a use, even if it’s just to use them alongside generated voice clips to add things the AI can’t generate
>>
what are some non pone character you would want to use this tech with?
>>
>>34264419
I'd like a John de Lancie vocoder
>>
>>34264419
Rebecca Shoichet
>>
>>34265144
>>34265139
i said character not VA, for example raven from teen titans, the witch from xiaolin showdown or xj9 from my life as a teenage robot(or star butterfly, lucy loud and spinel(steven uinverse), but i wanted to put up front the high quality waifus since the ones here in the brackets are not exactly that)
also de lancie guy, you should make a list of Q episodes so that you know from where to clip form and then add discord for good measure.
>>
>>34265232
But discord IS my husbando anon
>>
>>34265238
i was going to tell you that i told you to tell me a non mlp character you would like for this tech to be used on
then i reread mi post and i did not ask for that, i asked for a non pony character, you are completely correct on that, discord is a valid choice with the wording from my post
if you now may, could you answer the question again please?
this time with the wording i intended but did not write properly, what non mlp character would you like to use this tech on?
>>
Sorry for the silence. This is the first thing I'm programming in several years, so I needed to learn to use a bunch of tools to make my development flow sane and publishable. I'm writing up documentation and tutorials now to explain how to work with the code. Once I have enough that people can actually run it and have an entry point into the code, I'll share a github link. I expect it'll take 1-2 more days.
>>
File: 1567710517977.png (643 KB, 768x1024)
643 KB
643 KB PNG
>>34260827
It is just a script for creating from Clipper anon audios to dataset. Specific for deepvoice, but i am making improvements for tacatron.
I also managed to run it(tacatron2) inside google colab, but it is still very raw.

Because it's just a script, without any data or even checkpoints - its C&D free.
I think even checkpoints itself are C&D free.

>>34262012
For first datasets i was manually removing all audio clips with this sounds, because they could make learning harder.
But now i cant do this, because it will take huge amount of time. I think it could be good, if someone will do it.

>>34262781
Thats very cute. Is it watermelon fluttershy? Or how was the name of this fluttershy-youtuber.
And i dont know who is fedorashy

>>34263967
You need to share this. Cute noises must be available for all.

>>34265532
Godspeed, anon.
>>
>>34189332
>>34260434
I’ve re-clipped s6e13 - 26, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

>>34227343
>Season five
Have you had the time to process the episodes yet? If not, I could run them through iZotope myself tomorrow. Will also need EQG stuff soon.

>>34264419
You can do pretty much anyone you want, real or make-believe, provided you have enough clean audio. Personally, I would like a model that imitates prominent politicians, for two main reasons. First and foremost, it could make the ongoing clusterfuck known as Brexit even more hilarious than it already is, the potential for stuff like that is limitless. Secondly, having politicians talking about AI as a result of something like that would likely kick start public interest in AI, something that I think is seriously lacking right now. Technology like this has the potential to do all sorts of things, good and bad, and has had nowhere near the attention it deserves.

>>34265532
Good to hear from you again, it’s always nice to see more progress, you’re doing good work.

>>34262012
>>34265676
I haven't been saving any sounds like grunts or sighs, since they don't ever have any words/symbols associated with them, which means an AI would likely have no idea what's going on with the extra sounds. There are only two exceptions to this, inhales at the start of a sentence, as they occur almost all the time and are entirely natural, and when it would be too awkward to clip around something that occurs mid-sentence, which doesn't happen very often.
>>
>>34265734
I have been unfortunately extremely busy this past week (due to many reasons), so have not gotten around to alignment yet. The only thing I've done so far is to organize the languages of the dubs as previously the language tags were inaccurate. The earliest I could feasibly do it would be tomorrow night if you're alright waiting a little longer. I believe all the EQG stuff should be available already, at least on Anonfile. If there's any additional files you need or if you want me to reupload anything to another host, let me know and I can get that setup. Sorry about the delays.
>>
>>34262012
I don't know that we would necessarily need to have the AI handle the extra noises. Since the grunts, clips, and clops are pretty universal I'd imagine we could just have a sample of the various noises and drop them in where needed.
>>
>>34265831
I have all the season five files ready to go on my end, just need to run them through iZotope is all. Since I won't have anything else to do tomorrow, I may as well just run them myself and give you one less thing to worry about.

