>Special shoutouts go to cuckchannel's /mlp/ and its anons who have collectively spent hundreds of hours collecting, cleaning, and organizing clips of dialogue taken from the show My Little Pony: Friendship Is Magic. Thanks to everyone else who helped curate datasets for their favorite characters. Further thanks to Cookie for entertaining a couple illuminating discussions that contributed to the latest iteration of the model, and for proposing the idea of using DeepMoji embeddings to aid in context-based emotive generations; thanks to persona nerd for suggesting using a GAN vocoder as an alternative, leading to a significant improvement in audio quality. Finally, thanks to everyone with whom I've corresponded with regarding related issues within the past two months. (And as always, thanks to /v/ for the excellent content.) I don't remember any anon thanking an imageboard while working with machine learning tech, let alone cuckchan; it's usually disowned. Also, I've yet to see anything similar that relies on seiyuu clips in the animu side. Perhaps it's the way speech synthesis works for the Japs?

>>8740 Kinda, they have the vocaloid stuff which is pretty advanced but English Miku sounds like shit. And they have no fair-use laws so copyright will bite them in the ass.

if you can't get the inflection right, try adding a second sentence with the pipe symbol (it's that vertical line above backslash) for altered deepmoji analysis (I cannot believe I'm unironically using the word deepmoji). to wit, I got the inflection on webm related by adding "|Fuck You!" to the original line >>8738 every time I work with Carl the voice always sounds like they missed the last 1/8 second of each word. He talks way too fast. Scout too

>>8743 What are you trying to make him say?

>>8743 >if you can't get the inflection right Have you tried using other words like "buttox" or "behind"?

There's been some good ones floating around.

Oh yeah I tried it but it's shit, it doesn't let you input your own data. There are only so many comedic situations you can put Carl in.

>>8744 various memetic expressions >>8746 the way the analyzer works with the second sentence is it uses it for deepmoji analysis and ignores the first sentence. just changing it to buttocks wouldn't provide the analyzer with an emotional impact if you're suggesting changing the original sentence itself, surely you've at least seen pic related

>>8759 But Carl works in any situation.

>>8760 I mean if it helps smooth out how the Sponge voice says it while keeping it funny, then it couldnt be too bad. The sharp "butt" sound really breaks up the flow.

>>8753 This is great. Got any more to share? By the way, considering Rick May is dead, how good is the synthesis for soldier?

>>8766 I've been getting them from twitter.

>422 server error >422 server error >422 server error >422 server error this is getting tiresome

>>8767 >twitter Well at least you honest.

>>8780 It was either that or going to discord, tiktok or cuckchannel.

>>8781 Or you could have made your own

>>8763 the sharp "butt" was supposed to expressed shock. As sentences get longer they tend to lose energy without finagling it

>>8784 Steven Universe one needs the star iris zoom in at the end instead of a fade to black

>>8740 Japs had a meme on NicoNico for years where they would sentence mix the dialog of Broly and Paragus from the first Broly movie, all because they said every letter of Japanese alphabet at least once, the people making these videos could essentially make them say anything they wanted. Essentially making Japanese Youtube Poops.

>>8787 That and Touhou cookie along with the Homo men edits.

>>8784 True but that would require editing.

>like oldish PC game I played as a kid >its so obscure its not even on several lists of all video games lists (GameFAQs, Mobygames ect.) despite being commercially released and its two prequels showing up fine >filled with speech since it was a CD-Rom game when developers wanted to show how much data they could throw onto one >almost all the voices happens to be in unobfuscated .aif that's already mostly organized and clear of anything but the snippits of speech >submissions are open and want audio submitted in pretty much exactly the way it already is I'm quite tempted to spend a few hours transcribing this game's dialog just so I can make one of the characters say funny things that will only amuse me.

>>8823 Give it a shot, there's no real barrier on who's allowed in or not. Just whether the reference audio is well prepared.

Of all the work people are doing to upload voices, the focus should honestly go to characters that'll never be heard again.

>>8824 OK, I've got over a third of it done. Should provide 7+ minutes of dialog when done, and I think even more could be obtained from the parts that aren't so easily extracted and the previous game the character was in.

I couldnt figure out how to add audio tracks

>>9009 I think you need to send it directly to the developer.

>>9009 Yes. Read the contribute section: You have to transcribe all the audio and send a zip of it+audio files to the dev, with some requirements. The contact is in the about. >>8845 Finished. It's ultimately around ~10 mins in length (I know of no way to fully tally the exact length, but an average length of 2 seconds times the number of files gives me over 10 mins)

>>9009 >>9012 Are the devs FINALLY accepting new characters? Seriously it needs more if it ever wants to stay alive.

>>9013 I doubt 10 minutes is long enough to train decently the neural network, but tell me if you character gets accepted to have a reference.

/v/ related, but I just realized I can send them WC3 audio tracks and make Arthas memes for the rest of my life.

