[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] [Home]
Board
Settings Home
/mlp/ - Pony

Name
Spoiler?[]
Options
Comment
Verification
4chan Pass users can bypass this verification. [Learn More] [Login]
File[]
  • Please read the Rules and FAQ before posting.
  • There are 82 posters in this thread.

05/04/17New trial board added: /bant/ - International/Random
10/04/16New board for 4chan Pass users: /vip/ - Very Important Posts
06/20/16New 4chan Banner Contest with a chance to win a 4chan Pass! See the contest page for details.
[Hide] [Show All]



File: Triple-P v.png (1.19 MB, 2119x1500)
1.19 MB
1.19 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 favourite ponies.
Videos such as https://www.youtube.com/watch?v=GuJKTodX1FA or https://www.youtube.com/watch?v=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 roughtly 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 you technical level, you can help.
Document : https://derpy.me/uOlX0
Spreadsheet : https://derpy.me/EzAMm

>Active Tasks
https://clyp.it/mmxgq1yx [Do you like it?]
- Create a dataset for speech synthesis (clip and transcribe episodes TlDr : https://www.youtube.com/watch?v=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
-Attempt noise removal

>Latest Developments
https://clyp.it/xp4q1bru [Yay!]
- A group of anons (well mostly one Anon in fact) have completed full clip sets for 8 seasons worth of audio
- Multiple anons have created converters to help import subtitles and dialogues into Audacity
-SAMPA taggers under development by multiple Anons

>Voice samples (So far)
-https://instaud.io/3UAa
-https://instaud.io/3UAb
-https://instaud.io/3UAc
-https://instaud.io/3UAd
-https://instaud.io/3UAi

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

Last Thread: >>33963949
>>
File: ⚓.jpg (16 KB, 303x398)
16 KB
16 KB JPG
>>34019408
Anchor post.
>>
File: 1562487550555.jpg (65 KB, 512x512)
65 KB
65 KB JPG
>>
If you want to help clipping.
>>
>>34019739
I want to help aligning
>>
File: 1562599428797.jpg (42 KB, 589x431)
42 KB
42 KB JPG
>>
Anyone made a discord?
>>
>>34019769
Are you retarded?
>>
>>34019769
why would we need a discord
>>
File: profile_twilight.png (1.2 MB, 1500x2635)
1.2 MB
1.2 MB PNG
Reposting the voice profiles, because they're cool.
>>
File: profile_rarity.png (1.33 MB, 1500x2749)
1.33 MB
1.33 MB PNG
>>
File: profile_rainbowdash.png (1010 KB, 1500x2363)
1010 KB
1010 KB PNG
>>
File: profile_pinkiepie.png (1.59 MB, 1500x3322)
1.59 MB
1.59 MB PNG
>>
File: profile_fluttershy.png (921 KB, 1500x2194)
921 KB
921 KB PNG
>>
>>34019803
>English Trans-Atlantic

I thought it was perceived pronunciation
>>
File: profile_applejack.png (816 KB, 1500x2015)
816 KB
816 KB PNG
>>
>>34019817
No thats octavia.
>>
>>34019769
Literally read the OP
>>
>>34019763
That would be great.

Could you take an episode (maybe S01E01 as a starter?) and try the different tools we have on hand?

Gentle (https://lowerquality.com/gentle/)
Praat (http://www.fon.hum.uva.nl/praat/download_win.html)
Ans obviously, the two tools created by anons.

Don't do the whole episode, just do the first sentences, but try each tools to see which one is the best, and do tell us what you think about each.
>>
>>34019769
>discord
https://files.catbox.moe/3l1thv.webm
>>
>>34019900
I'll try this when I get home. If not today, then tomorrow.
>>
>>34019900
>>34020069
Also
>Ans obviously, the two tools created by anons.
Where can I get those?
>>
>>34020171
As far as I can tell, only one of both can be use, and it's still a work in progress :
https://anonfile.com/H8R039w6n2/Demo_7z
>>
>>34020181
May want to wait a moment on that. I've implemented the rising falling tone thing since that demo. I'll post an update in a few hours.
>>
File: Go and stay go.jpg (56 KB, 480x540)
56 KB
56 KB JPG
>>34019769
>>
>>34020346
Can you post a clip along with a pic or file of what you mean with the rising falling tone thing? I'm advocating for using the fundamental frequency and it might contain the same info.
>>
File: fundamental frq.png (100 KB, 2240x1135)
100 KB
100 KB PNG
>>34020463
>>34020346
Pics the fundamental frequency of the following clip:
https://mega.nz/#!2CAmUQjI!5F_Y874UspboSdvTih4ARVhZq-06bO4EyiC0yQk8_jg

The parts that are left out are silences and fricatives (th/s/sh), If what you mean with tone rising/falling is what I think it means I think you can get it from this graph.
>>
File: 1562691105614.png (150 KB, 791x691)
150 KB
150 KB PNG
Reporting in:
Dataset size does not correlate with minimum loss, but batch_size param doesi am using batch_size=32. I think someone mentioned it before itt, didnt know that i can change this param.
Currently continuing training 2 models with new dataset:
with restore_parts and just from checkpoint.
May be later will create third google account to play more with params of deepvoice.

In thursday i will add s5 to dataset.

Also for anons, who will try install pip packages - last version of dependency package lwh is bugged. Use lwh 1.2 with pip install lws==1.2, so you need to run like this:
pip install lws==1.2
pip install -q -e '.[bin]'
>>
>>34019424
Update on my progress - I’m now up to s1e18 in the normalisation process, and have uploaded today’s progress so far to the master file. I’ve also transferred all the episodes I did previously for seasons 3 - 8, since they won’t need updating. I should be able to do a few more later today as well.

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

With how I’ve reorganised everything, most of the links previously posted won’t work anymore, so I’ll advise you all to go via the new master file if you need to download any sliced episodes.

I mentioned this in the previous thread, but since we’re in a new one I’ll ask again. Does anyone have the Audacity label text file for s1e3? I suspect that the only person who would have it would be whoever clipped it originally, so if you’re still around, it would be very helpful if I could have it.
>>
>>34020463
>>34020493
An Anon in a previous thread recommended it. I'm not super familiar with the workings of SAMPA, but apparently it's a part of it.
>>
>>34020545
> normalisation process
I am not familiar with audio things. What does this do?
>>
>>34020557
Earlier episodes used a different file naming system and synonyms for emotions. I’m going back through them to update them to the new standard so it’s all consistent, and then putting everything in one place for convenience.
>>
>>34020578
oh, got it!
You're amazing, anon.
>>
>>
For now people can try aligning in audacity to see what ts like, doesn't matter if its praat or a custom tool its basically going to work in the same way, and I think audacity is easy to use (easier than praat). Heres an audio clip and a label file: https://www12.zippyshare.com/v/82zJX3bW/file.html

>Open the audio in audacity
>import the label file
>zoom in a bit so the phones dont overlap
>youll see 2 rows of labels, words and phones, thats roughly aligned
>first check the words:
>click on a word label, it goes into edit mode, so tap Esc twice and hit spacebar to play the word (I keep my thumb on space and middle finger on esc)
>listen, make sure the word is properly clipped, if not move the endpoints
>you can drag the play area bigger to make sure you got the whole word without moving the label
>then check the phones:
>same as words, but for 2nd row, a bit more subjective, dont worry about making it 100% correct, should just not be obviously wrong
>some words have missing phones (canterlot), these have to be inputted manually:
>check 2-letter column in https://en.wikipedia.org/wiki/ARPABET for sounds (why arbabet? because thats what gentle aligner uses)
>select a portion of the audio and start typing to add a label
>position them correctly as well
>when done, export the label file again

praat is basically the same, and the custom tools will also be roughly the same process, so people can check this out in the mean time. MEGA is slow as fuck right now, but ill try to upload the pre-aligned label files for s1e1 tomorow so poele can test out some other clips.
>>
File: 1418067864193.png (193 KB, 969x976)
193 KB
193 KB PNG
>>34020495
>but batch_size param does
Oh, ok. Before the next round of training ill change the batch size to 32. Thanks for the heads up anon
>>
I can't wait to port this into a digital assistant. I can just imagine it, coming home from a long day at work.
"Twilight, I'm home."
>"Welcome home Anon. You have three new messages."
"Thanks, Twi."
>"Of course, Anon! I'm happy to help."
"Twilight, what's the melting point of nickel?"
>"The melting point of nickel is approximately 1,455 degrees celsius."

sounds like paradise to me
>>
File: 1419819656614.png (85 KB, 280x280)
85 KB
85 KB PNG
>>34020673
It probably wouldn't be all that hard to do using mycroft since it uses a fork of tacotron it calls mimic2 for speech synthesis https://github.com/MycroftAI/mimic2. The only issue with training with mimic2 at the moment is that the datasets currently released are made for deepvoice, though it probably wouldn't be too hard to whip up a python script to convert to the proper format.
>>
File: 1562697128770.png (552 KB, 1024x715)
552 KB
552 KB PNG
>>34020783
Datasets are compatible with tacatron aswell.
I wonder if we could just replace mimic2 model with ours.
>>
>>34020803
Are you sure it's compatible? Checking the readme for mimic2 there doesn't seem to be a way to set up a json dataset, but maybe I'm just retarded and missed it.