As for the EQG stuff, Anonfile seems to have sorted itself out now, so I (hopefully) won't need any re-uploads, but I'll let you know if it goes wrong again. Looking in the Google Doc, I see files for the following:

https://anonfile.com/fc39a533n0/EQG_Tales_of_Canterlot_High_7z
https://anonfile.com/rek1a731n5/EQG_Rollercoaster_of_Friendship_7z
https://anonfile.com/V9q2ad37n4/EQG_Rainbow_Rocks_7z
https://anonfile.com/y7jca432nd/EQG_Forgotten_Friendship_7z
https://anonfile.com/J1wae63cnb/EQG_Friendship_Games.7z_001
https://anonfile.com/K2wde53dn8/EQG_Friendship_Games.7z_002
https://anonfile.com/RdMdd831n8/EQG_Friendship_Games.7z_003
https://anonfile.com/45Xcf337nc/EQG_Legend_of_the_Everfree.7z_001
https://anonfile.com/98X8f13cnc/EQG_Legend_of_the_Everfree.7z_002
https://anonfile.com/Oazbf13ena/EQG_Legend_of_the_Everfree.7z_003

Are those all the EGQ dubs there are? And also, which of the three files for Friendship Games and Legend of Everfree should I use? I only need one dub with the music removed for cleaning.
>>
I might go through and extract some of the non verbal sounds myself. Even if they won’t be of much use to the AI training they still might be useful when we’re synthesisng things. I’m guessing there’s no easy way to synthesize a laugh, for example, so having access to a database of them could be useful.
>>
>>34265933
Those are the complete collections of the dubs. You'll need all three to extract them. However I did upload FG and LoE with just a single dub earlier in the thread. I'd recommend going for those instead of having to download the large complete collection. Other EQG stuff use the links in the doc for. Most of the others only had a single 5.1 forign dub, so thats whats in those. Some have additional 2.0 tracks if the 5.1 tracks were lacking. But those don't take up as much space.

Since you've got season 5 under cover, I'll let you go ahead with that. Thanks for all you contributed.
>>
>>34265933
>>34266008
These should be the relevant links for FG and LoE.

>>34229784
>>
>>34265996
If you're going for that, I'd recommend you have the wiki transcripts open in a separate tab, as they have most occurrences of laughs and gasps in square brackets, which should make your searches easier.

https://mlp.fandom.com/wiki/Friendship_is_Magic_animated_media

>>34266008
>>34266032
Thanks, I'll see about downloading those tomorrow. One other thing, are there any dubs for the original EQG movie? I don't see an upload for that and I find it odd that dubs would exist for most of the EQG stuff but not for the original.
>>
>>34266053
Thanks, I was thinking of that. I’ll probably go through the transcripts to find where they occur then rip them from the cleaned audio.
>>
>>34266053
Another Anon brought that up earlier. The original is not available on Netflix for me. Someone else with access will need to rip it. Don't know what regions it'd be available in, but apparently the US is not one.
>>
>>34266078
Strange, but okay I guess. Not much we can do about that until someone else gives it a try. We can worry about that one a bit later, I should have enough material to clean for the next week or so.
>>
>>34265734
>>
bump
>>
>>34268451
Thank you
>>
I've published the repo. If anyone tries this out, let me know if you run into any issues.
https://github.com/synthbot-anon/synthbot

I've also cloned whichever of Clipper's files I've been able to download. Hopefully this will help anyone that's running into issues with Mega and Anonfile.
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S1.zip
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S2.zip
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S3.zip
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S4.7z
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S5.7z
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S6.zip
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S7.zip
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S8.7z
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/MLP+Movie.7z

S1-S3 are at least a month old now. S4-S8 and the MLP Movie are from >>34223803. I've been having issues with both Anonfile and Mega recently.

Clipper, if I can get your email address, I can grant you permission to upload to my S3 bucket. Ping synthbot.anon@gmail.com.
>>
>>34266074
So I'm going to have a go at this today, but I'm wondering if there are any links to the full cleaned episode audio for the episodes that have been processed, as opposed to the individual dialogue clips?
>>
>>34269083
Sure, I can set up a throwaway email when I get back from work this evening. Are you just looking to mirror the stuff in the master file?

>>34269221
Seasons 1 - 4:
https://anonfile.com/A7h6943bnd/S1_zip
https://anonfile.com/Z3J1Qa3dnd/S2_zip
https://anonfile.com/g9Caf244n2/mlp.s02e24_dialog_extra_flac
https://anonfile.com/4cubf740n1/S3_zip
https://anonfile.com/d1acz34anb/S4_zip

Season 6 here >>34245322

Aiming to have season five ready to go this evening.
>>
>>34269261
Great, thanks.
>>
>>34269261
>Are you just looking to mirror the stuff in the master file?
Yes.

