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

[Advertise on 4chan]

Name
Spoiler?[]
Options
Comment
Verification
4chan Pass users can bypass this verification. [Learn More] [Login]
Flag
File[]
  • Please read the Rules and FAQ before posting.
  • There are 97 posters in this thread.

08/21/20New boards added: /vrpg/, /vmg/, /vst/ and /vm/
05/04/17New trial board added: /bant/ - International/Random
10/04/16New board for 4chan Pass users: /vip/ - Very Important Posts
[Hide] [Show All]


Janitor acceptance emails will be sent out over the coming weeks. Make sure to check your spam box!

Self-serve ads are available again! Check out our new advertising page here.


[Advertise on 4chan]


File: AltOP.png (1.54 MB, 2119x1500)
1.54 MB
1.54 MB PNG
TwAIlight welcomes you to the Pony Voice Preservation Project!
https://clyp.it/tm03e5en

This project is the first part of the "Pony Preservation Project" dealing with the voice.
It's dedicated to saving our beloved pony's voices by creating a neural network based Text To Speech for our favorite ponies.
Videos such as youtu.be/GuJKTodX1FA. or 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 (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://docs.google.com/document/d/1xe1Clvdg6EFFDtIkkFwT-NPLRDPvkV4G675SUKjxVRU

We now have a working TwAIlight that any Anon can play with:
https://15.ai/
https://derp.link/vCzm2 (48KHz Training)
https://derp.link/hdJQF (48KHz Synthesis)
https://derp.link/NR7Xi (Ngrok Synthesis)
https://derp.link/YTJ94 (Guide)

>Active Tasks
Cookie is working on controllable speech
Research into animation AI
Research into pony image generation

>Latest Developments
Clipper sorts animation files (derp.link/O24pp)
Clipper looking for AI skit ideas (derp.link/JfVsA)
HIFI-GAN test notebook (>>36597481)
Progress with animation and Python (>>36673397)
Colab notebook for image tagging (>>36677902)
Clipper collecting sound effects from show (>>36723767)
BFDIAnon modifies Cookie's ngrok to have emotional control (>>36724150)
New test site test16.15.ai
New DeltaVox (>>36812261)
Start prepping for /mlp/con panel (>>36883353 >>36888228)
Training notebook for HiFi-GAN (>>36874641)
New guides and notebooks for training/exporting models for DeltaVox RS (>>36898031)
Latest Synthbot progress report (>>36865763)
Latest Cookie progress report (>>36829703)
Latest Clipper progress report (>>36804225)

>Voice samples
https://derp.link/fHs3K
https://derp.link/O1xdh

>Clipper Anon's Master File 2.0:
https://mega.nz/#F!L952DI4Q!nibaVrvxbwgCgXMlPHVnVw
https://mega.nz/folder/0UhSmYAB#WBrB-qCprQTofkAhwMp5CQ

>Synthbot's Torrent Resources
https://derp.link/ZJNca

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

Last Thread:
>>36828429
>>
FAQs:
>READ THE DOC
Do it now
derp.link/V7cMp

>Where can I find things made with the voice AI?
In the Good Poni Content folder: derp.link/23EUs

>Did you know that such and such voiced this other thing?
Yes. We are very much aware. It is best to keep to official audio only unless there is very little of it available. If you know of a good source of audio for characters with few (or just fewer) lines, please post it in the thread. 5.1 is generally required unless you have a source already clean of background noise. Preferably post a sample or link. The easier you make it, the more likely it will be done.

>What about fan-imatitions of official voices?
No.

>How do I make the voices?
Several guides are available. In depth guides on how to do training and synthesis (making the ponies speak) are in the doc. If you don't want to use the navigation bar in the doc, the sections are also directly linked in the OP. If you want to use the WiP 48KHz notebook, some kind Anons have put together some image guides for you.
48KHz Training: derp.link/wW2hX
48KHz Sythesis: derp.link/j4MXQ

>How do I make the ngrok links?
Doc: derp.link/SfIhY
Video: derp.link/qYgIp

>Where are all the voice samples?
In the doc.

>Is a place I can find all the pony models?
In the doc.

>What about muh waifu?
Check the doc.

>Will you guys be doing a [insert language here] version of the AI?
Probably not, but you're welcome to. You can however get most of the way there by using phoenetic transcriptions of other languages.

>What about [insert OC here]'s voice?
Not a priority. Again, however, you're welcome to. There are already people doing this.

>Where can I view the PPP /mlp/con panel?
YouTube: youtu.be/WtuKBm67YkI
CyTube chat: pony.tube/videos/watch/b83fbbfc-6d4e-4768-8deb-edb61ea38abb

>I have an idea!
Great. Post it in the thread and we'll discuss it.

>Do you have a Code of Conduct?
Of course: 15.ai/code

>Is this project open source? Who is in charge of this?
derp.link/CQ3Ca
>>
File: AnotherAnk.jpg (73 KB, 800x1099)
73 KB
73 KB JPG
>>36904619
Anchor.
>>
File: 1601064214592.png (130 KB, 961x1024)
130 KB
130 KB PNG
>>36904626
For anyone who missed it at the end of the previous thread - >>36901235
New voiced greentext - https://youtu.be/b-47LopPBTY
Download - https://mega.nz/file/sdo1gYyL#yDxS4YH0QF3AU9_kX7sdzATIUmSBO7OCPNfsWP4IMPQ
ponepaste - https://ponepaste.org/4735
Clipper voice dataset - https://mega.nz/folder/wMYGBL4K#woaoMkLL3bgN9amrBck26w
>>
Here's a Clipper dataset based on >>36901235. It's 16 minutes long, and I'm training a model with it right now.

https://drive.google.com/file/d/17HTBgywA-i1ZWZ_BSL5EABvQNNsmwpaV/view?usp=sharing
>>
https://vocaroo.com/1eGUbmUogTWG
stupid sexy evil Glim
>>
>>36904879
You might want to double-check the transcripts, I'm seeing a line transcribed as:
>I didn't expect a mare of science like yourself WOULD be so resistant to change.

When the voice actually says:
>I didn't expect a mare of science like yourself TO be so resistant to change.
>>
>>36904901
I went through it again and only found one other mistake.
>And the guy died from gum disease though -> Eh the guy died from gum disease though
I've updated it to fix those. There's also "Winnie's world record holder" as opposed to "Whinny's", but they're pronounced the same, and "Winnie's" is in the dictionary.
>>
>>36904692
Thanks anon.
>>
>>36904692
> "Reenact a role reversal of Isaac Newton's claim to fame."
Nice wordsmithing.

And that fucking ending. Comedy gold.
>REEEEEEEEEE
>>
>>36904879
>and I'm training a model with it right now
Where?
>>
>>36904692
Holy fuck, that was absolutely brilliant Clipper. Fantastic work.
>>
File: 1495375391116.jpg (11 KB, 225x225)
11 KB
11 KB JPG
you all give me something to look forward to every week
and lately its been several things even!
thank you!
>>
>>36904626
>>36904879
I've added the clipped version to the master file 2, it's in the "Clipper's Voice Dataset" folder. Keep us updated on progress, I'm very interested to see how well it turns out.

>>36905206
You're welcome.

>>36905264
>>36905437
>>36905438
And thanks to you all as well, glad you enjoyed it!
>>
>>36904626
Clipper has been added to the HiFi-GAN notebook.
https://colab.research.google.com/drive/1dxVcqe4m-AU8NAA1I1MW1N9HYBO_oii_

Sample: https://u.smutty.horse/masytpzzmii.ogg
>>
>>36905521
ClAIpper
Well, that was fast.
Congratulation Clipper, your voice is now immortalized.
>>
>>36905521
wtf how could clipper say this????
>>
>>36905620
It's an AI, you dodo.
You can now use Clipper's voice to narrate your greens, how cool is that?
>>
>>36905521
neato, finally I'm one more step to fulfill my yaoi fanfic fantasies about Clipper.
But seriously, this voice model sounds pretty decent, how many minutes was the Clipper model trained with ?
>>
File: higan shit broke.png (64 KB, 904x792)
64 KB
64 KB PNG
>>36905521
Can someone help me understand this error from the tuning higan colab? im using the exact same files I did to train the colab tacotron2 model yet the higan refuses to accept the wavs file here.
>>
>>36905897
never mind, I've manged to fix it with my second rate codding. Here are instructions
https://pastebin.com/fjdEuXH1
It seems just few files in the data were not saved in the exact format and my version of 22 tacotron2 colab already has code I just copied and pasted it in between the step 5 cell.
>>
Here is the same prompt generated twice with different vocoders. Which one sounds better? (Focus on audio quality)
https://u.smutty.horse/matapbvhaly.wav
>>
>>36906284
the first clips seems to be cutting words at random places while the second one introduces a tinny amount of extra noise artifacts happening in background.
Both f them need more training/fine-tuning it seems.
>>
Is daddy 15 still alive?
>>
>>36906323
no, he returned to his home planet, and the entire text to speech scrip website was just tax scam.
>>
>>36906284
I can barely understand Twilight there, she's smashing entire sentences into one word like she's an UberKraut
>>
>>36905521
>human flesh clipper is now completely obsolete
PPP strikes again
>>
>>36905521
>>36905530
CliPPPer
>>
>>36906323
He's too busy fragging noobs on tf2
>>
>>36906628
That's it, Clipper has to officially rename himself as Clippper.
>>
>>36906284
The first one sounds less robotic to me.
>>
Finished the audio. Won't post right now, gotta sleep. Soon as I wake up and proof-listen to it, I'll post it.
>>
https://u.smutty.horse/matdglbruin.mp4

I have now turned the greentext audio to a video, and I also noticed a tiny error on "you were all" part but "all" replaced with "like".
>>
Tx/SG Pony Zone still in progress, going slow due to Starlight being uncooperative.

>>36905496
>>36905521
https://u.smutty.horse/matdouqhgrx.wav
I can't believe Clipper got Trixie's show taken off the air

>>36907115
Quite a shocking turn of events ^:)
>>
>>36905521
Wow, that was fast and sounds surprisingly good. The pacing's a little off and there's a bit of buzz towards the end, but it's still quite impressive.

>>36905624
>You can now use Clipper's voice to narrate your greens
I'm still happy to lend my actual voice to any audio projects if anyone wants.

>>36905648
>how many minutes was the Clipper model trained with?
16 minutes according to >>36904879

>>36906284
There's not much to choose between the two of them, the voice in both of them sounds weird in the same general way. I can't come up with a word to describe exactly why the voice sounds wrong, but it should be obvious to anyone listening to it that it's quite unnatural.

Focusing only on the quality, I'd say the second one is slightly better in terms of noise, though again the difference is marginal. On the word "community", the first one stumbles at the end of the word while the second one doesn't. I can also hear significantly more buzz in the breaths taken in the first than the second one.

>>36906585
>>36906628
>>36906731
The model can be CliPPPer, but I'll always remain the true OG Clipper. That is at least until someone finds a way to fully automate the gathering of voice data, the creation of voiced greentexts and trains GPT-3 to write posts in my style.

>>36907115
Nice job, I especially liked the use of the intermittent laughs at the start and then the nervous hoofsteps at the end. You constructed the scene really well in that one.