>I wonder if we could just replace mimic2 model with ours
It would be fucking great if we could, it would basically mean that assistant twiggles isn't too far off
>>
File: 1562697849573.png (488 KB, 811x951)
488 KB
488 KB PNG
>>34020837
https://github.com/r9y9/deepvoice3_pytorch

>Building your own dataset, with metadata in JSON format (compatible with carpedm20/multi-speaker-tacotron-tensorflow) is currently supported.
This is basicly how i create all datasets. If understand this correct it means that this json is compatible with tacatron
>>
>>34020171
>>34020181
>>34020346
Here you go:
https://anonfile.com/703bM2yen1/SAMPA_Tagger_Demo_2_7z

JSON export still not done yet. Window snap/maximize bug still present (Resize by window corner to fix). Still only two episodes included. Give it a try and let me know what you think.
>>
>>34020952
Also, click on the sliders to change the tone to rising/falling/continuous.
>>
File: 1417299065793.png (157 KB, 871x1017)
157 KB
157 KB PNG
>>34020862
Ok, well as far as I can tell that specific line of tacotron starting with GSByeon's version added the json dataset support. Mimic2 doesn't have any reference to json based datasets anywhere in its documentation, so the dataset would need to be formatted the way ljspeech is which isn't too bad. Here is a quick conversion if anyone wants to try and train a mimic2 model https://anonfile.com/Q59eM6y7ne/metadata_csv
>>
>>34019408
>https://instaud.io/3UAa
I still cant get over how great this is , keep up the works anons.
>>
File: FaceGan.jpg (72 KB, 1200x486)
72 KB
72 KB JPG
For anyone who's worried about quality or if characters with a small amount of audio will ever work. Here's an image representing the progress made in AI for face generation.

It would not surprise me if 4 years from now you could take 1 minute of audio of someone speaking and have a better model than the one that produced the steamed hams video.
>>
>>34021463
scary stuff dude
>>
File: MayTheRideNeverEnd.gif (184 KB, 1200x568)
184 KB
184 KB GIF
>>34021470
>tfw people won't be joking when they ask if they trust you or your lying eyes and ears
at least I'll be able to have my waifu say "I love you"
>>
File: sfg.png (152 KB, 800x943)
152 KB
152 KB PNG
>>34021479
Maybe anon, in the future, you'll be able to respond.
>>
>>34019408
I can see it know
BREAKING NEWS HORSEFUCKERS STUMP SCIENTISTS BY MAKING AI WAFIUS
>>
>>34021488
I dream of the day I can respond to my purple princess. I dream...but for now, once this is over? I think I might finally be able to cast a smile unlike any other. You lot are some utter maniacs, I think I speak for most of the community here when I say you are doing literal work of the gods. Carry on mate, carry on.>>34019408 and co.
>>
>>34021581
We are our own Gods now.
>>
>>34021591
not quite gods, but close. bit too much hubris in claiming godhood for my liking
>>
>>34021479
>>tfw people won't be joking when they ask if they trust you or your lying eyes and ears
People already accept obvious bullshit from the media, and even the stuff that somehow become public knowledge while it shouldn't have get forgotten in a few weeks.
It will just make it even easier to bury embarrassing crimes evidences under "it's all deep fakes and conspiracy theories !".
>>
>>34021600
>bit too much hubris in claiming godhood for my liking
Maximazing hubris and getting away with it is how you know you ascended tho.
>>
File: 1418553786696.png (88 KB, 244x244)
88 KB
88 KB PNG
>>34021046
Update to this: I got the reformatted dataset to work with mimic2. If anyone else wants to try setting it up, don't bother using the docker version of mimic2 its complete trash. Make a virtualenv and set it up following the directions in the readme, just make sure to update librosa to its latest version since the version listed in requirements.txt causes the preprocessing step to fail. hparams.py also needs to be slightly edited, change the sample rate to 48000 and num_freq to 2049.
>>
I can't wait for when we get a proper Rarity. Like, Twiggles is easy cause of the amount of lines but also because Tara often does voices that sound similar, so you could just splice stuff together, but Rarity...Rarity so far has been impossible to replicate in any way or form.
>>
>>34021778
>it work
>just don't do anything like what is in the "ready-to-use" package
>and don't do anything like the fucking manual tell you to
Python in a nutshell.
>>
>>34021895
Don't blame a language for half-ass coding.
>>
>>34019771
>>34019772
>>34020376
Based. Fuck discord
>>
>>34022192
I want to fuck Discord.
>>
>>34022226
Calm down now, Flutters.
>>
>>34022110
This.
>>
Something someone can do thats bored is to take all the label files and make a table of all the unique words in it along with how many times they appear. We can do a couple of things with it:
>Check words that only appear a couple of times to make sure they arnt spelled incorrectly (I found 'their' spelled as 'thier' in s1e3 for example)
>Get all non dictionary words and give them a standard IPA/Sampa representation for alignment (like names)
>just cool stats (like which pony name is said the most/least)

Should be pretty easy with a python script or something.

>>34022110
I'll blame a language if it doesn't have built in protection against shit code and a shit package manager.

>>34020952
I'll check it out.
>>
>>34023211
>S1e3
I mean s1e1
>>
>>34023211
Good thinking. I guess i can do this for Twilight dataset.
>>
>>34023211
>if it doesn't have built in protection against shit code
That kind of thing is costly, and it's just encouraging people to hire shitty coders.
>>
>>34019424
Normalisation for season one is now done, except for episode three, still waiting to see if the label file for that one turns up. Will make a start on season 2 today.

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

>>34023211
>>34023214
Fixed, thanks for pointing that out. Let me know if you see any more errors like that.
>>
Changed the tick for noise cleaning, and added Teal for rework.
I will do something similar for phoneme decoding and alignment for Phase 3.

>>34023559
Anon, you are really doing a great job.

As for 0103, as said in >>34018640 , I started to rework it, but I didn't finished it.
So, here you have a label file with each sample at the correct rounded second and with a length of 0.5 second. Basically, I just recovered from filename, no further work. It's all I have.
https://pastebin.com/1hRf8JEY
>>
>>34023578
>https://pastebin.com/1hRf8JEY
Yeah, I remember that one. I'm keeping it in mind as a backup option if the original doesn't turn up, since it's much better than working from nothing at all. What I'm really after is a set of labels all positioned correctly, so all I then have to do is rework emotions in notepad++, and then the noisy/very noisy labels by listening to all the clips again.
>>
>>34023211
That's a good idea, and I may try to do something with that. Is there a simple way of downloading just the label files? Or do I have to go through the master file and select them manually?
>>
>>34020952
I'd like to help out with this but is there a tutorial anywhere on how to use it for dumb dumbs like myself?
>>
>>34023747
>just label files
At the moment, no. You'll have to go through it manually. The current standard is to name all label files "labels.txt" inside their associated episode file, so there would be no way to tell them apart if I put them all together. If it would be helpful, I could later create a duplicate folder with all the label files together renamed something like "s1e1labels.txt".
>>
>>34023307
>>it's just encouraging people to hire shitty coders.
IT is 99% pajeets nowadays. You are way late to care about that.
>>
>>34019408
I want to see this project fused with that one mad lad making a real sweetiebot

Imagine a replica ai robot of your waifu you can hold in your arms
>>
>>34023766
>If it would be helpful, I could later create a duplicate folder with all the label files together renamed something like "s1e1labels.txt".
That's what I'm doing right now. I'm downloading all the label files and calling them 0101.txt, 0102.txt...
I think it might be helpful for anons who want to do text analysis and stuff, but don't want to bother downloading over 10 GB of FLACs.
If you'd rather have a different naming conventions, please tell me and I'll modify them accordingly. I'll post a mega link once I'm done (and you can add it to the master file later).
>>
>>34023774
That standard should be fine, will just have to be careful later on to separate by category when we have more stuff from just the normal episodes, such as Equestria Girls. Thanks for doing this for us.
>>
>>34023771
> that one mad lad
The one that has been milking his patreon for 10 years and is still at the "Lego : My First Robot" stage, outside of making a custom 3D printed shell that help in nothing for the actual robot thing ?
>>
>>34023774
I'd say make them fims01e01, or fim_s01e01, for when other stuff gets added.
>>
>>34023780
>>34023792
There we go, I've just put them all in one big folder for maximum scripting convenience. I've called them "fim_sNNeMM.txt".

https://mega.nz/#F!wWhDGCLB!Q0o1yDS-S0Hjc2hmWX4flQ

Beware that these are only the labels.txt that I could find in the master file, therefore some (including the entirety of S2) are still missing. I plan to keep adding them as they pop up in the master directory.
>>
>>34023753
Not yet, as it isn't fully done yet. But I'll try and give a short write up.