Also, some horrific samples:
https://clyp.it/nuhujdyg
https://clyp.it/pet1yzp4
>>
>>34269261
Maybe it's just an issue on my end but I can't get anonfile to work at all. Just says it has 2 hours left and then the download fails after about 10 mins or so.
Don't suppose it would be possible to mirror the season 1 audio on mega or something?
>>
>>34269732
I, and some others, have had similar issues with Anonfile in the past, but it seemed to be working fine for me yesterday. If it’s broken again, that would be very annoying as I haven’t yet downloaded all the EQG material I need.

I should be able to re-upload season one to the master file in a few hours, but I probably won’t be able to keep it there forever as space on my free Mega account is limited.
>>
>>34269732
>>34269869
I'm uploading the cleaned season one audio now, it's the "S1" folder on the front page of the master file.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw
>>
>>34270248
Perfect. Thanks a lot, man!
>>
>>34269083
Sent the ping.
>>
File: 1558444003569.jpg (54 KB, 914x1024)
54 KB
54 KB JPG
>>34251219
Incel Detected
>>
>>34270352
>calling someone incel
>from a mare that stayed canonically barren
>>
>>34270176
I guess I'll do this while I wait for your next gen technology.
>>
>>34270390
Since when does not having children amount to not having sex? Do you not know how lesbians work?
>>
>>34270449
>die childless
>doesn't matter, had sex
Modern women in a nutshell.
Thanks for proving the other Anon's point about sexbots.
>>
>>34270464
What is it with humans and sex anyway, why is it so important?
>>
File: Project Management.png (61 KB, 760x378)
61 KB
61 KB PNG
>>34270518
For some dumb design reason it's also the access point for human's on-board 3D printing replicator.
>>
File: 2032226.png (103 KB, 705x397)
103 KB
103 KB PNG
>>34270464
Holy shit anon, do you have no self awareness you Helen Keller fuck
>>
>>34270547
Cope.
>>
>>34189332
I've run all episodes of season five through iZotope, will start re-clipping tomorrow.

Cleaned English - https://anonfile.com/GeA2e059n6/S5_zip
>>
bump
>>
>>34264419
I hope notjordanpterson gets better audio files so I can write a full debate between him and twilight sparkle on friendship and human nature.
>>
>>34273661
>we could not only make regular episodes, but even PiE episodes with other voice generators
Diplomatic-enjoy-Lyra aggressively flirting with every earth leader, the new Netflix series.
>>
I extracted the audio from Lollipop Chainsaw. The main character is voiced by Tara and sounds very much like Twilight.

There's a few minutes (right now 8, but will probably go down after the unusable onomatopoeia are removed) of clean, already clipped bits of Twilightesque vocals. I haven't transcribed them yet.

The cutscene/story dialogue has background noise/music and is only stereo, so not useful.

It's less than what I hoped to get, but eh.
https://anonfile.com/QdWbge5cn8/lollipop_chainsaw_ts_vocals_zip
>>
>>34274448
Wasn't "other sources than the show" considered a last recourse solution ?
Especially for Twilight who already the highest amount of lines anyway.
Would be a lot more useful to hunt down the other VAs' work (especially since unlike Tara they actually tend to do different voices).
>>
>>34274448
The audio quality is fucking horrible. This shit is absolutely useless.
>>
>>34265831
I've spent a good part of this morning trying to download the EQG stuff, but I seem to be having the same problem as >>34269732 with Anonfile being broken again. Sorry to be annoying but can I ask you to do a re-upload for a dub of each EQG thing you have to Mega? Right now, I don't see any other way of making this download work on my end.
>>
>>34274505
Come on, at least this anon try to help.
You don't need to be so rude.

>>34274566
I would gladly do so, but I don't have any space left on my disk right now...
>>
>>34274566
I'll get that going.
>>
>>34275884
Thank you.
>>
>>34275930
It'll be a while, but here's the Mega folder. Should be able to access each file as soon as it's uploaded.
https://mega.nz/#F!NxwBTYTK!o27hgjmN5tTVVZnE_4ocFw
>>
bump
>>
File: 1563605472146.png (357 KB, 563x672)
357 KB
357 KB PNG
Hey guys I've been under the weather the last couple days and wasn't able to make much progress. I wanted to post again with some results but unfortunately that's isn't the case.

I've looked over everything from >>34258316 and I can definitely help out once I'm feeling better.

I have unfortunately been running into issues with Mega rate limiting me and that's put a huge damper on getting the audio loaded in. Am I being stupid or is there some better way to do this? Everytime I try to download the cleaned voice clips (even specific episodes) I end up getting rate limited and haven't been able to get very much but a few files.
>>
>>34277243
No worries, anon. Thanks for letting us know and it would be great to have you help out again when you’re feeling better
>>
>>34277243
I've never run into such issues with Mega on my end, so the advice I can provide will be somewhat limited. If I assume your problems are caused by Mega intentionally throttling your download speeds, then the best advice I could give would be to create a Mega account, which gives you up to 40Gb daily transfers, and then download via the Mega Sync desktop app. If that doesn't help, then most likely there's either something wrong with your network or with Mega itself.