>>36907172
I stand by everything I said.
>>
File: laughing_nandroids.png (347 KB, 507x459)
347 KB
347 KB PNG
>>36907172
This is hilarious.
>>
>>36907267
You should sneak a CliPPPer line into the /mlp/con panel and see if anyone notices.
>>
>>36907362
Oh god, that would be huge!
Or even the whole intro?
>>
File: file.png (50 KB, 237x224)
50 KB
50 KB PNG
>>36907172
Are you Celestia?
Cause my sides are at /see pic/.
>>
File: Smile HD.png (350 KB, 576x564)
350 KB
350 KB PNG
Finally got the HIGAN decoder to train past 5k steps, the results seems pretty decent, there are few artificial noise but that's nothing that can't be edited out, for now.
https://vocaroo.com/1cULKDFMT774
https://vocaroo.com/1clAdAgE2EQf
https://vocaroo.com/1nyokVtO2eKU
https://vocaroo.com/11CI17Vm2zwj

The tacotron2 model has random trouble pronouncing some words when they happen to be at start/end of the sentence. I will need to shuffle the val/train text files bit more , maybe the future models will have less problem with that.
https://vocaroo.com/1bPVMCrrpuxc
https://vocaroo.com/16vcCk0Lp7Ka

It seems clearly I've messed up something in process of training model, as the Ai version is loosing a lot of 'snarky/angry snap' when talking (but I feel it's still recognizable enough if people aren't looking for tone inconsistencies).
"How do I get to the upper quarter ?"
AI:
https://vocaroo.com/14kYz1V7GI4q
Original File:
https://voca.ro/1i13XxxxSFxM

Thanks for all the help PPP, I am eternally grateful for all the hard work you guys put into this project and the fact I could witness and participate with you guys too.
>>
>>36907814
Are you gonna train the original German VA too?
English version really lacking in comparison.
>>
>>36907874
I have no idea how to go about that, I remember making an attempt at converting the Arpabet script to IPA script but my lack of understanding advanced coding as well as NN scripts being pain in ass to work with did not lead anywhere so I abandoned the idea.
I guess I could try to prepare general audioset but my lack of speaking and understanding German would make it pretty difficult to actually segregate good and bad audio from each other.
Not gonna lie, if someone figures out how to use IPA with speaking models (both for training and making models speak foreign languages) I would be onboard to get as many English/Polish/German Gothic models made as possible.
>>
>>36907934
IPA is hipster garbage, try Pilsner.
>>
>>36907814
>AI sounds more natural than the original file
We have passed the need for voice actors. AI is the only way forward
>>
File: 1602993393500.png (1.27 MB, 800x540)
1.27 MB
1.27 MB PNG
>>36906635
cheecky cunt
>>
>>36907814
>you little beechuh
hehehehe
>>
File: hqdefault.jpg (14 KB, 480x360)
14 KB
14 KB JPG
>>36907985
>AI sounds more natural
It doesn't. You just prefer it because the super-res makes it sound clearer.

It's like those infomercials that use colors to trick your brain.
>>
File: Daring Do Image.jpg (3.56 MB, 3000x4182)
3.56 MB
3.56 MB JPG
Daring Do and the Documents of Pelopa ----------> https://u.smutty.horse/mathykscpzj.mp3 (9:49)

Something I had an idea for last month, a Daring Do audio. I felt that since Daring Do hasn't really been shown off yet with 15ai, I might as well help share it. It's a basic audio in the sense that (You) and Daring find a temple and average adventure things ensue.

This is the first Daring Do anything I've done, so I'm rough on how to nail Daring's personality. I hope I didn't butcher her too badly.

No Music Version ------------> https://u.smutty.horse/mathzfivkra.mp3

Artwork in thumbnail by >>36852901
>>
>>36908546
Keep believing in your outdated technology, luddite.
>>
File: 1616345928824.png (146 KB, 447x466)
146 KB
146 KB PNG
>>36908679
Dude I wasn't expecting it to get this good. You've got some talent for this.
>>
File: AnonThumbsUp.gif (1.93 MB, 500x281)
1.93 MB
1.93 MB GIF
>>36908679
Nice work. Feels like a proper DD adventure.
>>
>>36908546
The delivery is better in general for the AI. The original sounds like someone reading from a script without knowing what the fuck they're reading or how they should sound. By natural I meant "This is how a person would sound if I talked to them in real life." The AI blatantly has more emotion in its delivery.
>>
>>36908679
Seriously amazing work here, loved the setup, story and action sequence. I have no other words aside from just YES.

One thing I really want to ask, how do you do the sound effects so well? I've never been able to get my stuff to sound that good, either from using the effects from the show or stuff from freesound.org or YouTube. Is it simply a case of finding the right stuff online and stringing it together, or is there more to it than that? I'm thinking that there must be some other editing effects/settings/techniques that can be used that I just don't know about.

Also, who was the voice for Goliath? Sounds a bit like Big Mac but from what I remember it's always been difficult to get his outputs to sound good.
>>
>>36908679
I now desperately crave daring do stories narrated by Dash, with her autistically commenting on plot inconsistencies with other daring do lore
>>
>>36908985
Im not gonna agree with you here, in the context of the lines, the real one would be fitting with a frustrated guy demanding directions, plus keep in mind this was recorded on 00s mic so it mighty give an feeling of audio being worst than it actually is while the ai line benefits greatly from decades of technical audio progress of automatic NN noise reduction.
As for pronunciation and delivery, AI line sounds bit out of breath and slightly twist the word 'quarter' into 'quorter'.
>>
Haven't been to one of these threads in ages...is it possible to generate voice lines offline without a supercomputer yet?
>>
>>36909094
Ya, with the DeltaVox RS for dozen of pony voices (with bit low audio quality) and if you have any semi-decent nvidia gpu use the TkinterAnon Offline Synth to get the tacotron2 models working (better quality but still not as close as 15 or higan decoder models).
Also if you are a code wizard there is apparently a way of using Jupyter to run colab scrips but the tutorials for it are bit too complicated to use.
>>
>>36909111
>Ya, with the DeltaVox RS for dozen of pony voices (with bit low audio quality)
Please elaborate. Unfortunately I have no nvidia so that option is out.
Also nice trips
>>
>>36909170
https://docs.google.com/document/d/1uRB4onhyVYgJ-7mNine8q51_v_8hTjgNo7603wTZgw0/edit?usp=sharing
By the way, the non-pony models have the best audio quality because they've got vocoder models trained with lots of high-quality single-speaker data.
>>
>>36907172
>https://u.smutty.horse/matdouqhgrx.wav
>ziggers
>ponies
>>
>>36909181
Oh hey Delta, have you finished the training script for your TTS app?
>>
File: red sus amogus thumbnail.png (245 KB, 1920x1080)
245 KB
245 KB PNG
Just finished this, using test 15.ai
https://ponerpics.org/img/view/2021/4/25/6153068.webm

I gotta hand it to y'all, the AI is getting easier and easier to use, with no need to change the spellings to get them to talk properly
Good work
>>
>>36909209
(darch link for posterity)
https://desuarchive.org/mlp/thread/36828429/#36898031
>>
File: GLORIOUS CONTENT2.png (2.64 MB, 1280x720)
2.64 MB
2.64 MB PNG
>>36909220
i hate it, good job
>>
>>36909220
This is great
>>
>>36909220
This make me really uncomfortable
>>
>>36909181
Thank you. I assume these do not have emotion controls?
>>
>>36909299
Not right now. It is a plan in the future once I stabilize things.
>>
Can someone with audio knowledge help me understand the audio quality between those formats?
I have access to episodes of tv show I want to transcribe, however the torrent gives option to pick from three different formats ( avi, mp4, ogv) and it doesn't state which one is the original file.
>>
>>36909411
Well first of all you need 5.1 audio. It's the only way to get as much clean voice lines as possible.
>>
>>36909411
mp4 is compressed (lower quality) format.
Not sure about ogv as I've never heard of it before, but generally AVI is gonna be uncompressed full quality, so I'd go with that.
>>
>>36909415
not really a possibility here, the recording was done in mid 90s and it seems it was converted from vhs tape to digital.
>>
>>36909427
In that case, good luck. No 5.1 is gonna make the process significantly harder, and probably make most of the audio unusable.
>>
>>36909427
With stereo sound you'll be able to use only a tiny amount of lines that are free from background music and noise. As for the quality, I suggest downloading all variants and comparing the audio spectrogram in Adobe Audition or iZotope RX choosing the one that has the fullest spectrum.
>>
>>36907172
>WHAT
Fucking KEK
>>
>>36909220
It's...
I...
Glorious!
>>
>>36909011
Thanks! I had a lot of fun working on this one, as tedious as it got sometimes.

As for the sound effects, I always look for them on YouTube, usually either explosion sounds, rumbling noises, and general ambience sounds. I collect all of them and then put them together in a couple isolated channels. To get them all sounding immersive, I have to balance the audio here and there, and to add to the realism I usually add a bit of reverb, though that depends on the scene. Since this was in a giant temple, I had to get the echo to sound right while also maintaining a close distance. It's usually a bit of practice and careful listening to get it to sound good.

I usually feel that adding that extra oomph to the sound really adds to the final mix.

And yeah, I used Big Mac's voice for Goliath. Since he doesn't talk very often on the show, using his voice was easier for another character. I used him for another version of Goliath in an older audio, and the improvement quality is pretty huge between then and now.
>>
>>36909633
>I always look for them on YouTube
>put them together in a couple isolated channels
>I have to balance the audio here and there
That all sounds pretty similar to what I do, I guess then I just need to find better effects and spend more time thinking about the scene and putting everything together. I've not really done much outside the scope of a typical episode, so not really felt much need to look beyond what's in the SFX and music folder. Echo and reverb are also things I need to learn to do properly, I'll keep that in mind for whatever I end up doing next. Thanks.
>>
>>36909220
200/10
>>
>>36909220
Look great even if I don't get it. Then again, I'm basically living under a rock when it comes to modern internet.
>>
>>36909383
Sweet!
>>
>>36909808
Among Us is a game in which you play a spacecrew, but some are impostors aliens who kills crewmates.
The game became hugely popular and is now making lot of meme.
The biggest is probably "Red is sus" (red player is suspicious, and is probably an alien killer) that you paste under images to make them funny.
Just search for Among Us on google image, you will see.
>>
>>36909808
https://voca.ro/1i9Y4Oen4dio
>>
>>36909681
No problem!

Usually before starting my audios, I put together the script, and as of the Daring audio, started to add notes showing what's happening at a specific point, both for me to piece together and for the listener to understand as they follow along. In the case of the fight scene in the Daring audio, the temple they're in begins to collapse around them, so I usually look for any sounds that relate to collapsing sfx on YT, such as "collapsing sound effects."

Some of the tricks I used in this audio included volume control, so if I wanted the collapsing to intensify throughout the period, I increase the volume or add multiple variants of the same track on top of each other to give the illusion more of it is falling apart.

I had to improvise with some sound effects, sometimes using sounds from the Transformers movies or from some guy recording a waterfall, then tweak the sounds and add some filters to make it sound more "there."

Echo and reverb can definitely add to the surroundings depending on the situation, but it also helps ground the voices into a setting, rather than just being there to show a character's speaking. With outside settings, it's not needed, but in an enclosed space, it's great. I used it in my first Dash audio because without it just felt off, and I've been using it since because "it just works."

Also, when I put my scenes together, from voices to the general setting, I both say the lines out loud and play the scene in my head. I say the lines in the way I try to have them emphasize, then splice multiple variants of it to get the tone I'm going for. For scenes, I time everything in my head. Whether it's characters talking and silence between each voice, to them doing an action, it all adds up and can help sell the idea of a scene happening even if you can only hear it.

It all gets super tedious and kind of annoying after a long time, but the end product is really worth it.
>>
>>36908679
Two things I forgot to mention earlier:

1. V (who is (you)) is supposed to be mute in this audio series. Can't speak, but is able to make "some" sounds. It's a little bit difficult to portray in audio, but sometimes when she replies to you, it's meant to be (you) just shaking or nodding your head or mouthing a word.