>>9017 When you're done mod the game and change all dialogue to https://invidious.048596.xyz/watch?v=7gCo8fajC7o

>>9018 There's an argentinian guy working on a remake of the original WC3 storyline with Wow models, this come really in handy form him.

>>9016 The majority of characters have under 11 minutes of audio. https://15.ai/characters

Some of these are really good.

>>9028 Those sound terrible...

Any idea of what kind of override sentences I need to make the AI sing?

>>9031 If its anything like moonbase alpha then something like this. Space odyssey: [dah<2000,17>][dah<2000,24>][dah<2500,29>][dah<200,35>][dah<2000,34>]

>>9032 This isn't a normal text to speech layout.

>>9033 I mean 15.ai

>>8835 I agree with you Anon, on the bright side if you can find clean voice tracks the guy is apparently willing to do it. but you may have to venture to cuckchannel's /mlp/

>>8835 Strongly agree, no one could contest the validity because while you could copyright strike shit like pre-recorded messages on movies and cartoons, how can you justify being able to mimic someone else's voice when they can't do it anymore? All reenactors would have to be shot on sight. Anyhow, does anyone have any idea if this guys ever released the source code or have found his original papers? >>9031 Have you tried just having them play the lines with a |happy deepmoji encoding and then just using something like audacity to give it higher or lower pitch?

>>9046 He hasn't released the source, I think it some sort of patent or something.

>>9084 15.ai truly has unlimited potential.

>>9084 HOLY FUCK I couldn't breath for a minute straight >>9086 >that ending Nice lol

When is someone going to make a moonrap feat. one or more of these characters?

>>9090 I could have sworn I saw a webm of Spongebob singing Notorious KKK.

>>9058 >>9059 I don't know what's so fucking funny to me about all of these videos of the Demoman being so done with his degenerate teammates and why "uwu" seems to be such an easy word for the NN to produce. Out fucking standing.

>>9058 >>9059 >>8767

Site's going down for maintenance again.

>>9017 Hearing the "suspenderoni your pepperoni" pasta in the 15.ai voice would be amazing.

kill me

Update coming soon (only date given was maybe by the weekend, but that was days ago). No new voices and processing of them will only begin over a week after new version is released. Creator does however have data for all Oblivion, Skyrim, and Fallout series voice sets.

Its back.

Got bored, made this.

Man, I'm almost scared to see how accurate this shit will be by the end of the year.

>>9090 >>9108 I aim to please.

>>9031 >>9713 >>9759 >>9799 >>9850 p good lads

While the site is down, I went through the twitter replies that contained "data set" "dataset" or "sent". The following (plus one group of characters I'm not including because the tweet has a screenshot showing it's in the wrong format) are confirmed to have a data set submitted >CD-I Link >HEV Suit >Juri Han >HL1 Scientist >Shadow the Hedgehog >Komaeda >Freddi Fish >Goro Majima >Wario >Vega (Doom Eternal) and I can personally confirm I've sent >Sydney Delgado These will probably be added eventually. Unfortunately no archive site I know of got the characters page, which mentioned some to be added characters.

>>9891 You think we could send one out for the memes?

>>9924 If someone wants to send one, I propose Captain Murphy.

I propose Troy McClure RIP Phil Hartman

>>9934 >>9925 >>9924 You'd have to get the audio ripped and transcribed in the right format. I had the audio essentially pre-ripped and it took a few hours to transcribe and format 10 minutes. If you still want to do it while waiting, here's what I remember of the submission requirements >put all audio files in one directory (if subdirectory is valid is unclear, but do it all in one folder to be safe) >put transcripts in a text file in following format (this style for file path, one file+line per line) /foldername/audio1.mp3|What the line says. /foldername/audio2.mp3|What the next line says. >time limit per clip is 15 seconds >only use the 26 basic latin alphabet characters plus basic puntuation (period, comma, exclamation point, question mark and hyphen) >like writing to generate something with 15.ai, punctuation is more about timing than proper grammar >do not use ellipses >like writing to generate something with 15.ai, write out numbers ("twenty") >leave grunts or other sound effects blank in the transcript (I did transcribe things like hmm, uh, aww, or huh that might show up in a dictionary but are arguably not actually words, but it was unclear on that point) >leave files in the original format unless you have to split them (I presume this is for video games where the lines are already broken up) >if you do use multiple sources make sure the audio levels are the same for everything (presumably because recording/output standards vary between a game and its sequels, or episode to episode)

>>9934 >Brought to you by Uber Eats

>>9934 >>9925 Both are good suggestions. I wanted Aku back to be honest but it's going to be a pain to remove the reverb from all the music of his theme. >>9939 >I had the audio essentially pre-ripped and it took a few hours to transcribe and format 10 minutes. Couldn't you use one of those speech-to-text automatic transcribers, then just work from there onward to spell check everything? Anyhow, good job anon

>>9971 No. I type quick enough the actual transcription is negligible time. The problem is pulling up each clip, listening to it a few times to make sure it says what it says it says, and getting the file pathes typed.