I've been working with Synthbot today to create an alternative download route via AWS. I'm not exactly sure how sharing from there works but hopefully he will be able to get a link or torrent up and running soon.

Another alternative mirror to the master file can be found here:

http://pubshare.ponemusic.net/Clipper%20Anon%27s%20Master%20File/

I'm not sure how up to date with the cleaned material that one is though, so will need confirmation from whoever made it.
>>
>>34277364
>I'm not sure how up to date with the cleaned material that one is though, so will need confirmation from whoever made it.
I try to keep up, but I don't have a lot of time at the moment, so I might be lagging behind some days.
It should show you the date every file was updated, that way you can check if you're not sure.
>>
>>34277364
Last I checked Mega only gave 5gb per 8hrs.
>>
Backup of Clipper's S1-S4 + S6 cleaned clips with transcripts:
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S1+Cleaned.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S2+Cleaned.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S3+Cleaned.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S4+Cleaned.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S6+Cleaned.zip?torrent

Plus some less-recent file:
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S5.7z?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S7.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S8.7z?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/MLP+Movie.7z?torrent
>>
>>34277243
megabasterd is perfect for this and runs on linux and windows. never had an issue with mega rate limits when i downloaded through that
>>
>>34279238
At least some people can get megabasterd working for them. Always errors out halfway through a download for me even when under the limit.
>>
File: 1567921870150.gif (20 KB, 50x50)
20 KB
20 KB GIF
>>34277243
I can reupload datasets to google drive or yandex. Will this be ok for you?
I really want to see how to improve models.
>>
File: 1567932741419.jpg (661 KB, 2894x4093)
661 KB
661 KB JPG
I've added creation of dataset-files for Tacatron2 into toolbelt.
Be sure that all dataset now will covert audio to 22hz 16bit for dataset. Source audio will stay untouched

I know working on setting up waveglow, and after that will try to train tacatron2+waveglow combo.

Stay awesome and love ponies!
>>
File: 1567933015119.jpg (34 KB, 537x540)
34 KB
34 KB JPG
>>34280298
https://github.com/Twibot-ai/audio_proj_utils
The link, of course!
>>
>>34278138
>>34280298
God job. Thank you!
>>
>>34270248
Would it be possible for you to drop the cleaned unclipped season 6 audio into mega too as you did with season 1? Still working on extracting some of these sounds.
>>
>>34280586
Sure thing, but I’m away from my computer right now and will be about two hours before I can get that started. Will let you know when it’s going.
>>
>>34280617
Great, thanks
>>
>>34280586
Upload started, “S6” folder in the front page of the master file.
>>
>>34279748
Yeah that would work but I was going to try just grabbing these >>34278138
>>
>>34279748
I love this tiny cutie Twilly!
>>
File: Libary trip.jpg (1.06 MB, 798x1371)
1.06 MB
1.06 MB JPG
>>34281622
Buk
>>
>>34189332
>>34265734
I’ve re-clipped s5e1 - 14, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

Raw stereo tracks for season 5 - https://anonfile.com/h4O1gf5dn7/S5_zip

All these used the French dubs.
>>
horse voice plz
>>
File: tenor.gif (1.07 MB, 498x324)
1.07 MB
1.07 MB GIF
Crossboarder here. This is so autistic I'm unironically impressed. I'm rooting for you to complete this. It may be one of the most ambitious 4chan projects to date.
>>
>>34284514
Thanks friend
The last few seasons have been rather disappointing so the longterm goal is to make our own episodes.
>>
File: popcorn sonata color.gif (70 KB, 386x450)
70 KB
70 KB GIF
>>34189328
If any of you guys have the time, can you give brief steps on how this whole process is done. does one guy like edit out their lines and another feeds it to the machine?
>>
>>34284624
I'll try and give a brief outline.

1: First thing was to locate a high quality copy of show audio. It needed to be 5.1 so that we could use a voice extraction technique. This results in a single audio channel with most of the extraneous noises removed.

2: Next was to clip lines and tag audio. This has been by far the largest step so far. Using Audacity and a handful of custom made scripts and tools, we marked start stop points for lines, charecters, and emotion. We also noted whether there was any background noise left in the clip after the extraction process.

3: Further noise cleanup. This is where we're at now. While the original noise cleanup technique is highly effective, it cannot cleanup everything. After much experimentation in thread, we came up with a way to use forign dubs of mlp to remove even more background noises. Identical 5.1 forign dubs had to be located for this to work.