You first pick an episode you want to work on from the first screen (can be returned to by clicking the episode button). This is also the screen you would save and load from once you've opened an episode. On the tagging screen, you have the line drop down on the left to pick what line you want to work on. Can also use the next and previous buttons to navigate lines. Play/stop should be self explanitory. You can also drag and drop the audio position marker to where you want it to be. The plus/minus buttons are to zoom in and out on the graph. When zoomed in, can drag the graph to move around. To start tagging, use the onscreen keyboard. Markers will appear below the graph. To move them around, drag and drop. To hear a preview of your selection, click the text next to the marker you want to hear. To change the symbol to rising/falling, click the marker (without dragging) and it should change into an arrow to indicate this. Can change the tagging language from SAMPA into IPA or Spelling (words with the sounds meant to be indicated by the symbols marked). What language you tag in will not effect the end result, is stored in SAMPA regardless.

If you have quedtions/suggestions, let me know. In the mean time I will continue to work on the tagger.
>>
>>34023806
Same anon. I've run a quick spell-check on the whole thing. I've found some errors which should be fixed:
1) In s04e01, in the line "But what if something else came up when we are gone?", the second "a" is not a regular ASCII character but rather a Unicode CYRILLIC SMALL LETTER A. This may cause problems, so change it to a regular "a".
2) In s04e04: same thing as above, in the line "a secret!".
3) Also in s04e04: "ingredient" is misspelled ("ingridient").
4) In s05e12: in the middle of the line "it didn't work out then it isn't gonna work out now" there is a REPLACEMENT CHARACTER which doesn't belong there.

There definitely are more typos, I've just reported the ones related to encoding problems (also, note that there are more UTF-8 characters in there, mainly French or Spanish words such as "pièce de résistance", "voilà", "piñata", "cute-ceañera", etc.).

Finally, since some anon asked, here are some rough figures for the frequency of the main cast's names (not including a very small number of misspelled occurrences).

>book horse
746 twilight
78 sparkle
72 twilight's
4 sparkle's
3 sparkles
>pink horse
529 pinkie
404 pie
68 pies
18 pinkie's
15 pie's
7 pinkies
>fast horse
504 rainbow
434 dash
30 dash's
8 dashie
8 rainbows
2 rainbow's
>gem horse
442 rarity
34 rarity's
>apple horse
375 applejack
25 applejack's
10 aj
>shy horse
356 fluttershy
26 fluttershy's

You may find the file with all the individual words and a file with the number of occurrences (as well as a super elementary Python script used to generate them) in this folder: https://mega.nz/#F!YPYQWKZA!CVfcUq438QNeIC9ooIMMaA
>>
Season two didn't take as long as I thought it would, I've just finished updating it and am uploading now. From here, I have the following episodes still on my to-do list - s1e3, s3e1, 5, 6 and s4e15. I should be able to finish them all tomorrow, and that should then be all of them.

>>34023914
I'm not surprised to find out there are some typos, but I am happy to see that there are only four significant ones out of what must be a few tens of thousands of files by now. I'm going out now for most of this evening, and will see to fixing those four when I come back. If you find any more, do let me know. The checking script should ensure that any typos are only found in the dialogue section. We were previously aware that words with accents like "café" would occasionally be in the dialogue, and I've been including then under the assumption that it wouldn't cause encoding problems.
>>
>>34023774
>>34023792
>>34023806
Too bad, I use MLP0101 for notation scheme
like "MLP000100031572.flac" is Twilight Sparkle, Excited, saying "Smile" (even if noisy) in the film.
or "MLP081000095888.flac" is Discord, Neutral, saying "Class is now in session." in ep 10 season 8
or "EQG000200180962.flac" is Fluttershy, Neutral, saying "But that was when Twilight was here." in Rainbow Rocks.
Mhe, I will keep my notation and convert if needed when I publish.
>>
>>34020673
>>34020783
>>34020803
>>34020837

I am not kidding, when I would THROW money at a Rarity personal assistant app.
Or even better, Desktop ponies, but with desktop voices.
The ride truly never ends.
>>
File: 1562787564400.jpg (138 KB, 614x406)
138 KB
138 KB JPG
Small update is here.
I managed to drop loss from 0.168 to 0.159.
I also checked deepvoice science paper didnt understand nothing and did >>34023914 after word counting i've found that current dataset has some mistakes and deleted words from quotes.
Mostly they are emotion-words. Also some typos.
This mistakes probably degrading our model.
Just for encourage anons, dropping current progress of ~92k iterations
https://instaud.io/3VIy
https://instaud.io/3VIz
https://instaud.io/3VIB
>>
>>34023914
I've deleted and retyped all the "a" characters in 1 and 2 and corrected the typos in 3 and 4, so those should be fine now. Again, do let me know if you find any others. Will also add the labels file when all the episodes are done.

>>34024358
>deleted words from quotes
>emotion-words
>Also some typos
With the normalisation I've done, the vast majority of those should now (hopefully) be fixed. If you think it's degrading the model, I'd advise redoing the dataset with the new stuff in the master file.

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

I should have the whole thing competed tomorrow.
>>
>>34024011
This. Just this.

It could just be a modification Alexa or Google assistant.
>>
>>34024768
Probably more likely to be hooked up to Mycroft, Google and Amazon would be leery of this.
>>
>>34024358
Sounding good!
>>
Bump.
>>
So what kind of machine learning is this using, exactly? Is it an adversarial network architecture, or what?
>>
>>34025819
Text -> feature extraction (depends on which one, I think some use LSTM layers, others have things like dictionary lookups) -> attention mechanism -> spectrogram -> dilated convoluted layers -> audio
>>
>>34025888
How does the training work? I know GANs can help solve the issue of blurriness in image synthesis, so if that's not something that's already being used here, maybe it could help? Comparing the samples that have been generated from pre-existing lines and the ones for new lines, it seems like there might also be some overfitting going on, but that's just a guess.
>>
>>34025958
I think deepvoice backpropegates the whole thing, but tacotron is 2 steps, it trains the first part to predict spectrograms from text, and the second part to generate audio from the spectrogram. I've seen this modified into a GAN, but I didn't find any code that works as easy as deepvoice does. And it should be overfitting right now, were using like 1/3 of the minimum audio other projects use.
>>
Just let it go guys.
>>
File: becoming an adult.jpg (60 KB, 600x800)
60 KB
60 KB JPG
>>34025973
Imagine being so defeated by life you go into a thread full of people doing cool stuff for fun and tell them to stop because it reminds you of what optimism felt like.
>>
What the hell? This is still going on here? I though this was dead because the discord I made back in the 2nd thread only has 2 active people left, and were not even done with S1, and you guys already have samples? I've spend fucking days working on this and you guys are just going to ignore it and do your own thing, why? Out of spite?

>>34024358
>>34023559
Please join us, it's a lot better than trying to coordinate in a thread.
>>
File: lWcF9GAfXfA.jpg (87 KB, 975x1000)
87 KB
87 KB JPG
>>34025996
And there we go again with the Discord fags. Please stop. Not gonna happen.
>>
File: 1562609978260.jpg (32 KB, 480x480)
32 KB
32 KB JPG
>>34025996
>see /mlp/ project
>stop going on /mlp/
>"wtfff why didn't I see any of the progress"
>>
File: Shocking realization.gif (2 MB, 240x180)
2 MB
2 MB GIF
>>34026012
Discord fags in a nutshell
>>
>>34025996
>>34019932
>>
File: Here We Go Again.png (663 KB, 640x621)
663 KB
663 KB PNG
>>34025996
>>
File: Mainman.png (78 KB, 803x796)
78 KB
78 KB PNG
>>34025996
YOU AGAIN!? I thought you were already told to get the fuck out. We have already discussed about discord quite a bit, we DO NOT want it and you need to understand that nobody is going to cite your stupid name on any credits page, and that if "people will know who [you are]", it will be for being such a fucking fag.
>>
File: Mainman(pony).png (146 KB, 1412x875)
146 KB
146 KB PNG
>>34026074

>>34025996 (you)
>>
File: 412937.gif (820 KB, 640x480)
820 KB
820 KB GIF
>>34025996
>>
>>34025996
Absolutely we're going to do it out of spite, because the collective work done here is far better and more significant than anything your pathetic grandstanding could produce. We have working samples of synthesized Twilight and you couldn't even get through a single season of audio, with all of your simpering dickwaving about MUH CREDIT. Discord faggots OUT.
>>
>>34025996
>providing the perfect example of why discord is fucking shit for this kind of project.
Thanks guy, we'll be sure to credit you as that one dumbass.
>>
>>34025996
Don't forget to zoom my words please
>>
>>34025996
>What the hell? People who said they didn't want to join my discord didn't join my discord? They made actual progress where they said they would?? Not where I wanted them to???? wtf??????
discord fags get out
>>
Guys why are we feeding him so many (You)s
If I knew it was this easy I'd have started creating random Discord servers for threads ages ago. He's got 11 (You)s and counting while contributing absolutely nothing except his ego. Most clipping fags get like 2-3 (You)s!
>>
>>34026254
https://instaud.io/3VSm
>>
>>34024358
>https://instaud.io/3VIB
She's almost here
>>
Some chinese dude got a very reasonably sounding GLaDOS with tacotron after 200k iterations, but I think the catch here is that the dataset wasn't even human generated - it was processed audio from a TTS
https://www.youtube.com/watch?v=2Hw5gCt6ItU
https://github.com/kairess/tacotron
>>
>>34019424
Normalisation for seasons 1 - 8 is now finished. Everything has now been run through the checking script, and so should therefore all use correct and valid tags for character, emotion, and noise level, and all be named according to the current standard as detailed in >>34019739.