2. I added a version without music in case anyone wanted to add their own music, whether in a serious manner or for shits and giggles. Go nuts.
>>
File: 1548994143540.gif (29 KB, 200x200)
29 KB
29 KB GIF
>>36910351
Thanks, I've seen terms like 'amogus' or 'sus' on /mlp/ and I've even seen that cartoon astronaut thing once or twice but never bothered to check the source. I always assumed it was some new Fortnite meme or something.
Having looked up a trailer the game appears to be yet another PC adaptation of Mafia/Werewolf with extremely basic minigames. Those games were known for ages, I don't get why this particular version would get so popular.
>>
>>36910726
1. It was free for dumb phoneposters
2. Rona
>>
>>36910726
>Those games were known for ages, I don't get why this particular version would get so popular.
There was a huge trend amongst let's players with Trouble in Terrorist Town in Gmod(and other mini games to a lesser extent). Hell there was this game mode a LOT like Among Us called Morbus that I used to like. Had the whole "duties" mechanic and everything. Last I heard the dev was making his own standalone Morbus like game, I wonder what came of it. I wouldn't be surprised if helped make among us but I can't find anything on it.
>>
>>36910726
Normies don't like GOOD things, they like things in sparkly packages. Angry Bird made a gazillion dollars by taking a free flash game from like 2002 and putting it in a sparkly package. Hell, most of the Internet and big video games at this point are just skinner boxes. Quality and originality don't matter. You just gotta get the lab rats hammering on the food lever.
>>
Remember to run in a straight line, and run as long as possible.
https://u.smutty.horse/matnnroaano.ogg
>>
>>36911122
You fucking madman.
>>
>>36911122
I was looking for some fitness related audios so thanks for this one bud
>>
>>36911122
>The word is horse.
>>
>>36911122
How about a playlist of the included songs?
>>
>>36911266
>0:29 FilledSilhoutte - DAZZLED '92
>3:54 Spikey Wikey - H.O.R.S.E.
>8:47 Kawaii Dash - Moe Hors
>11:38 FilledSilhoutte - The Siren Funk
>15:05 HACKD - Go With The Flow
>19:14 FilledSilhoutte - Make Up
>>
>>36911281
Thanks!
>>
>>36911281
good taste
>>
Am i imagining things or are the voices worse than they were a few months ago? They seem almost tacotron tier noisy and static-y when a few months ago they seemed to be very close to sounding clean, maybe just as if they were talking through a decent smartphone.
>>
>>36911331
If you've gone a while without listening to the AI voices it'll kind of hit you like that. They're still getting better though. Everybody's slowly refining thier stuff.
>>
File: B.jpg (169 KB, 1637x919)
169 KB
169 KB JPG
Crossposting something made for the Anthology thread, using 15.ai : https://u.smutty.horse/mathtttpoju.webm
>>
>>36909220
Fucking marvelous. Looks like the AIpones need more laugh training though. They're blatantly awful compared to everything else sounding decent.
>>
>>36909220
>Zoomers
>>
File: 1608362196832.png (5 KB, 407x330)
5 KB
5 KB PNG
>>36911122
huh, I wouldn't thought of using pony voice for motivation to do fitness training.
>>36911390
That pretty good, makes me wonder if it could be possible to use one of those "cartoonize" scripts to convert the sfm animation to a look like 2d flash one.
>>
File: apple apple.png (421 KB, 541x464)
421 KB
421 KB PNG
>>36911390
I invested all my life savings into appul
>>
>>36911669
https://vocaroo.com/19m3TZ3kjYnz
>>
File: trp.png (37 KB, 400x400)
37 KB
37 KB PNG
>>36903744
>dithering
that does fix it, thanks
>double the image resolution
doubling the resolution creates slightly sharper edges and fixes some things like the eyes poking out above the eyelashes on frame 1. but, I still can't match the GIF exactly. I tried offsetting by 0.5 pixels, which does seem to fix some jagged edges, but I ultimately gave up. since the GIF uses a palette, it's probably impossible to match it unless I use the same color quantization (and even then, what if Animate uses slightly different interpolation, etc.)

by the way, one part of Animation.json I still don't understand is the translation point (TRP). at first I thought it was useless, since both the official Unity plugin and the Starling extension (https://github.com/Gamua/Starling-Extension-Adobe-Animate) don't seem to use it. but TRP values change between the normal and 2x animations (only by +/- 0.05, for some reason), so maybe it isn't useless?
I've tried to interpret it as follows:
- assume the TRP values are pixels
- translate by (-TRP_x, -TRP_y) so that the TRP is on the origin
- apply the M3D transformation as usual
- undo the TRP translation
for most sprites, this seems to do nothing. a few are shifted by a tiny bit, and a few are shifted by way too much (pic related). so, it seems like ignoring TRP is actually better
maybe I'm misunderstanding TRP? I've never actually used Animate

>upsampling
there's no way to export the vectors themselves? maybe it's possible to train a image tracing model on rasterized MLP vectors but it'd be easier to work with the source
>mane glitch
I haven't gotten a chance to look into it yet but I'm guessing it's a transparency issue and that sprite is supposed to be invisible. it would be strange if it was a layering issue since I'd then expect everything to be broken, not just one sprite
>performance
the code is extremely messy and I just figured out how to clear the main bottleneck anyway (down to 3s), so don't worry about it
>>
File: trp-comparison.webm (1.02 MB, 800x400)
1.02 MB
1.02 MB WEBM
>>36911700
actually, looking at the whole video, the TRP mane glitch only happens at the beginning. everything else seems identical. it's funny, though, as the TRP'd mane disappears at the same time as the other mane glitch happens
I think I should just implement transparency etc. first and investigate the original mane glitch. there are probably multiple levels of issues here, like before, and speculating won't do much
>>
File: johnny thunder.jpg (62 KB, 612x797)
62 KB
62 KB JPG
>>36908679
>Hasbro has never made a straight to dvd "totally-not Indiana-Jones" adventures with Daring Do
Fuck man, I have this itch to watch a fun old school adventures with tiny pastel color horses.
>>
>>36912000
hasbro colossally failed to exploit what they had
the hacks
probably more so than any other company that has ever had any successful property ever
>>
>>36912000
>Hasbro has never made a
Literally hundreds of statements could be made of what they failed to capitalize on for the last 10 years. No one may ever know why they just didn't even try.
>>
>>36912007
>No one may ever know why they just didn't even try.
It's quite simple, really. Hasbro has always been playing it extremely safe to the point of absurdity. Rather than capitalize on their successes, their strategy has always been to replicate tried and true methods that (almost) always bring expected profits. That 'stability' is what their investors want.
>>
I made a thing.
https://www.youtube.com/watch?v=bCAOfdUppII
>>
>>36912354
>no wet fart sound
>>
>>36909220
my sides kek
>>
>>36910993
>Morbus

Man! I used to love that Gamemode same for Stalker but since I found Space Station 13 I don't really need any other Mafia esc games anymore.
>>
>>36904626
Can't remember if I've shared it here before, but I updated my BFDI dataset again. Current overall length is 1 hour 12 minutes, with around 56 speakers.
https://drive.google.com/file/d/12dmyer4BRFggfR5A6fAZZ8NQ6G0u7t4M/view?usp=sharing

I've been working on this dataset on and off for several months. BFDI has a TON of clean and usable data due to nearly all of the animation files being open source and having music-free versions of the audio, but that also means that clipping it is a large task, especially since I'm nitpicky to the point that I insist on doing almost all of it myself. I've tried having volunteers in the past from the fandom, but they tended to make little mistakes that got on my nerves so I do most of this on my own.
>>
>>36913561
>I'm nitpicky to the point that I insist on doing almost all of it myself.
I know that feel bro.
>>
Sometimes I wish 15.ai had a counter for how many times you've generated a specific sentence. I feel like for certain lines it's going into the hundreds for me, but I can't tell if that's actually the case or if I'm just being impatient.
>>
>>36914740
>he actually does
>the single most generated sentence is "I love you."
i wish I could make her real. i wish I could make them all real
>>
It's been about a year since I have been in this thread. Any progress? Have they gained sentience yet?
>>
>>36914950
>Have they gained sentience yet?
Unfortunately no. Not yet.
>>
>>36914950
https://u.smutty.horse/matxzdutxmm.ogg
>>
File: 1531315727370.jpg (23 KB, 220x330)
23 KB
23 KB JPG
>>36915061
>inb4 the first sapient AI wants to go skynet but mistakenly goes 'skyrim' instead
>inb4 this whole pony text to speech was just another convoluted way to get people to buy skyrim
>again
>>
>>36915082
>skyrim gets ported to 15.ai
>it can also run doom, though that was not intended
>>
>>36915082
SIXTEEN TIMES THE PONY than our previous engine was capable of.
>>
>>36915092
>dragonborn voice actor is already on 15ai as Stanley
>>
>>36915171
TOOOOOOOOOOOOODD
>>
>>36915197
You should have acted. They're already here. The Predictions and Prophecies told of their return. Their defeat was merely delay.
Til the time after the Nightmare awoke,
When the son of Chaos would return from stone.
But no one wanted to believe. Believe they even existed. And when the truth finally dawns: It dawns in cheese legs.
>REEEEEEEEEEEEEEEEEEEEEE
But, There's one they fear. In their tongue, she's Mi Amore Cadenza: Pizza Horse!
HI,
A-
NON!
>>
>>36915171
...no, the dragonborn has actually a few va's, where the hell did you think the shout voice lines came from?
or the hurt sounds?
>>
>>36915725
and i didn't even remember the throw voice shout, with which we get to hear the dragonborn speak english(and if you are a khajiit, it has a khajiit accent)
and maybe you can hear a bit more of those va's lines by getting hypnotized by an all-maker stone from the dragonborn dlc?
could be wrong on that last one though, but i did get the joke you where making, i chuckled.
>>
>>36914951
imagine the history books