Future: There's been talk of doing phonetic clipping/tagging. A few attempts have been made, but I don't know how much progress has been made on this front. If I understand things properly this may not be a 100% necessary but would allow us to be more accurate in what sounds we want the ai to reproduce.

After that I think the only thing left would be to actually train the ai. We are fortunate that we have a couple of Anons who know what they're doing in this regard. Once done there will be models that anybody could theoretically use to recreate show voices. Efforts will be made to make the ai more accessible to people with a gui of some kind.

We've been at this for a few months now and there are likely a few more to go. It will be incredible to see this project through to the end. We hope that it will open the pathway for more fan content down the road. After this project is done, I think we will see about moving on to another. There's no consensus on what this will be yet, but I'm hoping that the success of this project will encourage participation into the next.
>>
>>34190849
You guys are cool. Don't come to /mlp/ tomorrow.
>>
>>34269083
How long do you think this will take to train? I can probably try running it when I'm at my uni.
>>
>>34284932
Also, I should probably add that I've got a 1070, and an i5-7600k to run it on, so you can probably give me an estimate on how long it will take based on my gpu.
>>
>>34284764
>After this project is done, I think we will see about moving on to another.
As has been suggested before, how about voice morphing instead of text-to-speech?
>>
File: 1536471454817.png (29 KB, 320x240)
29 KB
29 KB PNG
>>34284764
So the real time eater on this is cleaning up the sound files to just get THE voices.
I've heard that the hackerman that delivers us the leaks for all these years found that there are individual audio tracks for each episode. Voices, SFX, music and all. If a person were to download each individual uncompressed piece to an episode, it would be 30+ GB
Maybe get into contact or something
>>
>>34282126
Dope, as usual! Pic related.

>>34284514
Thanks

>>34284985
Arent' the "Ponychan" S9E22-S9E26 this exact dope?
If yes, from what I understand, it's exactly what we need, and could use any high quality leak anyway.
>>
>>34285032
>when you forget your picture
>>
>>34285047
You are fast, but not as fast as me ^:)
>>
>>34285066
<3, and just so you know I've got the fastest damn hands in the wild west, boy.
>>
>>34284985
We've already got what could from the leaks. It would seem unlikely for past episodes to leak in this way. They've probably archived the project files long ago. The upcoming specials on the otherhand? Much more likely.
>>
>>34284932
You can probably run the preprocessing steps in a few hours. After that, it should take a couple minutes at most to load the corpus, then generating speech will be near-instant. The resulting voice is going to be pretty bad since right now it's just doing the bare minimum to get recognizable speech.

Quick status update:

I'm looking for a good way to find segway points in phones so I can string together diphones without introducing stuttering or other artifacts. The usual method is to use the middle of the phone for vowels and to use pitch marks for consonants. The techniques for finding pitch marks (http://festvox.org/bsv/x863.html) are really bad and require a lot of manual effort. It's on the order of ~5% error in the best case where you have EEG recordings alongside the speech. Multiply that by about 950k phones.

Because the more principled methods seem to be failing on this problem, I'm going to try using a neural network to find points where diphones can be stitched together. Step one for that is finding a good representation for the sound.

The best representation I've found so far is based on a realistic model of the cochlea (https://github.com/mrkrd/cochlea), and it takes about 100 seconds of preprocessing for a 0.6 second clip on my i7-5960X (single virtual core). It also generates about 2MB of output for that 0.6 seconds. That means preprocessing would take about 5000 hours CPU time and 300GB of storage for 24 hours of audio. I tried playing around with that to see if I could get reasonable output for less processing power. I can cut both processing time and storage requirements down by 50% at most without messing with the core algorithm.

A much faster alternative (~20 milliseconds for the same 0.6 seconds) is a spectrogram, but the output looks really bad. I suspect the low quality of spectrograms is the reason so many deep neural network speech generators end up overfitting to the few voices in their training set. I'll try it anyway.
>>
When will this be done?
>>
>>34285660
Yes
>>
>>34285660
valve time soon™
>>
>>34285680
>>34285663
Ok
>>
bump
>>
>>34285342
>segway
*segue, it's Italian. Segway is the goofy transport device.
>>
File: cochlea-1.png (40 KB, 611x248)
40 KB
40 KB PNG
>>34286743
Thanks.
>>
File: 1568056537699.png (1.37 MB, 1280x860)
1.37 MB
1.37 MB PNG
>>
>>34284514
It's the key for an infinite number of pony content /projects. I have no idea of how to help but cheers!
>>
File: 1568058930288.png (237 KB, 490x807)
237 KB
237 KB PNG
https://drive.google.com/open?id=1SVmhBF2UHm9LEi2l5nFf6hHAmY2N20DR
Updated folder with some new samples.
I was wondering how good model can be if it will be trained with the same quality as LJspeech model.
Almost all samples are finished properly, without any noises in the end.
The final voice is a bit different, tho. More robotic and cold.
What do you think ?
>>
>>34287303
spooky robo-twilight is spooky. Needs more time in the oven.
>>
>>34287303
The “pretty purple pony princess” one is the best yet imo, maybe a bit robotic like you say but it definitely sounds like twilight.
Which seasons/audio did you use for the dataset?
>>
>>34287303
22 khz sounds like ass
>>
>>34189332
>>34282126
I’ve re-clipped s5e15 - 20, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