For clarity, the emotion set used is as follows: Neutral, Happy, Amused, Sad, Annoyed, Angry, Disgust, Sarcastic, Smug, Fear, Anxious, Confused, Surprised, Tired, Whispering, Shouting, Whining, Crazy and Canterlot Voice.

The following episodes have been modified: All of season one and two, s3e1, 5 and 6, and s4e15. If you previously downloaded any of those, you should now download the updated version in the master file. All other episodes remain unchanged.

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

The master file also contains the audio from the mobile game, the character voice profiles, and the AI pony pictures from our wonderful drawfags. I’ve also added a separate folder with just the label files. If there’s anything else you think should be added to the file, or if you notice any missing files, please do let me know. I also can’t guarantee that the dialogue section is free of typos, as seen in >>34023914, so be on the lookout for those as well.

My question to you all now is - what would be most useful to work on next? I could move on to what we have for season 9, although the high-quality clips aren’t available yet, so the best I would be able to do is clip them and only post the label files to be used later for exporting. Or I could clip something else, in which case, what would be best?
>>
>>34026484
Nice
>>
>>34026484
Is there anything else aside from the episodes of the show we haven’t clipped yet? Like EQG or the specials/shorts? Maybe it would be best to get those done first so we have all currently available sources clipped.
I assume there isn’t a good enough version of Rainbow Roadtrip yet.
>>
>>34026508
According to >>34023578, there are several options available. My main point of thought is which would be the most useful to start with. While we're at it, I'd also suggest it would be a good idea to start putting more effort into completing the list of audio rips. If someone could also post links for other stuff that's been done already, I can add it to the file.
>>
Reposting the link to my upload of center channels for EQG movies and other specials extracted from the lossless BluRay or HQ Netflix sources in case if you're going to clip it.
https://mega.nz/#F!hIgR2STC!3_1uh5qHcaxk_X5JEe115g
>>
>>34026653
>in case if you're going to clip it.
Chances are I will at some point. Are there any subtitle SRT files to go with these, or I am I going to have to go hunting for them?
>>
>>34026676
Sure, I'll upload them o the same folder.
>>
So the endgame for this is to have a simple text to speech program where you click a button for the emotion and then let it do its thing, right?
>>
>>34026691
Thanks.
>>
>>34026695
Yes, my friend
>>
>>34026695
Yes, at minimum click which pony and emotion and it gives the audio, but I'd like to go even further and have optional settings to change the duration and pitch of each phone, and let the emotion change, starting happy and ending sad for example with a smooth transition. Another thing is that well most likely use a multi speaker model and they use an embedding vector for the speaker, you should be able to generate mixtures of ponies, so a voice that's 50% AJ and 50% rarity for example, which would be good for OCs.
>>
>>34026776
Very ambitious. I'll be happy when we have a working Rarity, she's been impossible to replicate by fan VA's so far.
>>
>>34026785
I’m guessing it’s the accent. It’s kinda hard to nail that transatlantic accent just in general. Hopefully the AI will manage to replicate it.
>>
>>34025996
Lol, imagine being this desperate for attention
>>
>>34026695
>>34026776

sounds awesome my dudes. even if I wanted to, I wouldn't be able to bring myself to stop following this project. y'all are amazing
>>
>>34026822
What? That's probably the easiest part, it's the pitch and emotions that no one can get right.
>>
File: 1562870455560.gif (143 KB, 600x520)
143 KB
143 KB GIF
All our audio is 48khz sample rate.
h_params are tuned to 22050hz.
There're also such params as fmin, fmax, as i asuming is min and max frequences. For 22050 they are 125 and 7600.
So question to audio fags:
Is it better to convert all audio to 22050hz or choose fmin and fmax for 48k ? What numbers should i choose?
I am a complete dumbass in audio technologies, but this will probably enhance our model and could remove noiseness.
I suppose that fmin should the lower frequence of voice and f max the higher. Keep in mind that we have like 50-55% of noised material.
May be i should start from searching how to clear our audio?
>>
>>34027260
22050 is definitely not enough, it sounds like absolute low quality shit. Can you use 44.1 or 33 kHz?
>>
File: 1562871089084.png (444 KB, 1200x850)
444 KB
444 KB PNG
>>34027273
Anon
> All our audio is 48khz sample rate
I am already using this, but i assuming that i need to tune our params of training, because they're for 22k.
Did i write something wrong, so you didnt understand me?
>>
>>34027260
I think mins and maxes really will depend on the voice, for example pinkie and rarity would probably need a higher max than applejack. We 100% need to fine tune the training params
>>
>>34023578
What do those blue squares in S2 mean?
>>
>>34027376
rework
>>
>>34019408
I can't wait until I can make my waifu say dirty things
>>
File: she will be real soon.png (724 KB, 2400x1800)
724 KB
724 KB PNG
>>34027260
Those hparams are given to the librosa library. If you check in audio.py in the _build_mel_basis() function, you'll see that hparams.{sample_rate,fft_size,fmin,fmax,num_mels} get fed directly to librosa.filters.mel().

This is entirely a guess, but the librosa documentation says that the default fmax is half the sample rate, so I assume fmax is supposed to be the maximum frequency used in the audio clips. In the case of S1, this never goes over 20000 and for S2, 20500, so I'd pick those as experimental fmax values if you want to experiment. For now, I wouldn't touch fmin.
>>
>>34027260
>>34028728

I absolutely wouldn't convert the audio to 22khz, though. The Nyquist sampling theorem means a 48khz sampling rate can only represent audio up to 24khz, and a 22khz sampling rate could only represent audio up to 11khz. 11khz sounds like garbage. Even shitty MP3s only cut off sounds above 16khz.
>>
File: 1562907781370.png (143 KB, 450x549)
143 KB
143 KB PNG
>>34028728
> This is entirely a guess
Yeah, i was guessing this too. I was interested if there're well known numbers for 44khz. I guess i could experiment one day, this doesnt hurt.
Model is learning, but the learning rate is around 1e-5

>pic names
Yes, this basicly the reason why i am doing this. This and also fun with science

>>34028731
Roger, that. Will not convert samples.
(Librosa still can do it, tho)
>>
may i request that the links to the previous threads archives be put on a permanent pastebin or the gdoc to save space on the op?
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)
>>
>>34022428
Fuck you. I'll fuck discord before she even meets him
>>
>>34028778
We're going to have to experiment in general. Hyperparameter selection is more trial and error than a science. When we have closer to a final dataset, we should try to get as many anons as possible training with slightly tweaked parameters and compare models.
>>
>>34026131
>Homer is behind 80% of all the drama on the board
put me in the next pts video please
>>
>>34026776
This makes my penis the big penis.
>>
>>34028824
Stuck them in the Google Doc above the resources. Is the manager of the doc still around anymore? Could use some updates. OP is pretty much at character limit.
>>
>>34027260
>>34027273
>>34027287
Before anyone messes with this sample rate shit, you need to know that when something says "Sample rate" and a number such as 44,100hz, it is saying that is how many samples the computer reads per second from a sound file. This is different from "sound frequency". The highest sound frequency that a sample rate of 44,100hz can produce is a 22,050hz tone. The same goes for a 48,000hz sample rate which can produce a sound frequency up to 24,000hz.

The highest frequency most adults can hear is up to 17,000hz, but some people (almost only kids) up to 20,000hz.
Lower sample rates don't distort the sound in a way that lossy compressed files do, rather, when the sample rate is low enough that you can hear it, it sounds muffled, like a low pass filter has been applied to the sound.
>>
>>34025996
>not using the trip
this is bait
>>
>>34026484
I've decided I'm going to make a start on the original Equestria Girls movie. While testing things this morning, I noticed that the SRT to Audacity converter and character tagger tools can't seem to handle anything over an hour long. I'm pretty sure this has been mentioned in the past, but now that more and more of the remaining material is over an hour long, I think it's becoming a bigger issue than it was previously. Is there any way that these tools can be updated to fix this?

Right now, I'm using the raw Audacity labels linked in the Google Doc, which are actually pretty helpful - https://u.smutty.horse/lpkoajzippt.zip
>>
>>34029298
Forgot to mention - the audio rips for the EQG stuff still has all the background music in them, unlike the FIM episodes, which means that pretty much every single clip is going to have significant noise. How big a problem is that? Is it possible to get rid of the background music as was done for the FIM stuff?
>>
>>34029310
Assuming the eqg stuff has a similar audio mix to fim, the same technique to clean up the audio should be able to be applied. Maybe tomorrow I'll see about tracking down/processing the audio if nobody else does it first. The process is linked in the doc if anyone else wants to give it a go. As for the srt converter, I know I've said I'd fix it a long time ago. As of late I've been putting my time towards the SAMPA Tagger otherwise it'd probably have been fixed by now. Maybe I'll see if I can get the converter fixed tomorrow as well.
>>
>>34029298
Also , even though the SRT Tagger doesn't work with stuff over an hour, I believe the character tagger shouldn't have any issues as long as it gets a proper Audacity label.
>>
>>34029338
>>34029340
All right then, I guess I'll hold off EQG for now and do season 9 instead. The character tagger seemed to work fine for EQG.
>>
>>34027498
Did something happen to them?
>>
>>34029603
Older episodes used different file naming conventions to the current standard. Rework means basically updating them to the new standard. see >>34026484
>>
Totally worth a watch https://www.youtube.com/watch?v=6bFN2YkN6bo
>>
>>34027376
Pic updated. as >>34027498 said, it was rework. But clipanon works too fast for me to update.