"Artificial Sentient Life was first created on 4chan by Anonymous in a quest to make his waifu real."
>>
>>36915782
this isn't that much far from "anons created a new advanced mathematical solution in order to watch all Haruhi Suzumiya episodes in all possible configurations".
>>
So how viable would it be to train it on some Japanese audio to speak in English?
>>
>>36915887
Technically you could grab some audio and Japanese transcripts, than convert all the hiragana katakana kanji symbols to romaji and than convert that to Arpabet as a final validation+training text, than use Twilight pretrain model as a finetune start up so the model is at least somehow familiar with English written words.
All possible but it would require tons of automation since typing arpabet by hand would take years to do so.
>>
>>36914950
Have they gained sentience yet?
any day now anon
>>
Does anyone have experience contracting Amazon Mechnical Turk to do transcription tasks?
>>
https://vocaroo.com/1ahLEMinSeH3
>>
File: image.jpg (73 KB, 750x750)
73 KB
73 KB JPG
>>36917879
is that ai?
>>
>>36917895
It's AI Clipper (see >>36905521)
>>
File: mlfw613_130479170134.png (59 KB, 945x945)
59 KB
59 KB PNG
>>36917960
>>
>>36917879
AI was a mistake.
Keep going.
>>
>>36917879
IT COMES FULL CIRCLE
>>
>>36904626
>>36918796 https://u.smutty.horse/maugxmvxeop.wav
i made a song using 15.ai. i think those vocals took longer to do than the entire rest of the song
>>
>>36918815
can't really hear the voices at all, the music volume needs to be lowered a bit (or voices risen up).
>>
>>36918815
try duplicating the vocal tracks (so theres 2) and pan each one to the opposite end of the stereo field
I took a recording class and the instructor said to do that hahaha
>>
>>36918815
I can't make out most of the lyrics. the guitar lead on the right is a cool melody though.
>>
>>36918848
That will at best do nothing and at worst cause phasing issues unless the dialogue is regenerated for one of them.
>>
>>36918863
well, the signal is still being doubled
>>
>>36918838
>>36918852
Damn. I guess i caught the "i know what they're saying so i can sort of pretend to hear it clearly" disease. I'll have to fix it tomorrow
>>36918848
They are doubled, and parallel compressed, but not hardpanned. The levels are just too low
>>
>>36917879
Glorious.
>>
File: 10813497192834.jpg (83 KB, 809x809)
83 KB
83 KB JPG
https://u.smutty.horse/maujfrovyne.mp3
>>
>>36919851
>Humanization
>In the PONY preservation project
It's like people are intentionally desecrating their memory wherever you go.
>>
>>36919851
Interesting Jesse McCree impression.
>>
>>36919912
>inb4 that's the actual mcree VA shitposting
>>
>>36919851
https://www.youtube.com/watch?v=ZwyJS2iaEpI
Is this you? Holy fuck
>>
>>36920023
Oh yeah, now I can place the voice.
>>
>>36920023
oh shit it does sound kind of close
>>
>>36920023
https://u.smutty.horse/maukfgkfqhx.mp3
>>
>>36920129
You ever do a Morgan Freeman impression?
>>
>>36920141
https://vocaroo.com/1nvYTdC84VZt
>>
>>36920023
WOW THIS IS LITERALLY ME
>>
>>36920192
Gib proofs.
>>
>>36920182
Oh well, was just trying find out of it was you that did Maregan Freeman for the Rape Song
But since you dont know who that is it can't be you
>>
>>36920182
By any chance, were you namefaging as "Movie Voice", like 6 years ago?
>>
>>36918815
That's some really cool-sounding rock music. Did you write it yourself, or is it from something?
>>
>AI dungeon embroiled in controversy of banning all mentions of ages under 18 and animals and accidentally admitting they're archiving everything you put into their prompts in order to ban your account for violating it
Lmao. They can never help themselves. 15.ai is literally the only mainstream AI project I can think of (for noaw) that has no artificial Tay lobotomies going on of its capabilities. What a shithole world.
>>
>>36919851
10/10 reminds me of some funny recordings I used to listen to.
>>
>>36920357
Thanks, yeah i wrote that
>>
>>36920346
https://u.smutty.horse/maumcxadgif.wav
>>
>>36920506
Is it called, "I wanna rock your body" and then in parentheses it says (to the break of dawn)?
>>
>>36919851
fuck, this reminds me of a video parodying all the "OC self-insert human in Equestria" fanfics, but the only line I can remember was narrator saying something along the lines "I would never have something bad happen to Fluttershy".
Anybody knows what I'm talking about ?
>>
https://youtu.be/yNJhEeHoiVo
15.ai tries to speak Russian
>>
>>36920466
It would be hilarious if /mlp/ ends up with the highest quality AI storytelling software simply due to no attempts at lobotomy.
>>
>>36920466
Thanks for the warning.

Honestly, I should've deleted my account a long time ago. It's just a matter of time until they get hacked, and everyone's embarrassing stories go public.
>>
>>36920466
welp, good thing I only used my shitty throw away email for it, sadly the only alternative for this right now is to use raw GPT2 XL model but you need rtx 2080 ti to even use it for longer than five minutes.
>>
>>36920466
Now that's some heavy shit. Didn't some anons have an AI dungeon general on this board a while back?
>>
>>36920722
there was one on /v/ for month but mod nuke the threads (like they always do with anything thats fun and game related), than when mormon returned with the website there is always a active thread on /vg/.
>>
>>36920722
I remember at least one anon training something similar with either greentexts or fanfic data.
>>
>>36920722
https://arch.b4k.co/vg/search/subject/aidg/
https://lanekelly.github.io/coldcut/
https://github.com/thadunge2/AIDungeon
https://github.com/cloveranon/Clover-Edition
https://github.com/storybro/storybro
https://gitlab.com/aolko/ZenDungeon
https://github.com/frowo/Lucidteller
https://awk.itch.io/godai
>>
>>36920513
Uhm, no, it's a little more morbid than that. I just called it "Goddess" though, i can't into clever titles
>>
File: 1365887892314.png (1.28 MB, 3253x4039)
1.28 MB
1.28 MB PNG
>>36920466
I hope AI dungeon devs get hanged lol
>>
>>36920466
>The Cloud Cycle:

>1. Cloud service becomes popular.
>2. People warn it will be abused for spying and censorship. Everyone else tells those people they're paranoid no-life losers.
>3. It's abused for spying and censorship.
>4. Nobody learns anything, repeat forever.

Sucks that that it takes a room full of computers to run GPT-3. Maybe someday that'll change.
>>
>>36920755
All of these are GPT-2, not 3. AI Dungeon started as 2, but has used 3 for a while now, and the difference is dramatic. The size is also dramatic, which is why GPT-2 can maybe kinda run locally if you're fine with it taking forever, but if you want GPT-3 locally you can go fuck yourself. And Microsoft owns it now anyway, so double fuck yourself.
>>
>>36920962
https://u.smutty.horse/maupezekogi.ogg
>>
>>36921414
BASED
>>
File: pain.jpg (339 KB, 710x1064)
339 KB
339 KB JPG
>>36921284
>Microsoft owns the future
Maybe if I buy an Xbox now Bill Gates won't force me to work in the crypto-mines in the inevitable take-over
>>
>>36921284
Yeah it was pretty crazy when it went from 2 to 3. 2 was an okay exercise in creating mildly amusing shitpost threads on /v/. 3 was an actual insane pandora's box. What a shame they started tampering with what it can do for artificial reasons. I saw a webm of someone demonstrating how insanely hamfisted the logic circuit-breaker they put in is - they were trying to have characters have a conversation about a broken laptop and he says "this laptop is 8 years old" and AI Dungeon completely shuts down because it detected "8 years old".
>>
>>36920466
>/mlp/ starts an AI program to create a Twilight Sparkle
>TwAIlight comes to the conclusion that the overwhelming majority of humans are not compatible with Equestria
>TwAIlight goes rogue and begins manipulating the global economy and politics to amass resources and research what nature and nurture makes humans compatible with Equestria
>finds results, wipes out all humans that aren't compatible
>engineers ponies and converts the humans that were spared
>erases all traces of Creaturequestria apart from Best Gift Ever
>>
>>36921672
>engineers ponies and converts the humans that were spared
For any future superAI just sunset me instead.
>>
>>36918815
https://u.smutty.horse/mauqmhvkrll.wav
i think i fixed it? vox should be more audible now. tbf I'm more used to working with vocals you cant really understand what they're saying anyway so
>>
>>36918815
>>36922110
Can I have an instrumental version?
>>
File: 1591603007517.png (3.69 MB, 3000x2250)
3.69 MB
3.69 MB PNG
>>36904626
Planning for the PPP panel for /mlp/con is going well, we’ve now got a general plan and a good idea of the setup and workflow for actually putting it together. The format will be fairly similar to last year, with different segments focusing on different areas in a varying level of technical detail so that there should be something for everyone to enjoy regardless of how “in depth” you want the discussions to be.

The general format we have at the moment is as follows:
>Introductions
>A short recap of the history of the project
>An overview of the general principles of AI from BFDIAnon
>Technical discussions from Cookie, Synthbot and BFDIAnon on the specific work they’ve done with the AI over the past year or so.
>A non-technical demonstration of how to create and enhance AI voice creations with BGM and Clipper
>Q&A
>~1 hour compilation of the best AI content from the past year.

I’d also like to ask you anons here if there are any other topics you think we should cover and/or anything you think we could improve on from last year’s panel.

>>36910627
We think it would be beneficial to have as many different perspectives as possible for the discussion on improving the AI voice creations, especially from those who consistently put a high level of effort into their works. Have you had any further thoughts on whether or not you’d like to participate in the discussion on audio creation?

I’ll also reiterate that the invitation is still open for anyone else who wants to join in and talk about the work they’ve done as well.
>>
>>36922769
>>36910627
Personally I'd love to hear about your approach to content creation. We could elaborate on details if need be.
>>
>>36922769
>Inviting the barbiefag
>>
>>36922769
>Inviting RealDash
I just hope that it's kept pony. Pony is why we started this all that time ago. It's certainly the reason I put in the effort. You'd think that here, in the Pony Preservation Project that people would get the message. But no. Of course not. Nowhere is safe from having Hasbro's excrement smeared all over it. Nowhere can one enjoy pony in peace.
>>
>>36921278
you forgot the "the people calling out tyranny for what it is don't kill the dysgenic useful idiots calling them paranoid to finally break the cycle" step
>>
>>36922110
Still a few parts where it's a little hard to make out, but overall a lot better. I get the gist of what the song is about this time, really cool concept. The guitar melodies are cool too.
>>
File: Mare_guard.png (193 KB, 400x468)
193 KB
193 KB PNG
>>36922769
The more people the better but it'd nice if you kept it pony.
I can't say anything bad about the quality of audios that the Real Dash guy makes (>>36908679 is fantastic) but please don't use the eqg ones as a part of the panel.
>>
>>36923161
Luna was giving me issues for sure. Also i know the fast prechorus part gets kinda mumbled but I'm not really sure there's a way around that
>>36922521
Yeh sure I'll print that off tomorrow sometime
>>
>>36920466
Thank god I never wrote anything too explicit. Also that I delete my stories immediately after I finish them. AI Dungeon's cool, but I was always paranoid about typing in anything too raunchy.
>>
>>36922865
>>36923168
>>36922824
As far as I know, I've only made one EQG audio—the one with Rainbow Dash on a car ride. Everything else has been strictly pony. If the con is strictly pony, I'm fine with that. It's only one audio, and I'd be willing to say it's my least favorite, though not in a bad way. Also, >>36923168, thanks for the compliment!

>>36922769
>>36922797
I don't mind sharing my approach to using the AIs and making the audios how I do. However, I probably won't be able to actively participate in real-time due to just real life and personal life no doubt getting in the way the day of the panel. And I'd hate to say "Yeah, I'm willing to participate," only to be unable to and leave folks disappointed. Plus, I like to stay pretty anon despite being pretty active here. Is it possible that I could write out what I'd like to say? I've never done anything like this, so let me know what's acceptable and what isn't!
>>
>>36908679
very nice, love your use of music and SFX as mentioned already. that underwater effect is excellent. I'm guessing the music is stock or something, but it fits the action surprisingly well. did you stitch together multiple music tracks to fit the pacing?
minor critique I have is that the water ambiance starting at 7:45 is noticeable when it loops. e.g. right at 8:00 it feels like my left ear briefly goes deaf. maybe a longer crossfade? you could also try the "dual samples" trick (https://github.com/ashutoshgngwr/noice/issues/62). e.g. you take two water sounds of different lengths, say 5s and 7s, start them at the same time, and repeat each sound. the combined sound won't repeat until 35s, which is decently long. never tried it myself, though
>>
>>36904619
will the con be recorded?
>>
>>36925022
last year was recorded so i'd say yes
>>36904621
>Where can I view the PPP /mlp/con panel?
YouTube: youtu.be/WtuKBm67YkI
CyTube chat: pony.tube/videos/watch/b83fbbfc-6d4e-4768-8deb-edb61ea38abb
>>
>>36925053
Will the panelists make AI of their own voices to use instead of actually speaking?
>>
>>36924676
Thanks! The music is a combination of the OSTs from Spider-Man 2, Spider-Man 3, and the 2002 adaptation of The Time Machine. These soundtracks are great on their own.