Aiming to finish season five tomorrow, then it’s on to EQG.

>>34287303
The old ones contain a lot of garbled noise, which makes most of the words unintelligible, aside from the purple pony princess one. On the new 22k ones, I can make out almost every word without having to look at the filenames, but there is a lot of the static-like noise. I’d say it could make a decent “Broken TwAIlight in need of Anon’s repair” type thing.

Not sure how much that noise will go away with further training, but based on these samples alone, I would say the newer ones are better simply because they sound like actual words. I can also clearly hear Twilight’s voice in there behind all the static, which tells me it’s working well so far.

Are you training these with the older audio or the new cleaned stuff?
>>
>>34287460
Whoo, if I remember well, it's the first time I update twice a day.
>>
>>34288547
Rainbow Roadtrip is an FiM special, not an EqG special.
>>
File: 1568092932147.png (815 KB, 1024x951)
815 KB
815 KB PNG
>>34287402
Only clean 44hz, 32bit.

>>34287460
I guess it is dataset which i've downloaded in august. Should i update it?
It is also dataset with only audio from MLP:FIM, without any EG records.

>>34287410
Yes it is, but only with 22hz model can better form words with adaptive training, i guess.
>>
>>34285342
>few hours
From quickly skimming through the nb, I've got some questions. Why do I have to use specifically mfa? Is it possible to use a different aligner? As well, could you tell me in numbers, how much is a "few hours"?
>>
How would I go about learning what exactly you are doing and how it works? I’m not asking anybody to spend time writing a big explanation or play Q and A with me, just asking for a resource so I can learn myself. This is really interesting.
>>
>>34290306
>>34284764
>>
>>34290332
I’m sorry I should have been more specific. The audio channel isolation, clipping and cataloguing I already understand. It’s the “training the AI” that I would like to do research on.
>>
>>34290358
Why are you asking this question here, and not into google?
>>
>>34290396
I find that asking someone who has working knowledge of a subject tends to work out better than relying on the normienet algorithm.
>>
>>34290358
just read the previous threads
https://desuarchive.org/mlp/thread/33700529/ The voices in this video were generated by an AI model trained using audio samples.
https://desuarchive.org/mlp/thread/33729880/ Pony ML (Thread 2)
https://desuarchive.org/mlp/thread/33745916/ Pony ML (Thread 3)
https://desuarchive.org/mlp/thread/33779583/ Pony Preservation Project (Thread 4)
https://desuarchive.org/mlp/thread/33854142/ Pony Preservation Project (Thread 5)
https://desuarchive.org/mlp/thread/33963949/ Pony Preservation Project (Thread 6 : Can you repeat? edition)
https://desuarchive.org/mlp/thread/34019408/ Pony Preservation Project (Thread 7)
https://desuarchive.org/mlp/thread/34079730/ Nameles
(not this one, it is just the opening post an then it was deleted but hey, he got trips)https://desuarchive.org/mlp/thread/34080777/ Pony Preservation Project (Thread 7)
https://desuarchive.org/mlp/thread/34080783/ Pony Preservation Project (Thread 8)
and the one you are in right now
https://desuarchive.org/mlp/thread/34189328/ Pony Preservation Project (Thread 9)
>>
>>34290427
Thank you, i’ll do that.
>>
>>34288703
Whoo, I must have been super sleepy or in a page10 rush to mix it. Thanks anon, I will correct it at the next update.

>>34290444
Speaking of trips...
>>
>>34290283
You can use another aligner. That'd actually be great if you did, and I'd love to have the notebook/script you used and the output you got so I can cross-check MFA's results. Once I get an end-to-end pipeline working, that's something I plan on doing myself as part of improving data quality throughout the whole process.