>>8738 Clearly it hasn't been fed enough lines, as some of them still retain that robotic noise to it.

Not only that, but also learning tone as well.

>>9980 The FAQ mentioned it does better with female voices than male ones, particularly super deep male ones. This is an issue because the most popular voices, bar maybe Scout, are all fairly deep voices, and most there wasn't enough female voices to really do much of note with (while just remaking voice spliced Team Fortress 2 stuff is massive showcase potential) and the female voices that are there either don't have much emotion in the first place (Daria, Rise), or are heavily robotized already (GLADOS, Turret).

Doctor who isn't a cartoon, but I noticed the tenth doctor voice sounds more Scottish than he is on the show. I know the actor is Scottish but he was putting on an english accent.

>>9980 >>9981 The problem is that you need a ton of annotated voice lines and starting off of previously learnt weights may be okay if the voices are similar but not so much if they're completely different. It may also be a case of overfitting to the training data. Also they're using Deepmoji to determine intention from the context of the phrases, it may be that it's specifically tiered towards one kind of pitch and ignores all the rest. I've tried making Soldier say the exact same phrase with different intent (happy, sad, angry etc.) and it sound extremely samey, but you can feel that small hint of emotion. Either way, it's an impressive improvement over what most other NN shitposters have done.

>>9986 >it does better with female voices than male ones No wonder its flooded with pony crap.

>>10060 Not to mention they had Halfchan /mlp/ going to autistic lengths compiling voice clips for them to use.

>>9891 Site did a small update confirming its still in progress. Mentions "hundreds" of datasets.

>>10541 And now says the new version of the AI is training now and implies there's nothing left for a human to do (or at least everything a human does is bottlenecked).

>>8766 Speaking of Rick May, there's the leaked uncompressed voice lines from StarFox 64, he really sounds like a younger version of The Soldier. https://youtu.be/SE1pKOJY66U

>>10060 The dev actually defended the pony crap on twitter.

>>10060 >>10870 The entire project was created to synthesize pony voices and create an advanced text-to-speech system with them. Of course it will be full of pony crap.

>>10871 No you faggot, the horsefuckers found it and are the most proactive contributors. They basically did the grunt work in the early days of the project by scrubbing through hours of dialog just because they got pissed the final of their show was a shitty epilogue.

>>10872 Oh did MLP end? How come horsefuckers were pissed?

>>10885 Rushed two-part final and an epilogue that barely answers anything, while pushing a retarded ship.

The first two ones are fucking gold. Who made them?

>>8784 >Fucking Discord denouncing the yellow man's penis in favor of the master race's enormous bulging member

>>9934 >>9949 This would be an act of evil. Spare him, anons. He'd still be with us today if Andy Dick wasn't a complete cunt and gave Mrs. Hartman those drugs. No, I'm not kidding. Jon Lovitz beat the unholy fuck out of Dick years later at a bar

>>11006 Andy Dick has always been fucked up and beyond. That's why he got a #MeToo.

Twitter now says >I was way off with my prediction for v14. This one took over 50 retries to get it just right – the last several versions took around 10 or 20 retries each. So, final version to be released soon™

>>11385 Well, while you're waiting for it, someone else made it for other voices: https://vo.codes/

>>11571 >software filters words the developer doesn't like and generates audio of the character saying "you should not say something hateful like that"

>>11573 >it's real I hope you'll find it in your heart to forgive me for mentioning it in the first place. Though at least you can circumvent it by using or combining words with a similar pronunciation. Why do people like this exist.

>>10872 >No you faggot, the horsefuckers found it and are the most proactive contributors Wrong. 15.ai started as part of /mlp/'s own Pony Preservation Project, and 15 himself was a huge contributor. https://desuarchive.org/mlp/thread/33700529/ (first thread about the project) https://desuarchive.org/mlp/thread/33729880/ (first thread with a proper OP) https://desuarchive.org/mlp/thread/34611670/#q34627577 (first post made by15) https://desuarchive.org/mlp/thread/34996397/#q35030139 (first time 15.ai is mentioned) https://docs.google.com/document/d/1xe1Clvdg6EFFDtIkkFwT-NPLRDPvkV4G675SUKjxVRU/edit (a huge document detailing everything they have done so far)

>>9084 jesus christ this is both amazing and scary how accurate it is.

>>11697 I want to know what Dave Willis thinks of this.

>>11698 im all up for getting a perfect imitation of sweetie belle's voice on my sex doll but this has far more potential for evil than it does for good or leisure.

>>11699 Given the choice between this tech being developed in secret and in public, I'll gladly take harsh reality over blissful ignorance.