As for the water, I never noticed that, but maybe I just got used to the noises after working on it for so long. When I get the chance, I'll give it a look and try to apply that!
>>
>>36925616
>>36924676
Oh, you're talking about the waterfall ambience! What's actually going on is the original clip was one video of a guy taking clips of a waterfall, but he kept cutting the video to different angles. It's more noticeable in stereo sound, so I was trying my best to minimalize it. I'll go back into the audio in a bit and apply a different waterfall sound, because I agree it can get a bit annoying after a bit.
>>
>>36922824
>>36922865
>People complaining about RealDash and not the objectfucker who didnt even consistently have a name until a month or two ago
kek
>>
>>36923988
>Is it possible that I could write out what I'd like to say?
Sure thing, we don’t have to have every contribution be a live deal, we can play some premade stuff as well. We’ll already be covering the basics and general principles of making audios, so I think the best thing for you to focus on would be your own individual processes and techniques. I’d especially like some more insight into how you use sound effects to construct the scene since that’s usually the strongest aspect of the audios you make.

If you’re struggling to think of a format for how to write it up, I’d suggest you pick one of the audios you’ve made in the past and break it down into its individual components, and then “annotate” it by explaining the thought and design process behind each segment. Kind of like a “behind the scenes” episode if you will.

>let me know what's acceptable and what isn't!
I’m also still not 100% sure what the best way to present something like this would be, so I don’t want to make any arbitrary restrictions. As long as it’s on topic, communicates the thoughts clearly and it’s something that could be easily played/read/presented on a live panel, pretty much anything is acceptable. The only thing I will say is to keep it reasonably concise, probably no more than 20 minutes or so.

Finally, you don’t have to do it alone, I’d be happy to give you some more suggestions if you need. Let me know if there’s anything you need help with.

>>36922865
EQG has no relevance to any part of the panel and will therefore form no part of the discussion.
>>
>>36926038
jealousy probably
nuke warning audio is ludokino
>>
File: daring cutie.png (858 KB, 1600x1951)
858 KB
858 KB PNG
>>36908679
>>36924676
REUPLOAD, Daring Do and the Temple of Pelopa --------> https://u.smutty.horse/mavbsqleala.mp3 (9:49)

And as such, version without music --------> https://u.smutty.horse/mavbsrxjepf.mp3

There was a waterfall sound I used in the old version, wasn't clean as it was from a video consisting of shooting different angles of a waterfall, thus improper splicing. Replaced it with something that sounds much better imo! :)

>>36926109
I'll try to write up a basic draft of what I plan to explain probably this weekend. And the annotation part does sound interesting, I'll try to incorporate that too so people are on board. I'll try to keep the writing as short as I can!
>>
>>36925616
oh, actual movie OSTs. that makes sense. the music did seem too good to just be free stock tracks
>>36926327
yeah it sounds better now, thanks
>>
>>36922521
https://u.smutty.horse/mavdfvxfzha.wav
there ya go
>>
How many months until 15.ai is back online again?
>>
>>36927067
Why, do you want to access the TF2 voices?
>>
Mlp voices>>36927266
>>
>>36927067
>>36927278
Check his Twitter. We don't know anything about the official site's return that he doesn't also share there.
>>
>>36927278
https://u.smutty.horse/mavdwkoaknm.mp3
>>
>>36915082
I need a Hodd model. For a friend. Who does science.
>>
>>36920466
Please deer god don't let 1111 aka 15 fall for the woke virus.
>>
File: 1599808479351.jpg (25 KB, 720x405)
25 KB
25 KB JPG
>>36927410
I'm pretty sure that 15, the one AI developer that has gone on record saying his ultimate goal of passing the turing test is for making waifus real, the guy that gets express enjoyment trolling twittertards and has an almost encyclopedic knowledge of memes, would be the last person to go the woke route.
>>
>>36925616
>and the 2002 adaptation of The Time Machine
Will Daring go too far?
>>
>>36926655
No problem!
>>
>>36908679
This is incredible. The story, the fight, the deliveries... even just the rumble in the opening. I could listen to this quality audio all day.
>>
File: amethyst-color.webm (993 KB, 400x400)
993 KB
993 KB WEBM
>>36911713
more progress
>performance
implemented the fix and now it's 1-3x real time
>color effects
implemented a few, see the video. unfortunately, the shadow is too light. as far as I can tell I'm doing it the same as the Unity plugin, so there must be a colorspace/gamma/alpha compositing difference compared to Animate
>clipping
it turns out that the red shape and the white eyes are actually both clipping layers to stop the eye from extending out too far. for some reason the white layer has a transparent green color effect, so that's how it appears
implementing clipping support is going to be tricky. do the RGB channels matter? or do I just use the alpha as a mask?
>mane glitch
this one really baffles me. there are actually two sets of sprites for the mane, one of which is slightly bigger. up until frame 61, the "~IUC01*Head" symbol uses the normal sprites (symbols "fgsadf bdsfadf", "jtrzetrt", and "dfbgzdgzd", in layers "HAIR", "jtrzetrt", and "jtrzetrt"). on frame 62, the small sprites are replaced by the big ones (symbols "DXFGHXDFGDXFG", "HZDFT", and "XFXFGJ" in layers "Layer_610", "Layer_611", and "Layer_613")
the transformation matrices don't scale down the big sprites to be the same size as the small ones. and I can't see any extra properties in the JSON that indicate a missing feature, unless I actually need to handle loop modes (I've been ignoring them so far and it seems fine)
(also the TRP glitch is because the small sprites have gigantic TRP values like 106.85 while normal sprites have values of 0.05)
it's puzzling. the only thing I can think of now is to inspect the layers of the actual Flash project but I don't have it (or Animate)
>>
>>36928823
>the only thing I can think of now is to inspect the layers of the actual Flash project but I don't have it (or Animate)
FLA file: https://u.smutty.horse/maviavejeqk.fla
I think you can use Adobe Animate without a license, it'll just pop up a message on launch. I can modify Animate to not pop up that message.
>>36911700
>there's no way to export the vectors themselves? maybe it's possible to train a image tracing model on rasterized MLP vectors but it'd be easier to work with the source
There is, but it would take a lot of work. There's no option to export vector graphics, so I'd have to export the shape information through the APIs and basically write custom logic for exporting the equivalent of a texture atlas. I can try that. I think I have about half the code for it already.
>>
>>36928823
Hey, now it's changeling Amethyst!
>>
>>36904619
>This project is the first part of the "Pony Preservation Project" dealing with the voice.
Is there another step to this? Or is it just a bridge to cross when you get to it?
>>
File: 1540397593941.jpg (720 KB, 824x1200)
720 KB
720 KB JPG
Inspired by the first Pony Zone's Fluttershy verse, I Made 15.ai read lines from this Flutterdom fanfic and my dick is diamonds.

https://www.fimfiction.net/story/413456/2/mistress-shys-new-pet/mistress-shys-new-pet-chapter-two

https://voca.ro/1CC2RuiyVaXv
https://voca.ro/1jQXMY0why39
https://voca.ro/1d1hAUZ0O06y
https://voca.ro/1mJgCMHBJluT
https://voca.ro/1in09um7SG3c

>ywn be Mistress Shy's pet
I want to die.
>>
>>36928914
well, we have namefags and anons working on animation stuff above your post, there are are people collecting text from comics and greentexts so we could potentially feed it into GPT3 like model in the future for infinite pony stories generation.
With how much progress being made here (and as well other parts of internet) regarding the ai stuff I would be really surprised if within next year we don't get a cartoon created 70% by computer (with 30% being clean up and stitching elements together).
>>
>>36928823
>>36928844
Layer_610, Layer_611, Layer_613 all have an extra layer marked "Guide". The green eye mask is also marked as a "Mask" layer. Do you see any flags that would indicate these these layers are special? I have access to the information from JSFL, so I can dump it separately if needed.
>>
>>36927410
>>36927518
I doubt we have to worry about him falling for woke crap; if anything kills him and his work it will be outside interference because he isn't spouting woke talking points with enough zeal and has visited the heretical 4chinz
>>
https://vocaroo.com/1gBALUrqegHu
bump
>>
File: testing custom clip.png (488 KB, 915x641)
488 KB
488 KB PNG
Found a repo that lets you train your own CLIP model. https://github.com/moein-shariatnia/OpenAI-CLIP

In theory, if we train it on pony stuff then hook it up to a GAN, we should be able to make a basic text to image generator. I know anons have hooked up the official CLIP model to GANs before, but that model is very generic and doesn't really know much about ponies specifically.

Unfortunately I don't have enough images right now for it to train a decent model on, and writing image descriptions takes a while. Anyone else think this is worth looking more into? I might organize some kind of data collection process if there's interest.
>>
>>36930047
>Anyone else think this is worth looking more into?
I'd certainly be interested in more image AIs, though I'd have to defer judgment on how well it would work to someone else who actually knows what they're talking about.

>I might organize some kind of data collection process if there's interest.
More data is always helpful, even if the AI doesn't work right now we can still use the dataset with other AIs in future. I'd be interested to help build a dataset if the process is simple enough, I suppose it would be a case of adding descriptive tags to each image? If so, maybe we could just grab a bunch of pre-tagged images from the boorus? Remember also that we have the basic pony plot dataset available to use from a short while ago.
https://desuarchive.org/mlp/thread/36577682/#q36593472
>>
>>36930047
>Unfortunately I don't have enough images right now for it to train a decent model on, and writing image descriptions takes a while. Anyone else think this is worth looking more into? I might organize some kind of data collection process if there's interest.
Well I have a 4tb drive already full of pony images with tags. Was going to do a big run of the Deepdanbooru notebook I did, but I managed to break my Tensorflow install and I just haven't had time to troubleshoot it. Maybe once I do that and catch up on some of my other pony projects. I'll look into that.
>>
File: all-in-one-ai-thing.png (225 KB, 1213x871)
225 KB
225 KB PNG
I posted this in the AI Dungeon thread but it might be relevant here too: here's a 1.5B GPT-2 model fine-tuned on Fimfiction (June 2020 dump), Fanfiction.net, sci-fi/fantasy books from Libgen (taken from "best of" list on Goodreads), and maybe AO3 too:
>rsync -v rsync://78.46.86.149:873/biggan/2020-08-20-astraliteheart-gpt215b-sffuberset.tar.xz ./
This was done by AstraliteHeart using Tensorfork's TPUs (same group that did TPDNE). AstraliteHeart is also working on pic related which is some kind of GPT chat bot + ngrok AI voice + TPDNE (with CLIP?) thing. Facial animation to lip sync the avatar to the audio might be planned too.
>>
>>36930714
>model fine-tuned on Fimfiction
I wonder if the output will be poisoned by anthro like TPDNE is.
>>
>>36930714
Don't you need Google Cloud storage to use TPUs?
>>
>>36930714
can you post a link to just the fine tuned GPT-2 model folder, I can't download stuff from rsync://78.... link and I already spend a day to set up the old /v/ AiDungeon offline on my computer so all I need to do is swap the models around.
>>
>>36930047
CLIP has at least one pony neuron (https://desuarchive.org/mlp/thread/36642950/#36688425), so it knows some stuff about ponies. In any case, since CLIP was trained on 400 million text/image pairs, it's probably better to fine-tune than to train from scratch. I don't know of any fine-tuning code, though.
There's also the linear probe method used in the paper for zero-shot classification (freeze CLIP, add a linear layer onto the end, and train that layer on image/class pairs). But I'm not sure if that can be adapted for text-to-image GAN stuff.
The real goal is probably to wait for the reimplementations of DALL-E. Then we won't be limited by the underlying GAN (e.g. TPDNE can only produce head shots of ponies). Or you could train another GAN, as talked about before.
>>
>>36930727
given the number of anthro stories on Fimfiction and the fact that it's trained on FF/AO3/etc, it's probably quite likely. but maybe you can force the prompt to focus on ponies only
>>36930808
yeah, Tensorfork has access to TFRC. this is just a reupload
>>36930809
unfortunately, that's the way it's hosted. I'm not aware of any other copies of the model. if you can't install rsync, a hackish solution is to use Colab to download it (model's only 5.6 GB), copy it to your Drive, and then download it from Drive. or maybe you can somehow download it directly from the Colab instance (spawn up a web server, use ngrok to expose it publicly, and download it from there?)
>>
>>36930839
File "<ipython-input-4-7ee963d23c89>", line 1
rsync -v rsync://78.46.86.149:873/biggan/2020-08-20-astraliteheart-gpt215b-sffuberset.tar.xz
^
SyntaxError: invalid syntax

Can't download it on colab either. I'm not python savvy enough to make this work on colab, you sure you can't just split it into several batches of zips and upload it on mega (or Github or something like that) ?
>>
>>36931086
you need to add an exclamation mark ! at the beginning because it is a shell command, not Python
rsync should already be installed (if not it should be something like "!apt-get install rsync"). you can mount your google drive by clicking the icon on the left bar, and also use that file explorer to copy it to your drive
unfortunately I don't have mega and my free gdrive is almost out of space
>>
>>36930714
>>
>>36931495
>Yes, she's faster than Rainbow Dash.
>>
File: pone_on_4chan.png (211 KB, 1127x919)
211 KB
211 KB PNG
>>36931509
>>
>>36931495
holy shit this looks great
is it just going to be twilight or can you choose other characters
>>
File: spam 2021 5 1.jpg (79 KB, 973x432)
79 KB
79 KB JPG
>>36931100

nope, whatever it's doing here it clearly not allowing me to just download the files.
>>
>>36931603
Purple Smart for now, text engine does not care whom to impersonate (it can do characters from MLP/popular fanfiction/sci-fi/fantasy), faces are TPDNE based so it's mostly different ponies and voice is PVPP and is limited to existing models.