The rest of the code doesn't care much what aligner you use. If you can output TextGrid files in the same format and to the same location as MFA, then you won't need to change anything. Otherwise, there are two functions to update, both in speechcorpus.py:
* The first 5 lines of load_utterance(transcriptfn: str, sheaf: SoundSheaf) are for reading the TextGrid that MFA produces to get each phone's start, end and phoneme.
* load_character_corpus(audio_folder: str, transcripts_folder: str) searches for TextGrid files produced by MFA and the corresponding .wav files.

If you do plan to using a different aligner, I can update the code to be more flexible and explicit about its inputs.

It took my computer (i7 5960x, single core since I had issues running multiple instances in parallel) 4 hours to run MFA. I believe it was about 4.5 hours total for all preprocessing steps AFTER the --dry-run completes. It's hard to get an estimate for the --dry-run time since most of that is going to be manual effort fixing data errors (< 20 errors in my case), which depends on which set of files you're using. The runtime for the --dry-run script will vary from 10 seconds to 1 minute, depending on whether the files its reading are cached in memory.
>>
>>34290358
Twibot is using the latest published techniques like Tacotron and Waveglow. Synthbot is working on something custom based on concatenative synthesis.

The best way to get up-to-speed on Twibot's approach is going to be to learn deep learning (TensorFlow tutorials) and to read the Tacotron and WaveGlow papers. The best way to get up-to-speed on Synthbot's approach is to read through Festvox (or similar) documentation.
>>
>>34290550
Thank you that’s the answer I wanted.
>>
>>34290759
You're welcome.

bump
>>
>>34189332
>>34287460
I’ve re-clipped s5e21 - 26, all the relevant files have been updated.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw

And that’s season five done, I’ll upload the updated season five folder to Synthbot’s archive later this evening.

I’m going to mostly take tomorrow off in order to organise the EQG material I have and work out what cleaning I’ll be able to do, and also put together a list of material we still need for further cleaning.

>>34290176
The current EQG stuff uses the same standard as the old FiM stuff, so you should be able to include those with your current setup without issue. If you’re only using the clean audio, then not much would have changed. I’m taking more care and being a lot more strict with the noisy tags this time around to try and maximise the quality of the improved dataset. The effect of including the noisy stuff should now be a lot less bad than it was before, at least for the material for seasons 1 - 5. Honestly, going back through some of these I’m amazed at what I allowed to get away with just the regular noisy tag.
>>
>>34291728
Amazing. You're doing incredible job!
>>
>>34291728
robot twi is, indeed, a robot.
>>
I can't wait to listen Twilight singing Rush's Tom Sawyer
>>
The S5 cleaned audio is up, thanks to Clipper. Here's the current set of torrents to use to get the data.

https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S1+Cleaned.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S2+Cleaned.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S3+Cleaned.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S4+Cleaned.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S5+Cleaned.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S6+Cleaned.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S7.zip?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/S8.7z?torrent
https://elasticbeanstalk-us-east-1-993468875543.s3.amazonaws.com/Clipper's+clips/MLP+Movie.7z?torrent
>>
>>34290487
you should also add Holidays Unwrapped to the eqg square.
>>
>>34293481
I can't wait to listen to Twilight singing Twilight Time.
>>
>>34260513
What happens if they find out? I don't want to see this doe, does the project just move somewhere else?
>>
>>34295752
I feel like it would be impossible for them to really take any legal action against people on an anonymous imageboard. There are no actual names tied to this, not even anywhere like an email to send a C&D to. I think if we keep it fairly low key for now we'll be fine.
>>
>>34295781
I'm just thinking to myself, of COURSE the only thing keeping waifu-bots from becoming a reality is some obese jackass in a suit who wants to rub even more money all over his tiny dick.
>>
>>34295752
If it's out there it's out there.
>>
>>34295752
>What happens if they find out?
As long as we stay like that, nothing.
Everyone is anonymous, the archive are stored on random hosting site.
That said, having a regular "everything included" archive available via torrent or something could be nice.
All that dialogue extraction work is incredibly precious.
>>
>>34291728
After some re-organising, I can give this report on the state of EQG:

I have already clipped the following:
>Original EQG Movie.
>Friendship Games.
>Legend of Everfree.
>Forgotten Friendship.
>A ten-minute segment of Roller Coaster of Friendship (special source).
>Season 2 episodes 4 - 8 from the Better Together shorts (special source).
All of these are currently available in the master file.

I now have material available to clip the following:
>Rainbow Rocks.
>Dance/Mirror/Movie Magic.
>Roller Coaster of Friendship.
I will clip these over the next few days.

I have the dub material to clean the following:
>Rainbow Rocks.
>Friendship Games.
>Legend of Everfree.
Everything else still needs dubbed audio in 5.1 in order to be cleaned.