>>34029350
Great.
I'm still working on EQGRR even if I did't touch it for a week.
It will be done by this week end tho.
>>
>>34019424
I’ve finished the first two episodes of season 9. As before, we only have the iTunes rips right now, which (as far as I’m aware) aren’t as good quality as the Netflix rips used for seasons 1 - 8. For now, I’ve just put the Audacity labels in the labels folder, but I could export the audio and transcripts as well if you guys want. Let me know. Will also do more season 9 tomorrow.

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

>>34029972
>11:50, “I can’t think of a more painful thing than manually transcribing audio.”
Pure truth right there, thank God for all the tools we have for doing it automatically. The results he got also sounded pretty promising if we could do the same with the Twilight clips.

>>34030894
>All that wonderful green.
Truly a wonderful thing. Keep up your good work.
>>
Here is the processed audio for EQG:
https://anonfile.com/FdJ3obz5n3/EQG_Processed_flac
https://anonfile.com/idJ9o5zanf/EQG_Extras_RAW_7z
https://anonfile.com/i1K1o8zbnd/EQG_RR_Processed_flac
https://anonfile.com/dfJ3odz1n5/EQG_RR_Shorts_RAW_7z
https://anonfile.com/X2M0o2zdn5/EQG_LoE_Processed_flac
https://anonfile.com/X2M0o2zdn5/EQG_LoE_Processed_flac
https://anonfile.com/j5K0odz8na/EQG_MMN_Processed_flac

Those labeled processed have undergone the treatment to remove background noise. Those labeled as RAW have the original audio. Those were only stereo and could not be processed in the same way. All audio sources from yayponies. BluRay remuxes used for all except LoE and MMN as these were not available as remuxes. LoE audio is from the BluRay rip and MMN audio is from the "Scootaloo Remastered". Will post FG audio once I can finish downloading the remux. Links added to document.

I have also got the SRT converter figured out and will update it as soon as I can figure out what's up with my java install.
>>
>>34031184
Also, before using the unprocessed shorts audio here, check the doc to see if it's available in the shorts special source. Should have cleaner audio.
>>
>>34031184
Why haven't you use these? >>34026653
>Scootaloo Remastered
Fucking really? That shit was probably compressed as fuck.
>>
>>34031196
It's all they had. I can look for a better source later, but should be serviceable for clipping at the very least.

>>34031184
Goofed on the second LoE link. Meant to be:
https://anonfile.com/B4G5o7z4ne/EQG_LoE_Bloopers_Processed_flac
>>
>>34031184
Thanks for your work anon!
It will be handy.
>>
>>34031196
Didn't use those because he only uploaded the center channels. Need 5.1 audio to do the processing.
>>
Here is the audio for EGF FG:
https://anonfile.com/Sbt5paz1n2/EQG_FG_Processed_flac
https://anonfile.com/Mfq6p5z0nf/EQG_FG_Deleted_Scenes_RAW_flac
https://anonfile.com/K0r4pfzen6/EQG_FG_Shorts_7z

Same deal as before. Added to the doc.

The SRT converter tool has been updated. Theoretically it should be capable of SRT files up to 100 hours long or up to 10,000 subtitle entries. Whichever comes first.
>>
>>34031184
>>34031218
>>34031588
you are a gud boi anon-sama
>>
A bit late to the party. How hard is this on the computer? Can an old potato run the learning software or does it have to be semi-recent hardware?
>>
Time for a far larger than necessary progress update:

A. K. Yearling - 00:03:37
Apple Bloom - 01:05:35
Applejack - 02:31:46
Autumn Blaze - 00:04:58
Babs Seed - 00:01:30
Big Daddy McColt - 00:01:37
Big Mac - 00:09:13
Bow Hothoof - 00:01:41
Caballeron - 00:02:52
Cadence - 00:10:34
Capper - 00:05:01
Captain Celaeno - 00:01:13
Celestia - 00:35:30
Chancellor Neighsay - 00:05:57
Cheerlee - 00:08:33
Cheese Sandwich - 00:03:08
Cherry Jubilee - 00:01:06
Chrysalis - 00:08:10
Coco Pommel - 00:01:42
Coriander Cumin - 00:01:20
Countess Coloratura - 00:06:46
Cozy Glow - 00:09:10
Cranky - 00:03:07
Daring Do - 00:01:09
Diamond Tiara - 00:12:51
Discord - 00:52:42
Double Diamond - 00:01:36
Dr. Hooves - 00:01:48
Dragon Lord Torch - 00:02:02
Ember - 00:05:02
Fancy Pants - 00:02:11
Filthy Rich - 00:02:55
Firelight - 00:02:23
Flam - 00:05:55
Flim - 00:04:49
Fluttershy - 01:56:32
Gabby - 00:04:40
Gallus - 00:05:28
Gilda - 00:03:57
Gladmane - 00:03:39
Goldie Delicious - 00:02:24
Grand Pear - 00:01:04
Granny Smith - 00:22:53
Grubber - 00:01:50
Hoo'far - 00:01:55
Iron Will - 00:04:38
Lightning Dust - 00:03:13
Limestone - 00:01:10
Luna - 00:15:14
Ma Hooffield - 00:01:57
Mane-iac - 00:01:55
Matilda - 00:01:25
Maud - 00:08:02
Mayor Mare - 00:04:46
Minuette - 00:02:03
Miss Harshwhinny - 00:02:48
Moon Dancer - 00:01:59
Mr. Cake - 00:01:43
Mrs. Cake - 00:02:09
Mudbriar - 00:01:30
Night Light - 00:01:13
Nightmare Moon - 00:05:19
Ocellus - 00:04:05
Party Favor - 00:01:40
Pear Butter - 00:01:59
Pharynx - 00:01:43
Photo Finish - 00:01:42
Pinkie - 02:49:29
Pipsqueak - 00:01:47
Pony of Shadows - 00:01:18
Prince Rutherford - 00:04:34
Prince Skystar - 00:02:53
Queen Novo - 00:01:22
Quibble Pants - 00:06:32
Rainbow Dash - 02:50:35
Rarity - 03:04:16
Rockhoof - 00:05:01
Rumble - 00:04:10
Saffron Masala - 00:01:30
Sandbar - 00:04:07
Sapphire Shores - 00:01:20
Sassy Saddles - 00:04:30
Scootaloo - 00:41:03
Seaspray - 00:01:09
Shining Armor - 00:08:04
Silver Spoon - 00:01:48
Silverstream - 00:05:38
>>
>>34033039
Dont need much hardware to clip or sort or clean, but you need a decent rig if you want to help produce AI audio.
>>
>>34033042
Sky Stinger - 00:01:30
Sludge - 00:05:07
Smoulder - 00:07:23
Soaring - 00:02:21
Songbird - 00:03:33
Spike - 01:58:11
Spitfire - 00:08:49
Star Swirl - 00:05:36
Starlight - 01:16:50
Stellar Flare - 00:02:04
Sugar Belle - 00:03:42
Sunburst - 00:09:53
Svengallop - 00:02:56
Sweetie Belle - 00:41:47
Sweetie Drops - 00:01:04
Tempest Shadow - 00:06:01
Terramar - 00:01:32
The Storm King - 00:02:45
Thorax - 00:09:02
Thunderlane - 00:01:09
Tirek - 00:05:03
Tree Hugger - 00:01:22
Trixie - 00:26:12
Twilight Velvet - 00:01:05
Twilight - 04:34:57
Vapor Trail - 00:01:51
Windy Whistles - 00:02:00
Yona - 00:04:24
Zecora - 00:12:35
Zephyr - 00:05:31
Zesty Goumand - 00:01:18
>>
>>34033047
>Sweetie Belle - 00:41:47
Either the squeaks will be extra-random, or they will be averaged into nonexistence.
Please be the former, please be the former, -
>>
>>34033112
Likely averaged out, but adding them back in would be relatively simple once you have her baseline voice.
Just a relatively simple 'voice crack' roll everytime she speaks, which gets weighted if shes speaking excitedly, and a successful roll briefly changes her pitch.
>>
File: J U S T.jpg (31 KB, 456x320)
31 KB
31 KB JPG
>>34033042
>>34033047
>tfw waifu had less than 5 minutes of audio and wasn't even clipped at all
>>
>>34033139
who was it, flurry heart?
>>
File: 1833914.png (1.19 MB, 1125x1080)
1.19 MB
1.19 MB PNG
>>34033177
No, my precious Cajun belle with the sweetest voice in the universe.
>>
File: hey colt how you doin.png (1014 KB, 1246x1024)
1014 KB
1014 KB PNG
>>34033139
>>34033186
I know that feel.
>>
>>34033186
my condolences, brah
maybe some day the tech will be there
>>
>>34033042
Is this clean audio or total audio?
>>
>>34033399
Total.
>>
>>34033042
>>34033047
>Trixie - 00:26:12
>Twilight - 04:34:57
Impressive, the whom started as a one shoot character now have 6% of the most speechy one.
It's even more than Granny Smith.