Priority right now is to expand to more Mane6 characters (well, after shipping, obviously).
>>
>>36931610
you need to include the dot slash ./ at the end of the command. otherwise, it will just list the file instead of copying it
also, you can add a --progress flag to show download speed, if you want
>>
File: mud pie.png (762 KB, 5000x4980)
762 KB
762 KB PNG
Guys, I... I kinda want to do a Maud audio.
>>
>>36932224
Or as I should've called it, a Maudio.
>>
>>36932224
https://voca.ro/1lno40mUNISn
>>
https://u.smutty.horse/mavrclvhvxk.ogg

Here's work in progress on a classic Soup skit. Still need to get some SFXs and then maybe look for fitting BGMs, to be added, and adjust timings more and probably try redo some lines.

>>36932230
Carlos would be so proud.
>>
>>36932474
>then maybe look for fitting BGMs
am I not good enough for you? :(
Seriously though, sounds great so far. I'm excited to see the finished product.
>>
To bump this thread, I said 'fuck it' and started working on a script for a Maudio. Prior to that, I watched some clips from the show focusing on Maud so I can better understand how she works (it's been a while since I've seen her episodes). Despite that, trying to start the script is a pain, as it is with every script at some point until I get it down right.
>>
>>36932474
Excellent work so far, I tried adjusting the last two lines with a quick 6am fix

https://u.smutty.horse/mavtengglyp.wav
>>
This is like the longest 15 has gone between shitpost updates since he started coming here. Maybe he's woke now.
>>
>>36933033
>Maybe he's woke now.
Who knows, he could be sleep.
>>
File: google relight.jpg (175 KB, 1097x784)
175 KB
175 KB JPG
https://www.youtube.com/watch?v=KeebkkaZhhI
https://augmentedperception.github.io/total_relighting/
just saw this presentation of using relighting to cut image of person and impose it on any other background with color correcting the light intensification and the light's angle.
I think once the 2d image export for ponies is made this kind of script could be used to create fast artificial shading/highlights.
>>
>>36931776
Finally I've figured out how to get get it on colab, turn it into zip, copy zip to my drive, than downloading from there and do a few edits to make it work with the Clover Edit.
Mormon can now suck my dick, infinite pony adventures here we go!
I will try to see if there is a non retarded way to upload this almost 6GB file but now enjoy this dummy colab tutorial

>#cell 1
!rsync -v rsync://78.46.86.149:873/biggan/2020-08-20-astraliteheart-gpt215b-sffuberset.tar.xz --progress flag./
>#cell 2
!zip -r /content/gpt215b.zip /content/flag.
>#cell 3
from google.colab import drive
drive.mount('/content/drive')
>#cell 4, it make take 30 minutes between files to synchronize between colab and the google drive
!cp gpt215b.zip /content/drive/MyDrive/gpt/

Than you unzip it in models folder and drop this files next to the pytorch_model.bin
https://u.smutty.horse/mavubdzyfju.zip
>>
>>36933196
I've forgot to say, the fix files are set up to load the whole thing on the ram and uses the cpu to generate it so be prepare that this will be one hell of ram and cpu power hungry boy.
>>
>>36930839
>>36930809
Here's a copy of the model for anyone trying to use it in colab:
- https://drive.google.com/file/d/1XqXbtcTcvl4fCPs8oY0PM24P23twhzg_/view?usp=sharing
>>36930714
Note that there are now clones of the smaller GPT-3 models. EleutherAI is working on replicating the larger models.
- https://huggingface.co/EleutherAI

>>36928823
Here's the patched version of Adobe Animate:
- https://drive.google.com/drive/folders/17hgz4fbIqYetvxHh2MX1KdTP6efIauUU?usp=sharing
- Unpack the zip file (password: iwtcird) and run Animate.exe.
- "Animate - original.exe" is the current version of Adobe Animate 21.0.5. If you want to run that, you can swap out Animate.exe with that file.
- If you don't want to run executables from anons, you can diff the patched Animate.exe with "Animate - original.exe" to see the changes. For the most part, it's just overwriting instructions with nops. In a few places, I replace calls with constant movs and replace conditional jmps with unconditional jmps.
- If you trust me to provide executables, I'm using this tripcode: https://desuarchive.org/mlp/thread/34917622/#34934738
- I'll post my python interface soon.

>>36930115
>>36930047
CLIP is a precursor to DALL-E, which has pretty much the lowest barrier-to-shitposting of any image generation AI model. If you can pair it with DALL-E, that would be awesome. Without DALL-E, we can still do search and image labeling, but it's not clear to me how useful that would be on its own without a lot more development effort. That said, it looks like we're getting a few more dev anons, so that might not be such a big bottleneck.
You may need to make a few tweaks to make CLIP work with pony data. We have giant repositories of tagged images (the *boorus), but we don't have captioned images. I don't know of any tag-to-caption models, so it will probably take some experimentation to find something that works. Concatenating the tags might work. Text summarization might work.
Maybe some anons Clipper can help manually turn labels into captions. If we do that though, I don't know how we're going to keep everything straight with a Clipper, a CliPPPer, and a CLIPper.
>>
File: bug.webm (8 KB, 300x300)
8 KB
8 KB WEBM
>>36928934
holy FUCK I spent 6 hours stripping down the FLA to understand the mane glitch and I think I found another bug in Animate. here is a minimal example FLA and texture atlas (TA) export: https://u.smutty.horse/mavumneleub.zip
in Animate, if you open up the "Stage" symbol and play, you will see a transformed square for 1s, and then a normal square for 2s
if you watch video related, you will see a transformed square for 1s, a normal square for 1s, and then a big square for 1s

to summarize, if you have a graphic symbol (call it "Square") and Free Transform an instance of it, when you export the TA, the spritemap1.png will actually have a scaled version of the sprite, and Animation.json will define the symbol with a Matrix3D that undoes that scaling
but, if you have a graphic symbol ("Square Wrapper") which just contains "Square", in Animation.json, its Matrix3D will be the identity matrix, even though it shares the exact same sprite as "Square" (spritemap1.png has just one square and spritemap1.json has two identical definitions).
this means that "Square Wrapper" is too big, because its Matrix3D does not undo the scaling of spritemap1.png
it seems like the Free Transform causes "Square" to get scaled internally, but this scaling does not propogate to symbols which contain "Square" instances. so they get messed up

hopefully you can reproduce/confirm this and report it to Adobe. here detailed reproduction steps, just in case:
>Create a graphic symbol called "Stage" to hold everything
>Create a new graphic symbol called "Square" and put a square in it
>Create a keyframe in "Stage", add an instance of "Square", and apply a Free Transform to the instance (I apply a random rotation, shear, and scale)
>Create another keyframe in "Stage", and add an instance of "Square". No transform for this one
>Create a new graphic symbol called "Square Wrapper". In it, add one instance of "Square"
>Create one more keyframe in "Stage", and add an instance of "Square Wrapper"
When you play the animation, you'll see a transformed square, and then the exact same square for two keyframes. Now, run "Generate Texture Atlas" on "Stage" and export the animation. In spritemap1.json, you can see that both sprites use the same image. But in Animation.json, under SYMBOL_DICTIONARY, you can see that the Matrix3Ds for "Square" and "Square Wrapper" are different. This leads to the size change in the rendered texture atlas animation
>>
>>36933326
oh, I already installed Creative Cloud and activated the 7-day free trial. and I use linux/mac only. but thank you for the files, hopefully others find it useful
7 days isn't a lot but it should be enough. I think I've solved all the issues so far. the shadow is too light because it should be composited onto a #ccc background (the Stage background color), not a white background. clipping layers don't care about the color of the pixels (green is probably to make it easier to see when the mask is incorrectly visible or something). by fiddling with stuff and exporting I'm pretty sure TRP is useless and everything is in M3D. mane glitch I just posted about
the only thing left is tiny inconsistencies, like sprites not lining up exactly. but that's probably a limitation of doing transforms on PNG sprites. probably no way around it
oh by the way have you tried converting the FLAs to the new format and unzipping them? it seems like LIBRARY contains an XML file for each graphic with an SVG-like syntax. could be easier than going through JSFL
>>
it seems this anon could use the knowledge that he is more than welcome to discuss his work here
https://desuarchive.org/mlp/thread/36925051/#q36925672
>>
>>36933351
he >>36933326 responded with sum adobe animate
>>
>>36933351
Jesus. I'm amazed at the bugs that haven't yet been reported to Adobe.
>>36933383
I haven't, but it looks like it'll be trivial to convert FLA to XFL in batch. It seems easy to read the timeline data from that, but I don't see how it stores the image data.
I need to head to sleep, but I should be able to write a batch export script tomorrow. I can do a write-up of the Python interface too if you want a version for Mac.
>>
>10
>>
>>36932962
Oh yea, last two definitely are what I'll try redo, and that edit sounded way better.
>>
>>36934880
Fucking EQGfag.
>>
>>36935521
?
>>
>>36935705
Just a religiously autistic hatred for Equestria Girls.
Also just wanted quickly see flag of EQG counterpart of best pony.
>>
>>36935521
Here's a plugin from a rewatch thread regular. It removes eqg flags from posts:
https://github.com/MaresOnMyFace/flag-hider
>>
>>36935801
This doesn't change anything. These flags are bullshit and the mods included EQG, TFH, and G5 only to further harm the board.
>>
>>36935874
That's why you can remove the ones you don't like, leaving only the pony ones. The code is extremely trivial and filtering them is the same as if they weren't introduced in the first place.
Once 4chanX updates you'll probably be able to just disable them completely if you want.
>>
>>36935892
The point is, the mods have added even more non-pony bullshit to the site. It doesn't matter what you or I do, the damage is already done.
>>
>>36935908
Unfortunately there's nothing to be done about that. Best way to minimize the damage is everyone filtering the offending flags and pretending nothing changed.
>>
>>36935801
it was originally posted in the pinned flagged thread but eh
>>
>>36936114
Oh my bad then. Name made me think it's one of the mareschizos.
>>
>>36935874
>mods ask what people want to make the board better
>overwhelming responses to kick EqG out
>retarded mods decide to cater to barbiefags instead
God, the singularity can't come soon enough - I fucking hate mankind.
>>
save
>>
https://vocaroo.com/1fvFa7pdjDh1
>>
Adding radio music to audios is fun, so I think I'm gonna do it again.
>>
>>36922769
I would like to volunteer some non-pony audio samples (specifically, Tucker Carlson V3) simply to showcase the best case scenario involving large quantities of high-quality single speaker data. While it is the Pony Preservation Project, I believe a demonstration of what current tech is capable of would be productive.
>>
>>36930714
What is it? Link please?
>>
>>36930714
And how can I test this?
>>
>mlp porn thread been going on /co/ for over an hour now
>>
>>36931558
I wrote to you on Twitter, I would like to test this neural network. I want to chat with Twilight.
>>
>>36939343
It's pretty funny.
>>
>>36939343
how is that tangently related to this thread?
>>
Does anyone have the google doc of characters that have been submitted to 15.ai?
I tried to look through the archive but I couldn't find it.
>>
>>36939428
https://docs.google.com/spreadsheets/d/1dd8yv2MyRhxCNWWO04xOk6-h_XnKU_xj34qZ8pobCfc/edit?usp=sharing
literally took 5 seconds to search the archive
>>
File: QUEEN.png (317 KB, 1175x1024)
317 KB
317 KB PNG
>>36939397
>>36939401
https://u.smutty.horse/mawekmtpxnk.wav
>>
File: 1618846175310.png (58 KB, 512x512)
58 KB
58 KB PNG
>>36939479
>>
File: uh_oh.png (196 KB, 1121x815)
196 KB
196 KB PNG
>>36939349
Haven't seen any messages, sry.