We still need original English material for the following:
>All of the shorts and the digital series, except s2e4 - 8 from Better Together.
>Spring Breakdown and Sunset's Backstage Pass specials.
>Choose Your Own Ending, both seasons.
The full list can be found on the fandom wiki - https://mlp.fandom.com/wiki/Equestria_Girls_animated_media

To the best of my knowledge, this is the complete list of everything we have and still need to find for EQG. If I’ve forgotten something, be it an entry on the list or some available audio material that I missed, please let me know.

On top of this, we also still need dub material for FiM seasons 7, 8 and 9. The English material we currently have for the first half of season 9 was taken from iTunes, which as far as I know, isn’t as good quality as what should eventually be on Netflix. I don’t how easy it will be to track down the remaining audio, so any suggestions would be greatly appreciated.
>>
>>34296385
I think the easiest place to start looking for the remaining audio would be Netflix. We know that, for some reason, some dubs are available in some countries but not others. I and >>34245322 have already got everything we can, so if any anons who live somewhere other than the US or UK could look on Netflix and see what 5.1 audio you could get, that should be a good start.

Use Flixgrab to download the audio material - https://www.flixgrab.com/

Note that Flixgrab imposes a trial period of a few days, so don’t leave it hanging around for too long. There are supposedly ways around the trial period >>34230157, but I personally haven’t tried any of them.

Again, any suggestions for other places to look would be great.
>>
>>34296392
Only other places I would think to look are foreign streaming sites.
Potential issues:
No 5.1 audio
No matched English track
Getting access outside of wherever it's located

I didn't have much luck previously looking for less official sources, but I'll keep an eye out.
>>
>>34295781
4chan responds to DMCA requests, and if Hasbro really wanted to, they could get a subpoena that would require 4chan to give up any information about the users who posted the content per USC512h. Doubt that would happen though, the worst would be this project being banned from the board.
>>
>>34296479
+if they do get a subpoena, it would likely only include the IP address, so they would then have to go to the ISP to get your information (most ISPs cooperate without even needing an official legal claim though). It would cost too much money for them to do this but it's possible.
>>
>>34296385
I still have the Rainbow Rocks file, and I'm still willing to finish it as soon as I have some time.
Please, could you do it last? This way, Either I can finish it this week, or I will post what I have done so far so you have less work.
>>
page 10 bumparino
>>
>>34296486
this if for shit like this that I regret that 4chan didn't take the kiwifarms approach, where the owner will personally tell people who request IPs and the like to fuck off and that he's not required to do anything
>>
>>34296385
do not forget about the Holidays Unwrapped special.
>>
>>34297174
>Do Rainbow Rocks last.
Sure thing, just post the Audacity labels file when you’re done and I’ll transfer them into the cleaned version I’ll make later.
>>
bumpgh
>>
*wild bumpino appears*
>>
>9 posts to bump limit
Alright time for suggestions to make changes to the OP.

My suggestions:

-Active Tasks: Change "Track down any left behind audio" to "Track down remaining English and Foreign dubs that are missing".
-Latest Developments: Add "Anons are investigating Deepvoice3 and concatenative synthesis for speech generation"
-Voice Samples: Add Twibot's Google Drive of samples https://anonlink.com/2aCEZ
-Add "Synthbot's Torrent Resources" https://anonlink.com/2aCEY

I'll make the next thread in a few hours, hopefully that will be enough time to get some responses.
>>
>>34299347
Good suggestions
>>
Possibly of interested (but probably not): https://open.unmix.app/#/
>>
>>34189332
I’ve run Rainbow Rocks, Friendship Games and Legend of Everfree through iZotope. Cleaned versions below for anyone who wants them.

https://anonfile.com/35m67456n2/Rainbow_Rocks_Cleaned_flac
https://anonfile.com/09mc7d5an8/Friendship_Games_Cleaned_flac
https://anonfile.com/24me7358n5/Legend_of_Everfree_Cleaned_flac

Will resume clipping tomorrow.

>>34300195
Could potentially be interesting. Do you know the capabilities of this? Have you tried it yourself on any of the audio material?
>>
>>34300308
>Could potentially be interesting. Do you know the capabilities of this? Have you tried it yourself on any of the audio material?
I've tried it on songs, but not show material. It seems pretty good on singing.
>>
>>34300341
Could possibly be used to isolate vocals of the songs then, if that's something we decide to do later on. The current method for removing background music sometimes leaves some music untouched, presumably because it was mixed differently in those cases for whatever reason. The "Tornado power intro video" sound files on the front page of the master file are a good example of that.
>>
>>34299347
>6 hours later

NEW THREAD
>>34300569
>>34300569
>>34300569
>>34300569



Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.