Here, take a quick and dirty graph of it.
>>
>>34033595
>Rarity has a lot of lines to work with
Yes, very good.
>>
File: 1563023349205.png (76 KB, 333x246)
76 KB
76 KB PNG
>>34033595
GRAPH CHART
>>
>>34033595
That's a nice graph, but I think the axis is mislabeled
>16000 minutes of speech
>>
File: Main6.png (2 KB, 172x82)
2 KB
2 KB PNG
The main 6
>>
>>34033723
RIP flutterfags
>>
File: 1528928204084.jpg (508 KB, 2560x2143)
508 KB
508 KB JPG
>>34019408
You guys are doing god's work, here. I don't know any other community who is so dedicated to the characters and the impact they had on their lives that they are all working together - contributing hours and hours of work - in order to preserve these beloved characters via their voices. Years ago, I thought that once the show ended, everyone would move on and nobody would give a single shit about it. But here you are, working with neural networks and AI and other black magic sorcery to make a sick text-to-speech program.
>>
>>34026474
That not Chinese bruv that Korean.
>>
>>34033639
Why do you think I use purple?

>>34033645
Oh. I guess I'm retarded after all.
It's seconds of course.
>>
>>34033801
Thanks bro.
If the ride never ends, it's because it's fuelled with autism. In the best way possible.
>>
File: 1563030825384.jpg (70 KB, 988x808)
70 KB
70 KB JPG
>>34033878
Based purple-friend.
I am downloading all s8 clips, so expect new big dataset tomorrow.
And also a bonus one

>>34033801
This project and our autism is representaion of our love. Weird, strong, honest love.
>>
Do we have any updates on how the current training is going? (If it still is)
>>
File: 1563032307186.gif (404 KB, 378x489)
404 KB
404 KB GIF
>>34033907
For now, i stopped training due two things:
1) Mistakes in quotes
2) Doubts in h_params.
Currently, i am downloading amazing work of Clipper anon >>34026484.
Then i create dataset from all this clips (may be even today). Tomorrow i want to re-read all quotes and delete all "bad" qoutes with long pauses, screams, etc.
Then i will create new dataset from it and also a clear version with clips without noises.
From monday i will start training, to see where i will stuck again.
After all that i will start exepiments with h_params, so yeah, sorry not much to update, actually.

I wish i could spend more time on this project. In august i will have vacation from job, so i will dive into neuronetworks more deeply, so in future i could provide a more easier way to use our model.
I think even if we end up with not perfect model of Twilight to september, it is better to start doing application for everyanon to use.
>>
File: once apon a time.png (26 KB, 2540x301)
26 KB
26 KB PNG
Heres S1E1 automatically aligned data.
https://mega.nz/#F!DT4iFIqZ!8FVxgpaiXU1rAKyl7iLwAg

Each clip has 3 files,
>json straight from gentle align with all the data
>audacity label for words
>audacity label for phones

the phones are in Arpabet, underscore, then a letter showing if its at the (B)eginning, (I)nterior or (E)nd of a word, offsets are based on how many phones are in the word, rest should be self explanatory. I used that to make audacity labels so you guys can easily look at it.

Even the short ones will need some adjustments, the first line (pic related) has about 6 small adjustments, but it will be easy to do since the starting positions are pretty close.

This isnt finished data, just an example, theres some pre and post processing that needs to be done:
>non dictionary words need to be added back in with basic phone positions
>json converted to better format
>non aligned words added back in with basic phone positions
>Convert phone representation maybe

But if someone wants to they can go over these and align them properly, it will be slightly more work since you'll have to manually add non dictionary words, but the work wont be wasted.
>>
File: 2088258.jpg (116 KB, 900x1024)
116 KB
116 KB JPG
>>34033907
Also stopped training, mostly because my 1050ti doesn't have enough memory for 32 batches. I set up a copy of the google colab so that once a bigger dataset is done I can get to training again.

Also regarding the digital assistant voice with mimic2, basically no pretrained models exist in a way that would allow me to continue training them. Training a model from scratch for 8 Hours on the existing twilight dataset for seasons 1 - 4 was able to get intelligible speech though, so that might be enough to train from scratch on. If not then training on LJSpeech for a while then training on twilight's dataset would probably work.
>>
>>34033953
>>34034006
Nice to see that even if you stopped training for some reasons, you still continue to document and enhance the process for you and other to continue the training!
Based work!

>>34033983
Great!
I will download it and have a look at it next week.
I probably won't have much time this week end.
>>
>>34033186
The multi-speaker thing could potentially be used for her. You could find someone who has the same accent and has appeared in a lot of media, and clip their audio. Someone would train a "cajun accent" model with that, and then pass it the pony audio as a new speaker.
>>
>>34033595
Very exciting that Rarara has so much.
>>
File: 1509790449101.png (504 KB, 537x766)
504 KB
504 KB PNG
>>34034439
Are there even any popular actresses with such accent who starred in many things that were released in high quality 5.1 sound? Sounds fucking impossible. I'll totally try if it's possible though.
>>
File: 2018860.png (349 KB, 800x450)
349 KB
349 KB PNG
>>34019424
>>34030942
I've finished s9e3 - 7, and added the Audacity labels to the master file. Also turning off my name now since it's not really necessary as I'm back to just regular clipping.

>>34031184
>>34031588
Thanks for doing all that, will certainly be helpful once I'm done with season 9.

>>34033047
>Tempest Shadow
I forgot that the movie had been done, I'll see about adding it to the master file. Have I forgotten anything else?

>>34033953
>>34034006
Good to hear you're still around. If you're going to be going through all of season 1 - 8, please do let me know if you find any significant mistakes/typos in the data.
>>
>>34035025
Forgot some of the youtube stuff like the fundamentals of magic, behind the boutique and some other mobile games. One thing I noticed when counting the character time, fluttershy is tagged regularly when she had the poison joke and the voice was different. Also, I think the stick/mean version of the mane 6 can be counted regularly as they're not much different. Some characters are also under two different names. Noticed Neighsay and Chancellor Neighsay for example.
>>
>>34034864
I don't know. You could use a male actor, too. But I can't think of any examples of media with prominent Cajun accents. Well, I guess Waterboy, but I don't think you want her to sound like Momma.
>>
So what happens if the AI doesn't have enough voice samples? Does it guesstimate what the character would sound like in an unknown situation (I don't know, maybe a character who never vocally expressed anger has been finagled and now someone is trying to create a voice clip of her getting angry) based on what it knows? What's the cutoff for where the voice would start to sound noticeably robotic (compared to the norm) as the AI struggles to figure out what noises to make?
>>
>>34036381
this happens
https://clips.twitch.tv/AlluringNimbleNeanderthalTakeNRG
>>
>>34035229
>fundamentals of magic, behind the boutique and some other mobile games.
I intend the master file to be used for processed audio only, so unless those have been clipped, I'll leave them out. This is mainly because I'm using a free Mega account that puts limits on how much I can upload, and I don't want to have to make and manage multiple accounts if I don't have to.
>fluttershy poison joke and Neighsay.
I'll look into those. Thanks for pointing them out.
>>
>>34036765
The mobile game audio quality was awful.
And I did some work on it, but, clearly, if we add it to the dataset, it will be worst than without it.
>>
File: 1417737729430.gif (2.33 MB, 992x950)
2.33 MB
2.33 MB GIF
Fluttershy
>>
>>34037274
We should use an AI to clean up the audio
I'm sure it should be possible with modern techniques to take a noisy voice sample and output clear voice
>>
File: 1563123420502.gif (365 KB, 648x864)
365 KB
365 KB GIF
>>34026484
>s5e1
redundant file from netflix.

Big dataset from s1 to s8. Thank you Clip anon!
https://mega.nz/#!SowzUKqQ!-tqR-1pkM0fdD4DJj8ZBeHA-m6DcTEG6rQU0xb4Z-B4
I didnt check it on mistakes, tho. Currently generating only clean version without noises.

Happy learning, my lads!
>>
File: 1417995221257.png (173 KB, 851x938)
173 KB
173 KB PNG
>>34037515
Oh fuck yes. Gonna continue training my old deep voice model on this to see if it improves results at all. Thanks for your hard work twibot
>>
File: 1563123755064.jpg (231 KB, 1024x1365)
231 KB
231 KB JPG
>>34037525
For our waifu!
>>
File: 2057210.png (372 KB, 4145x5771)
372 KB
372 KB PNG
>>34019424
>>34035025
I’ve finished s9e8 - 11. I haven’t been able to find any subtitle SRT files for episodes 12 and 13, so this is as far as I can go with season 9 for now. The label files should be good to go for when the higher quality rips become available.

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

I’ve now also added the clipped MLP Movie the master file. It all looked fine, so I didn’t change any of it. The anon who worked on it used a few different emotion labels such as “nervous” and “excited”, which weren’t part of my list. Not really a problem, just something to be aware of.