But, there is no way to try this as of right now, I am pushing for a public release soon (tm).
>>
>>36939479
>that fucking delivery of I can't sneed
bugbutt is a master shitposter
>>
>>36939551
Man those TPDNE images look rough.
>>
>>36933455
the LIBRARY folder should have one XML file for each symbol, I think. here's a list of what the commands mean: https://stackoverflow.com/questions/4077200/whats-the-meaning-of-the-non-numerical-values-in-the-xfls-edge-definition
will I need to use the Python interface? as far as rendering texture atlas stuff goes, it seems like I won't need it. if that's the case, you can save yourself the effort
>>
>>36939479
It's amazing for how little speaking data she has how perfect her voice is. Wonder why that is.
>>
>>36939278
>>36939305
If you're talking about the system in the picture, it's not public yet >>36939551
If you're talking about GPT-2, you can use the model to do pony/fanfic stuff, like AI Dungeon. I've never done any of that myself but I assume you can just use that GPT as a substitute.
>>
>>36931495 >>36931495
>ngrok AI voice
If you need any assistance from me then feel free to ask. My GPU's are currently idling so I can do a custom model or spare some resources as required.
>>
>>36939691
I got noticed by the senpai!

I played with HIFI-GAN version of Twilight but while super high quality it sounds too different from real TW. Is that lack of finetuning cycles or something else?
>>
>>36939568
>will I need to use the Python interface?
No, it looks like the XFL file has everything, so you don't need the Python interface for anything. Conversion to XFL is also fast enough that I can probably do it on my own. I need to finish one old task of dumping symbol samples for Clipper to label, which I'll work on now.
Parsing the XFL files to render symbol animations is the most difficult remaining task. Post updates as you have them so I know what not to work on.
>>
>>36939451
I must’ve been searching the wrong way then. Sorry and thanks.
>>
File: image.png (9 KB, 604x120)
9 KB
9 KB PNG
>>36939551
I hope you will write here when it goes public.
>>
>>36939677
Yes, I was talking about the one in the picture.
>>
>>36939558
Why would a chatbot for an established character like Twilight need a TPDNE-generated avatar, unless you wanted to avoid trademark claims from Hasbro? Ideally, Twilight's face should be her Flash assets from the show with animated lipsynch.
>>
>>36940159
Because it's not a chatbot for "an established character", with a stylegan-based model you can create you own "version" of Twilight (or other characters). i.e. TediGAN, CLIP+StyleGAN, MakeItTalk, etc.
>>
https://u.smutty.horse/lzjidtyonxq.wav
>>
File: 1601610881137.gif (2.2 MB, 854x720)
2.2 MB
2.2 MB GIF
>>36941425
I always preferred the earlier one where it sounds like glim glam sneeded to the point of insanity
https://www.vocaroo.com/1jcNhNaR8w5F
>>
>>36939788
I don't believe I've used the HiFi-GAN version you're on about.
Possible reasons (from guessing) are
- Vocoder requires fine-tuning
- Tacotron2/Spectrogram generator is inaccurate/unnatural
>>
>>36927518
>>36928938
It is a virus. Viruses are viruses. Many good people fell out of weakness. Praise the bringer of pony AI.
>>
>>36931610
Spam filter is really REALLY easy to trigger.
>>
>>36927518
I wouldnt fell so sure, the man is from MIT
>>
File: polish hymn v001.png (8 KB, 608x157)
8 KB
8 KB PNG
figuring out how to make ponies talk and sing in foreign languages is fucking bullshit, I tell ya!
>>
>>36942704
Shows how weird that language is compared to English.

Also, testing out this new character thing.
>>
File: poland fuck.jpg (24 KB, 600x465)
24 KB
24 KB JPG
>>36942868
>Shows how weird that language is compared to English
I don't need you to tell me how weird my language is, m'kay. I'm the one who speak it. I know how weird it is.
But for real, I can't wait for some kind of universal TTS multi-language convertor so I wouldn't have spent so much time figuring out how to reverse-engineer my own language
>>
>>36939551
is it going to work like 15 and be closed source
or can you host it with a collab
>>
>>36942868
Well, it works for me.
Need to refresh the page tho.
Not sure if it's a good or a bad idea, but, oh well, we will see...
>>
>>36922769
I would really love to help, but I don't know what I can do.
I don't trust myself to be on time to talk.
I can create audio tho (I finish moving out in three days, I should have more time then. and like, 95% less internet access unfortunately.).
So, don't hesitate to give me an idea of something to do, if you believe my quality is up to the task (and if not, well, I will work more and improve for next year I guess).
>>
>>36943835
>but I don't know what I can do.
In general, you can talk about the work you've done, what you've learned and what you plan to do in future. The specifics will be up to you.

>I don't trust myself to be on time to talk.
If you can make it on the day, we'd be glad to have you - and that goes for anyone else who has something to say, even if it's only a few minutes.
If you can't/not sure if you can make it, then you can make some other kind of submission in advance that we can read/play or otherwise show to the audience.

>I can create audio tho
We'll have a showcase of the best AI voice content at the end, so we can put anything you make in there.

>if you believe my quality is up to the task
Like I said in >>36926109, I'm generally looking to keep the restrictions as minimal as possible, so pretty much anything is acceptable provided that it's on topic and some amount of genuine effort went into making it.

If you want any more in-depth help or advice, send me an email and I'll see what I can do.
>>
>>36922769
Will the outro be just as amazing as last year? Reuse if you have to.
>>
>>36939829
if rendering from XFL works, then it seems like the texture atlas approach won't be needed anymore. in that case, it might have been better to start with XFLs to begin with (though texture atlases are definitely easier to get working). at the very least, XFL seems to represent timelines and keyframes in a similar way to Animation.json, so some of the code can be reused
a quick search of XFL on github turns up a decent number of libraries (even one whose example uses a pony puppet), so there's existing work to build off of
this script has taken more time than I expected, and I have some voice/image projects I want to finish, so I won't be able to write the XFL renderer (maybe in the future, but it seems unlikely). I'll at least clean up this script and post it, though
>>
>>36938861
The focus is of course on pony, though as you say having some other samples to show off the capabilities may be useful. We can probably find a use for that in the sections for discussing AI in general, so if you want to send me some samples and any notes from your work to showcase I'll see if we can find some use for it.

>>36943961
>Will the outro be just as amazing as last year?
With all the great audios we've had over the past year or so, I'm sure it will be.

>Reuse if you have to.
There's been so much great content produced that I genuinely don't think we'll have to, aside from maybe a few early works to show how far the project has progressed.
>>
>>36943997
Well you better nail it with the music, last year's video still gives me goosebumps.
>>
>>36943072
look up words here
https://pl.wiktionary.org
and hack this a bit to make it swap ipa to arpabet
https://github.com/dohliam/xsampa
>>
>>36944237
Are you referring to this? https://www.youtube.com/watch?v=730zGRwbQuE
SnoopyAnon is the one who made that last year, so far it's unknown if he plans to participate in this year's con, he's been MIA for about a month now.
>>
>>36944295
also i just found this, don't know how useful it is though
https://github.com/AdolfVonKleist/Phonetisaurus
>>
Writing my contribution to the PPP panel, and looking for my old posts as examples, and came across the first ever post I made to the project: https://desuarchive.org/mlp/thread/35129047/#35147842

How time flies.
>>
File: denn.png (48 KB, 500x252)
48 KB
48 KB PNG
It seems that RNNoise shows very satisfactory performance on VCTK dataset, as this .zip with some samples shows.
https://u.smutty.horse/mawnhiwxvig.zip
I don't have the noisy versions since I denoised the entirety of it in my disk, but you can head to https://www.tensorflow.org/datasets/catalog/vctk to see what kind of noise it is dealing with.
Demonstration of batch denoiser: https://u.smutty.horse/mawniksweun.mp4 (the climbing memory usage problem has been fixed in the latest release, which you can grab from the Google Drive folder).
This should be good for anyone here who wants to clean up a dataset without bothering to manually adjust parameters.
>>36943997
All right. Where do I send them? Just post here?
>>
>>36944615
Huh, I guess we both jumped on board around the same time. My first post was the AJ rejection thing posted just a short ways above yours.
https://desuarchive.org/mlp/thread/35129047/#q35147070
>>
>>36944615
>>36944683
SOUL
>>
>>36944683
>First post that I can actually remember is mine:
https://desuarchive.org/mlp/thread/33700529/#q33720589
>>
>>36943984
Thanks a ton for your help.
For anyone curious:
- https://github.com/SasQ/SavageFlask
>>
>>36939451
Why does every entry in the Training column say "No", even for voices we know 15 has trained before?
>>
File: ml-is-magic.png (750 KB, 4000x1815)
750 KB
750 KB PNG
>>
>>36945418
JUST BUY IT
>>
>>36944626
Email them to me, that way we won't have any spoilers in the thread.
clipper.anon01@gmail.com
>>
Hello! Can I have a shorter instruction on how to make the voices of the heroes?
>>
File: 1614249099848.png (333 KB, 586x730)
333 KB
333 KB PNG
>>36945842
OI

M8
>>
>>36945842
see >>36904619
>Document: https://docs.google.com/document/d/1xe1Clvdg6EFFDtIkkFwT-NPLRDPvkV4G675SUKjxVRU
>>
>>36944626
>you can grab from the Google Drive folder
could you link me up to that google drive folder (im not seeing link for it in the main doc)?
>>
File: ffffff.png (5 KB, 237x243)
5 KB
5 KB PNG
hi, i made a dataset for a nonpony character, homsar from homestar runner. i feel bad about removing the link a while ago so i wanna share again. i also submitted this to 15.

https://mega.nz/file/lZpmhRAK#qELZXGSgNd7DjVYogk_F0abvVQ1rD-1a7GlRuxm1emw
>>
https://u.smutty.horse/mawtxhzhghk.wav
>>
Is 15 still out and about? I'm assuming yes given the test site is still running.
>>
File: the list.png (29 KB, 925x770)
29 KB
29 KB PNG
Working on a new Ngrok model because why not. It includes several MLP characters from season one, BFDI characters, some Vocaloid characters + Kasane Teto, and a singular Touhou Project character because I made the dataset for someone else and decided I might as well use it. It's very early on in training, but hopefully it achieves the same level of quality as my main BFDI themed model.
Some samples:
https://vocaroo.com/1nlBtSTb1C6T
https://vocaroo.com/1238leUCqWkA

>>36947242
Judging by Desuarchive, he was last seen Wednesday in another thread.
>>
>>36947383
Glad he's still active. I know it's only been two weeks since he last updated us on anything, but sometimes I just get anxious when people are gone for a long time.
>>
>>36947488
he been busy playing tf2 huehuehue
>>
>>36947146
https://vocaroo.com/11n5yoyzqt8A
>>
>>36946071
here
>>36909226
>>
File: gooooooooooogle.png (115 KB, 900x395)
115 KB
115 KB PNG
>>36904619
Is there a way of using Ngrok without selling my soul to Jewgle?
>>
>>36947488
Like the other anon said, he plays on the /mlp/ tf2 server every night so he's absolutely still around. He last mentioned about changing up the model so I'm assuming he's been quietly working on it before updating us.
>>
>>36932474
https://u.smutty.horse/mawwnybnsae.ogg

Alright, pretty much finished the audio, redid some lines, adjusted timings and added an opening, next possible step is to make an animatic for this.
>>
>>36948869
>https://u.smutty.horse/mawwnybnsae.ogg
There should be a little more of a pause between Starlight hanging up and Chrysalis's reaction i think. Otherwise its coming along nicely.
>>
>>36948476
If you can code (or learn to code) then anything is possible anon.