I’ll make a start on the original EQG Movie tomorrow.

>>34035229
>Chancellor Neighsay
I’ve replaced all occurrences of “Chancellor Neighsay” with “Neighsay” in the character tags of s8e16 and 25. The modified audio clips and transcriptions have been reuploaded, along with the new labels in the label files folder.

>Mean Six
I’m a bit conflicted on this one. I think they sound different enough to justify different tags, but they are still technically the same characters. I’ve kept the tags as they are so whoever’s working with it later can make their own decision. They only make up a tiny portion of the dataset anyway so it’s not that big a deal.

>Flutterguy
There’s a lot of Fluttershy audio to recheck. Could you specify the episodes?

>>34037515
>>34037525
>s5e1
Removed the redundant file, and am very much looking forward to more TwAIlight updates.
>>
>>34037973
I think for Fluttershy it's mainly episode s01e09. There may be a few more instances of it, but I'll have to look around to see.

There was also Celestia/Princess Celestia in the character names. This was also the case for Cadence and Luna. I think there might have been another instance of this, but I would have to check.
>>
>>34038065
Alright, I'll run back through s1e9 and recheck the Fluttershy lines. The only other instance of Flutterguy I recall is s4e14, where Fluttershy sings as Flutterguy in the pony tones. I'll recheck that one as well. Let me know if you find any others.

The checking script should have replaced all occurrences of "Princess" in the character tags, so not sure what went wrong there. Again, let me know where you find them and I'll take a look.
>>
>>34037973
Damn, you did a really huge chuck of work for this project.
Thank you based anon!
>>
anyone have a clip of fluttershy saying 'inside'?
>>
File: autism.png (407 KB, 650x582)
407 KB
407 KB PNG
>>34038883
She said it once (as opposed to Twilly that says it more often).
Season 6 Episode 2, 5mins 41 sec
Fluttershy, Neutral (Noisy) : "still, when you think about it, the view is just as good a little further back like, inside your house."
>>
File: 569384.jpg (35 KB, 500x371)
35 KB
35 KB JPG
Thanks to this new technology we now possess, prompts, CYOAs, greentexts, first person posts/threads, etc, will never be the same ever again.

In fact, they'll be even better and far more entertaining than they would normally be thanks to this new voice mechanism.

This shit is gonna be so fucking awesome.
>>
This looks very ambitious but very interesting.
>>
What if when you guys are finished this, and if it's not system resource heavy, we convince Moot to integrate it into the board with a "Read Post" button in the arrow dropdown thingy beside the post number?
>>
>>34039811
one moot works for googgle he has been for the last 4 years.
also the new 4chan guy is a asian eyed dog eating earth monkey
>>
>>34039811
It doesn't need to be integrated, if someone is willing to run a server a greasenmonkey script is all that's needed to add a button like that.
>>
File: 1563218405312.jpg (100 KB, 1140x701)
100 KB
100 KB JPG
My eyes are hurting, but i reread all s1 quotes and found some typos. I hope i found them all... Will try to make it more efficient tomorrow.


S1E1
Legend has it that on the longest day of the thousanth year => thousandth
That we are on the precipus of disaster => precipice
Oh, kay => Oh, Okay
S1E10
qi got it! => I got it!
S1E11, E13
Oh, kay => Oh, Okay
S1E22
Then you have to mak her take it => Then you have to make her take it
S1E6
Twilight just let her go => Just let her go

S1E9
Dear Princess Celestia, my friends and I all learned an important lesson this week, ever judge a book by its cover =>
Dear Princess Celestia, my friends and I all learned an important lesson this week, never judge a book by its cover

S1E8
It's hard to believethat two ponies who seem to have so little in common could ever get along =>
It's hard to believe that two ponies who seem to have so little in common could ever get along
>>
>>34041571
Very nice.
>>
File: 343710.png (1.11 MB, 1500x1293)
1.11 MB
1.11 MB PNG
>>34019424
>>34037973
>Make a start
Turns out I finished. Delivering the audio clips and transcriptions for the original EQG movie.

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

I’ve also added the various songs to the song folder, and a separate copy of the labels to the labels files folder named “eqg_original movie.txt”

I’ll do EQG - Friendship Games next.

>>34041571
Wow, I wasn’t actually expecting that you would trawl through the whole thing, so thanks for that.

>Oh, kay
I named those like that because that is how it’s pronounced, so double check the audio files and make sure that you want them changed like that. I’ve fixed and re-uploaded all the others. I haven’t got around to rechecking for Flutterguy in s1e9 and s4e14 yet, I’ll try to get that done tomorrow.
>>
Anyone started Rarity training yet?
>>
>>34042013
We’re doing Twilight first as she has the most audio and then will start working on the others I assume
>>
>>34042304
I mean it makes sense, but Twilight's voice isn't all that special, considering other voices Tara's done that are similar, when compared to how inimitable Rarity's is.
>>
>>34042338
I think the idea is we can likely perfect Twilight's voice first more easily. A lot of this is a learning process, and learning on the low-hanging fruit will be more beneficial in the long run.
>>
>>34042304
>>34042338
>>34042569
After Twilight's voice is done, we should do a small Anon x Twi greentext story to test out her voice algorithm.
>>
>>34043136
Introducing another character, especially one that doesn't have any defined voice will only complicate things in a wrong way. Let's focus on the show's characters first.
>>
>>34043336
What about a "Message in a bottle" situation where she writes about how everyone is doing after the show ended and how much she misses you
>>
>>34043407
Well that'd break my heart and give me feels I don't know how to deal with.
>>
>>34043427
Not gonna lie, Anon. When this thing eventually comes into fruition, I'm going to pathetically make my favourite mares say affectionate things towards me.
>>
>>34043336
Huh?? Are you referring to Anon? But he doesn't have a voice to begging with. It's not like we're forced to give him one, so why should it complicate things?
>>
>>34043616
couldnt you just make a voice based on literally every persons voice to make the most generic "hey that kinda sounds like me" kinda voice? better yet couldnt we just use an already existing generic sounding high quality speech synth?
>>
>>34043775
Better still, just use some basic TTS like Microsoft Sam.
>>
To walk the path and reach the destination... God bless you sacred pilgrims.
>>
>>34044349
I'm going to spend HOURS voicing my fucking greentexts with each character's proper voice.

You people are actually amazing.
>>
>>34043407
It could be seen as a stress test where the message starts very simple because she wants it to get through but starts rambling about a complex topic also testing how it handles words that don't exist.
>>
>>34043437
That was the plan from the start, my dude
>>
File: 1563304754514.png (443 KB, 594x743)
443 KB
443 KB PNG
I copied all quotes to Office Word, so now i can work much quicker, but still i looked only one season.
Sorry guys.
I hope i will end with mistakes in this week and start training. I am not rushing, because other anon should train in parallel, so i have free time to make dataset better and may be to play with h_params.

Cut the crap, mistakes from S2 are here!

S2E3 - Item 1, create checklist of the things i have to accomplish by the end of the day =>
Item one, create checklist of the things i have to accomplish by the end of the day

S2E25 - please, call me princess mi amore Cadanza => please, call me princess mi amore Cadenza

S2E25 - Princess celestia cordially invites you to the wedding of Princess mi amore cadanza, and =>
Princess celestia cordially invites you to the wedding of Princess mi amore cadenza, and

S2E21 - We're just a little early, and i'm glad we are. This way, we can watch every moment of the migratiom Without bringing any unwanted attention to ourselves =>
We're just a little early, and i'm glad we are. This way, we can watch every moment of the migration Without bringing any unwanted attention to ourselves

S2E20 - But if he's here, then all the ancient evil creatures that have been imprisoned therec ould escape and destroy equestria! =>
But if he's here, then all the ancient evil creatures that have been imprisoned there could escape and destroy equestria!

S2E20 - What happend to you the future must be awful => S2E20 - What happened to you the future must be awful

S2E3 - I got to think of something. thinkthinkthinkthink, twilight think! =>
I got to think of something. think, think, think, think, twilight think!

S2E3 - Nonononononono => No, no, no, no, no, no, no
>>
File: 994505.jpg (495 KB, 842x1200)
495 KB
495 KB JPG
>>34019424
>>34041942
Delivering the audio clips and transcriptions for Equestria Girls: Friendship Games.

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

I’ve also added the various songs to the song folder, and a separate copy of the labels to the labels files folder named “eqg_friendship games.txt”

I’ll do EQG - Legend of Everfree next.

>>34045075
Fixed and re-uploaded all, thanks again. Did you double check the “Oh, kay” files from earlier? >>34041942
>>
File: screenshot.png (136 KB, 1920x1041)
136 KB
136 KB PNG
>>34045186
While doing the setup for EGQ - LoE, I’ve noticed a potential problem with the audio. LoE seems to be much louder than anything before, using up all the vertical space on the audio track, where all others before never really went above/below 0.5/-0.5. Pic related shows a comparison between EQG - FG (top) and EGQ - LoE (bottom). Both of these are the processed audio.

Is this a problem, and if so, is there anything I can do in Audacity to fix it?
>>
>>34045257
I think it should be set to about the same volume as the rest, you can change it with Effect -> amplitude. Someone who knows more about audio should say if it will be fine or if audacies amplitude filter fucks it up in some way.
>>
>>34045257
You sure you decoded it into .wav properly? Is it the same in this >>34026653 anon's upload?
>>
>>34045327
>decoded it into .wav
They're .flac files, from >>34031184
>>
>>34045257
As >>34045320 said, you can lower the amplitude. The only concern would be if clipping occurs because of the high amplitude. Basically, keep track of any audio that may have issues so that better sources can be located in the future for those that need it.
>>
>>34045334
Can't even check it cause the filehost is fucking dead
>>
>>34045320
>>34045338
I don’t want to mess with things I don’t understand, so I’ll just turn my volume down and clip as normal for now. The label file should be fine to transfer into a different audio source later on.

>>34045345
I’ve temporarily re-uploaded the processed audio for EGQ FG and LoE to the master file.

https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw
>>
File: wsfsdfdsf.png (1.64 MB, 1680x1020)
1.64 MB
1.64 MB PNG
>>34045390
Yep, it's completely fucked up with clipping. It's unusable.
>>
so how did you guys go about doing the original derpy voice+the one that substituted it?
also reminder of the existence of baking with pinkie(movie extra).
>>
File: faith.png (404 KB, 1512x868)
404 KB
404 KB PNG
>>34043811
>My God... what happened to you?
>DO YOU THINK MY VOICE IS PRETTY?
>>
>>34045402
Alright, fuck that then. I'll just take tomorrow off and find something else to do later.
>>
>>34026653
why not use this?
https://mega.nz/#F!hIgR2STC!3_1uh5qHcaxk_X5JEe115g
or ask that anon(if he shows up again) to provide the raw 5.1 audio?
>>
>>34045417
That stuff still has all the background music in it >>34029310. If any other better audio exists, I'll gladly take it.
>>
>>34045414
it might sound i bit dumb to you, but you could do other dubs, for example the japanese dub since the fucks never got season 3 and beyond and quite a few episodes(if not all) had some scenes/lines removed entirely, if i remember correctly rome has the german and polish dubs, as for the pasta and baguette dubs i can't remember right now or you could even do gen 1 and 3(but i would much prefer for you to take a break, you have earned it with flying colors, you glorious horsefucker).
>>
>>34045439
That does sound a bit dumb, mainly because I don't really see how the dubs would be helpful to the dataset. I only speak English, and the dubs would presumably have used different voice actors.
>>
>>34045426
i think yay has the bluray iso's, might be wrong though(if i am wrong rome might have them, although i might be wrong on that as well).
>>34045439(me)
going on that japanese note i remembered that they also had the eqg movies and even the shorts on youtube(you can find the other languages either on the playlist of the channel or by doing a quick search for "hasbro eqg languageofyouchoice" without quotes)
https://www.youtube.com/playlist?list=PLNXV9Mb7egjFnPuL9VN5F3OBvelQS453t
also i make no promises but i think i could find you the japanese eqg movies, i do not know how big the are or even if the links still work(let alone if the page that had the links is still up) bit it might prove useful.
also delivering from the eqg tread their mega for the shorts
https://mega.nz/#F!baBQTKwB!pddUvRszMyT1KIUFvUqA3Q
>>
>>34045439
Is it possible to "teach" the synthesized English voices how to speak other languages?
>>
>>34045483
They only have some of the isos. The best quality available at yayponies was used. Other sources will be seeked out later.
>>
>>34045459
bro... there will come a moment in which you will have done all the clipping that can be done for that english release so i thought of offering you audio to clip so that your clipping muscles don't ache from idleness or as i said you could do gens 1 and 3, which will be useful if for example the greens include a previous gen pone.
and it would obviously be for making datasets of other languages, i'm with you on that, i apologize if i came across as saying that for example, spanish pinkie pie clips would be of use for the english dataset.
>>
>>34045515
rome has them
https://kwn6i5jng6bsow5z.onion.pet/Movies/Equestria%20Girls%201/English/Blu-ray%20rip%20[1080p]/
https://kwn6i5jng6bsow5z.onion.pet/Movies/Equestria%20Girls%202%E1%8D%A1%20Rainbow%20Rocks/English/Blu-ray%20rip%20[1080p]/
https://kwn6i5jng6bsow5z.onion.pet/Movies/Equestria%20Girls%203%E1%8D%A1%20Friendship%20Games/English/Blu-ray%20rip%20[1080p]/
https://kwn6i5jng6bsow5z.onion.pet/Movies/Equestria%20Girls%204%E1%8D%A1%20Legend%20of%20Everfree/English/Blu-ray%20rip%20[1080p]/
i think the shorts are already on that mega >>34045483(me), but just in case
https://kwn6i5jng6bsow5z.onion.pet/Movies/Equestria%20Girls%202%E1%8D%A1%20Rainbow%20Rocks/English/Shorts%20%28Normal%29%20[1080p][Bluray%20Encode]/
https://kwn6i5jng6bsow5z.onion.pet/Movies/Equestria%20Girls%203%E1%8D%A1%20Friendship%20Games/English/Bluray%20Bonus%20%28RAW%29%20[1080p]/
>>
>>34045483
>>34045540
Alright, I'll look into all that tomorrow. I think it may be useful to spend some time organising all the audio we have left and put together a treasure hunting list for stuff still to find.

>>34045526
>gens 1 and 3
Maybe later, but that's way into the future right now.

>other languages
Again, I don't speak other languages, so the stuff I produce for that will likely be sub-par. For now, let's stay focused on actually making it work in English before making it more complicated later.
>>
>>34045439
I doubt good transcripts even exist for most languages. The only subtitles I've seen in my language are translations of the original English subtitles, not actual transcripts, and as you can guess, different translations made by different people are often too different to each other.

>>34045504
Kinda.
https://ai.googleblog.com/2019/05/introducing-translatotron-end-to-end.html
You don't need the non-English dubs for this though.
>>
>>34045564
well, that just sounds like a job the guys that love their dub and want to preserve it can take: transcription.
but all in due time.
>>
>>34045564
Fucking hell. It would be really nice to hear ponies speaking my native language in their original voices instead of talentless local dub cunts.
>>
Here are subtitles for season 9 episodes 12 and 13:

https://anonfile.com/K9UaNaz2n9/S09E12_and_S09E13_Subtitles_7z

Added to doc as well.
>>
any progress on training?
>>
I know this is thinking a bit far ahead, but I wonder if this technology could be applied to something like voice changers or voice masks of some kind. Like once we perfect this, could we eventually make it so that we could speak into a mic and have it come out as a pony's voice in real time?
>>
>>34046433
>COD
>"He's using a soundboard!!!"
Nah, this is just my voice mate.
>>
>>34046669
Oh god, it would be a new golden age of pony trolling.
>>
File: 1563342637447.png (62 KB, 283x389)
62 KB
62 KB PNG
>>34045186
> Did you double check the “Oh, kay” files from earlier?
No, sorry. I forgot. Will check todays evening.

>>34045408
We probably haven't so many lines of original derpy to reproduce voice.

>>34045504
Yes, but no. That will require new model and i didn't find any existing for this task. Basicly as i understand the proccess, you need somehow match phonems from other languages, which is a lot of work.

>>34046100
Only if other anon can report something. I am currently working on typos.
>>
What do you guys plan on doing as soon as this works? Even if the first version is buggy and sounds like a robot is giving a speak-and-spell a blowjob, what will you do with it first, and what do you want to be the first words you hear out of your waifu's mouth?
>>
>>34046784
Try to get Mycroft working on my system and hook it up to this so I have my own Twilight personal assistant.
>>
>>34046796
This.
>>
File: DyyqH1aX4AAcLuV.jpg (811 KB, 2562x2340)
811 KB
811 KB JPG
>>34046784
>what do you want to be the first words you hear out of your waifu's mouth?
I want her to say hello to me and say my name.
>>
>>34046433
>>34046669
Already exists, and since it has a complete voice to work with instead of just text it sounds a lot better. Heres the first shitty example i found
https://soundcloud.com/andabi/sets/voice-style-transfer-to-kate-winslet-with-deep-neural-networks
>>
>>34046837
Holy shit boys, we are living in the future.
>>
>>34046784
Use it in combination with Oculus' LipSync and the SFM models in Unity, to make ponies say dumb and OOC shit.
>>
>>34045540
>>34045483
>>34045439
It's a secret
https://drive.romesilvanus.io/wl/?id=BCICU6tNX2kOZhLIy9DkkMYquulYbvky&mode=regular
https://drive.romesilvanus.io/wl/?id=BCICU6tNX2kOZhLIy9DkkMYquulYbvky&mode=list

There's an sftp server available through TOR:
sftp://kwn6i5jng6bsow5z.onion:22
UN: whinny
PW: neigh
>>
>>34046720
>We probably haven't so many lines of original derpy to reproduce voice.
even then we should archive everything.
>>
>>34045483
so on that promise i did not make, i uploaded them to rome so even if a jap dataset is not going to be made say, tomorrow in the afternoon at least you guys can enjoy some eqweeb.
https://iwiftp.romesilvanus.io/Shared%20by%20other%20people/Upload%20form/



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.