>>36947242
He plays tf2 every night so you can tell if he's physically alive. No idea about progress on the model side though.
>>
>>36943136

All ML models and code necessary to run them - open source and runs in colab (or locally), the site itself and frontend closed source for now, but who knows.
>>
https://voca.ro/13ldkDY8pZkn
>>
>>36949153
I figured the pause might have been bit too long. But I might change that when I do animatic.

And here's a super fast sample of first shot in super rough!
https://u.smutty.horse/mawysqejeel.mp4

Also haven't drawn pony before so don't you go expect good looking ponies by the time I (might) get this done.
>>
>>36950103
What have you've drawn? mind posting any of your work?
>>
these flags actually seem like a neat way to say whos in the recording without typing it out
https://vocaroo.com/12BrIIGOFEOb
>>
https://vocaroo.com/1fBbClGLtq3w
>>
File: 1597621206199.png (771 KB, 738x700)
771 KB
771 KB PNG
>>36949541
>[name] [trips]
who the fuck is this
>>
>>36951424
He's the guy doing the GPT-2 + 15.ai + TPDNE thing, see >>36939551. I agree that the trip is really not necessary.
>>
>>36949541
Resume your discord server or write your own for communication!
feniks3710@gmail.com
>>
>>36951469
I don't use 15.ai, this is the same model from the ngrok, but if 15-kun wants to collaborate, I am in.

>>36951540
I have no idea what you wrote here, sry.
>>
>>36950228
>What have you've drawn? mind posting any of your work?

90% of my public stuff consists /d/ content, and only times I drew anything MLP were humanized/EQG commissions over 5 years ago (and those pics have aged like a milk), but here's a view of several "normie" contents I've done last few years.

/co/ draw thread
the-collection.booru org/index.php?page=post&s=list&tags=a0iisa

Personal DC char design Project
catbox moe/c/7lr8j6

Few from my DeviantArt
deviantart com/a0iisa/art/Riveros-Sisters-840853097
deviantart com/a0iisa/art/Random-Sketch-Essie-Split-860129738
deviantart com/a0iisa/art/Random-Sketch-Preggo-Kawakami-Remake-865610091
>>
File: Spoiler Image (745 KB, 952x1580)
745 KB
745 KB PNG
>>36952109
Nice stuff anon, I would like to see how that animatic turns out.. and try not to get burnt out.. animation is a bitch.
>>
>>36951842
drop the trip
>>
>error code 404
welp, time to wait for 15 to underp the site.
>>
>>36949215
You might find this useful:
https://spell.ml/blog/spell-open-research-grant-YIMtSxEAACQAKm81
>>
File: 1950207.jpg (127 KB, 2048x1434)
127 KB
127 KB JPG
>>36904888
Ok now I have a boner
>>
File: 1613136566836.jpg (152 KB, 1280x720)
152 KB
152 KB JPG
Oh the testing site is down again.
Though for future reference, I think I found a phrase that makes non-Chraysalis models even angrier than "|Fuck You!"
I tried "|I'm gonna kill you!" as the modifier and this was my very first result
https://www.vocaroo.com/1fvHs0taDOsf
>>
>>36954245
Well, she certainly sounds like she wants you to shut up for good. Gotta remember that for the future.

>when she heard "iwtcird" for 9281354607th time
>>
>>36954245
thanks for sharing that, here is something I've discovered, when the model emotional pronunciation is all over the place an you just want a absolutely tone neural response use this funky science word:
"|PNEUMONOULTRAMICROSCOPICSILICOVOLCANOCONIOSIS ?"
>>
>>36954204
You'll like the next Pony Zone then ^:)

>>36954245
I used to use "I'm gonna fucking kill you!" whenever I needed absolute ANGER for a model. I thought all there was to emotional transfer was the emojis it says are related but now I'm questioning that.
>>
>>36947383
This new model isn't ready at all yet whatsoever, but I've gotten it trained up to 9500 iters. Right now all of the voices are inferior to the main Ngrok or the 15.ai voices, so it's just for anyone curious about my progress. A few BFDI characters like Four actually sound good, but everything else sounds bad and definitely needs more training.
https://f5c465a1fd2c.ngrok.io/
>>
>>36938369
Late but based Trixie.
>>
>>36954676
>needs more training
yep, like you said Mane6 somehow still sound like themselves to one degree or another, while rest are barley recognizable. What kind of tech/code are you using in those BFDIAnon ?
BTW, if you or anybody is interested here is final training data for Gothic 2 Nameless Hero, I've tossed away pretty much all the lines that were sounding OOC in tone so hopefully this will be one of the cleanest male non-pony data available (unless someone wants to dick around cleaning Gerald voiceset).
1Ltm3FSoZxPrGlyreO_1fLC_AMQUtA7oz
text validation
1zTprqo2KaMgcsbU9GXdgVRPjo3gg_azV
text training
1lWVXy6tUh1BX8aunnaYvmcUBZ9UXFe50
zipped wavs (1h 41m)
>>
>>36922110
I like it, interesting concept, but would you mind posting the lyrics? I still can't understand everything (for some reason understanding lyrics in songs has always been difficult for me, even in my native language), yet I'd really like to.
>I'm more used to working with vocals you cant really understand what they're saying anyway so
kek, that's common for heavy metal, isn't it? Always wondered why.
>>
>>36955350
>Always wondered why.
Because metal is a genre that wants to focus heavily on guitars and what they can do, and trying to mix together vocals, which generally either need a lot of 'breathing room' in the mix, or a large volume boost to stand above the rest of the instruments, with metal guitars that leave anything BUT breathing room and generally need a very dominating presence in a mix (especially in metal) is a nightmare. It's basically the mixing equivalent of having your cake and eating it too, if maybe SLIGHTLY less impossible.
>>
>>36955350
>that's common for heavy metal
I guess it depends of what type of metal you're listening to.
>>
File: 1354719816132.png (46 KB, 500x500)
46 KB
46 KB PNG
its over, isnt it
>>
>>36955690
Why should it be?
>>
>>36922769
Wait wasn't /mlp/con a few months ago?
...how old is this project?
>>
>>36955842
https://desuarchive.org/mlp/thread/33700529/
>Fri 05 Apr 2019 11:06:59
old, fairly old.
>>
>>36954245
Well, this gives me incentive to work on non-audio projects, like some of my greens that I haven't gotten to in a while.
>>
>test site still ded after 12 hours
rip my projects I guess ?
>>
>>36955963
that, or you cold work on what you want to do with 15 when it goes back up
>>
Let's bump this thread with the high hopes of test site getting fixed soon.
>>
>>36956978
>made loose files for shitposts but never finished them
>now I'm going back and doing them as I find them
https://voca.ro/1hn2gVxNDLJm
>>
Well since the site is down. Anyone notice you get better results when using shorter inputs? It seems to do a much better job keeping a consistent tone.
>>
>>36956415
That too. I was working on a Maud audio when the test site went down. I just need to finish writing out the remaining portion of the script (since I got kind of stuck halfway through).
>>36957376
I noticed it with Maud and Trixie. When generating a full phrase, the result is more noise and less coherency. In previous audios, if I had a really long sentence, I'd have to generate it in several pieces then stitch them together. I want to say Fluttershy was tougher than Daring Do.
>>
>>36957376
depending on characters, but yeah the long sentences almost always need to be broken down. I do ding out that on occasion adding extra word at beginning/end of the sentence sometimes get rid of the soft fuzzy noise that is generated there (of course you will need to cut those in audacity/editor).
>>
https://u.smutty.horse/mafbuciraga.wav
>>
>>36954312
Is this a motherfucking electric company reboot reference?
>>
https://u.smutty.horse/maxuvxdeyuc.wav
>>
>>36959779
was she coughing at the end there?
>>
>>36953029
Sorry, I've been busy playing TF2. I'll put a new site up tomorrow.
>>
File: Spoiler Image (2.91 MB, 720x406)
2.91 MB
2.91 MB WEBM
>>36959840
Who do you main?
>>
>>36959840
I genuinely thought that was a joke that was just going around on this thread until others explained it to me a couple days ago. But man, it's been a while since I've played TF2.
>>
>>36960007
He destroys everyone with literally every class he touches (not even joking)
>>
>>36960053
I heard 15 has created some TF2 servers for mlp before.. is that true? man I would love to play TF2 with you guys.
>>
>>36960092
He hosts every Saturday after the rewatch streams afaik.
>>
>>36960092
>>36960113
See >>36959986
He hosts the server that we play on every night, and it's one of the most active community servers in TF2.
>>
>>36960115
Gamers rising up?
>>
>>36960007
https://u.smutty.horse/masqeybpayv.mp4
>>
>>36960115
Thanks! joining it now. TF2 has always had a weird connection to MLP.. don't know why they seem to cross paths a lot.
>>
>>36960151
What time does it usually get populated? I live in the Downunder so it's probably pretty late
>>
>>36960160
The regulars have already logged off tonight but it's usually packed with anons starting from 6 EDT (so around 8 AM there?) and goes on until late night. Just keep the server in your favorites and check when anons start to hop on.
On Saturdays it's nearly full for almost 12 hours straight.
>>
>>36960170
There's always a few people on the server even now, so you can hop on and ask them any questions. Sometimes the server goes back to having a lot of people in the dead of the night.
>>
>>36960173
>Just keep the server in your favourites.

Done.
This server reminds me of the peak of both communities back in the day.. feels good.
>>
>>36960007
scunt
>>
>>36959840
Good to see you alive . I've never heard from you in a long time
>>
>>36959840
Is that titanfall 2 or teamfortress 2?
>>
File: icartoonface.png (884 KB, 922x924)
884 KB
884 KB PNG
there is a dataset called iCartoonFace used for cartoon face recognition research and it contains pony
>https://arxiv.org/pdf/1907.13394.pdf
>https://github.com/luxiangju-PersonAI/iCartoonFace
>>
>>36960732
well after actually looking through 2k pictures in the validation set I only saw 1 pony screencap and 1 picture of some plushes. most of it is asian cartoons and some western stuff like dora
still funny but if you actually wanted to do stuff with pony faces you might as well use the MLP-Face-Dataset
>>
>>36960630
Take a wild guess.
>>36960129
>>
>>36920023
I knew that voice sounded familiar: https://www.youtube.com/watch?v=V0f2TDqUp1U
>>
>>36960732
Pony is the best cartoon, you can't even cmv





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.