Welcome to the Pony Voice Preservation Project!youtu.be/730zGRwbQuEThe Pony Preservation Project is a collaborative effort by /mlp/ to build and curate pony datasets for as many applications in AI as possible.Technology has progressed such that a trained neural network can generate convincing voice clips, drawings and text for any person or character using existing audio recordings, artwork and fanfics as a reference. As you can surely imagine, AI pony voices, drawings and text have endless applications for pony content creation.AI is incredibly versatile, basically anything that can be boiled down to a simple dataset can be used for training to create more of it. AI-generated images, fanfics, wAIfu chatbots and even animation are possible, and are being worked on here.Any anon is free to join, and there are many active tasks that would suit any level of technical expertise. If you’re interested in helping out, take a look at the quick start guide linked below and ask in the thread for any further detail you need.EQG and G5 are not welcome.>Quick start guide:docs.google.com/document/d/1PDkSrKKiHzzpUTKzBldZeKngvjeBUjyTtGCOv2GWwa0/editIntroduction to the PPP, links to text-to-speech tools, and how (You) can help with active tasks.>The main Doc:docs.google.com/document/d/1y1pfS0LCrwbbvxdn3ZksH25BKaf0LaO13uYppxIQnac/editAn in-depth repository of tutorials, resources and archives.>Active tasks:Research into animation AIResearch into pony image generation>Latest developments:GDrive clone of Master File now available >>37159549SortAnon releases script to run TalkNet on Windows >>37299594TalkNet training script >>37374942GPT-J downloadable model >>37646318FiMmicroSoL model >>38027533Delta GPT-J notebook + tutorial >>38018428New FiMfic GPT model >>38308297 >>38347556 >>38301248FimFic dataset release >>38391839Offline GPT-PNY >>38821349FiMfic dataset >>38934474SD weights >>38959367SD low vram >>38959447Huggingface SD: >>38979677Colab SD >>38981735NSFW Pony Model >>39114433New DeltaVox >>39678806so-vits-svt 4.0 >>39683876so-vits-svt tutorial >>39692758Hay Say >>39920556Haysay on the web! >>40391443SFX seperator >>40786997 >>40790270Clipper finishes re-reviewing audio >>40999872Synthbot updates GDrive >>41019588Private "MareLoid" project >>40925332 >>40928583 >>40932952VoiceCraft >>40938470 >>40953388Fimfarch dataset >>410279715 years of PPP >>41029227Audio re-up >>41100938>The PoneAI drive, an archive for AI pony voice content:drive.google.com/drive/folders/1E21zJQWC5XVQWy2mt42bUiJ_XbqTJXCp>Clipper’s Master Files, the central location for MLP voice data:mega.nz/folder/jkwimSTa#_xk0VnR30C8Ljsy4RCGSigmega.nz/folder/gVYUEZrI#6dQHH3P2cFYWm3UkQveHxQdrive.google.com/drive/folders/1MuM9Nb_LwnVxInIPFNvzD_hv3zOZhpwx>Cool, where is the discord/forum/whatever unifying place for this project?You're looking at it.Last Thread:>>41064811
FAQs:If your question isn’t listed here, take a look in the quick start guide and main doc to see if it’s already answered there. Use the tabs on the left for easy navigation.Quick: docs.google.com/document/d/1PDkSrKKiHzzpUTKzBldZeKngvjeBUjyTtGCOv2GWwa0/editMain: docs.google.com/document/d/1y1pfS0LCrwbbvxdn3ZksH25BKaf0LaO13uYppxIQnac/edit>Where can I find the AI text-to-speech tools and how do I use them?A list of TTS tools: docs.google.com/document/d/1y1pfS0LCrwbbvxdn3ZksH25BKaf0LaO13uYppxIQnac/edit#heading=h.yuhl8zjiwmwqHow to get the best out of them: docs.google.com/document/d/1y1pfS0LCrwbbvxdn3ZksH25BKaf0LaO13uYppxIQnac/edit#heading=h.mnnpknmj1hcy>Where can I find content made with the voice AI?In the PoneAI drive: drive.google.com/drive/folders/1E21zJQWC5XVQWy2mt42bUiJ_XbqTJXCpAnd the PPP Mega Compilation: docs.google.com/spreadsheets/d/1T2TE3OBs681Vphfas7Jgi5rvugdH6wnXVtUVYiZyJF8/edit>I want to know more about the PPP, but I can’t be arsed to read the doc.See the live PPP panel shows presented on /mlp/con for a more condensed overview.2020 pony.tube/w/5fUkuT3245pL8ZoWXUnXJ42021 pony.tube/w/a5yfTV4Ynq7tRveZH7AA8f2022 pony.tube/w/mV3xgbdtrXqjoPAwEXZCw52023 pony.tube/w/fVZShksjBbu6uT51DtvWWz>How can I help with the PPP?Build datasets, train AIs, and use the AI to make more pony content. Take a look at the quick start guide for current active tasks, or start your own in the thread if you have an idea. There’s always more data to collect and more AIs to train.>Did you know that such and such voiced this other thing that could be used for voice data?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-imitations of official voices?No.>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 phonetic transcriptions of other languages as input for the AI.>What about [insert OC here]'s voice?It is often quite difficult to find good quality audio data for OCs. If you happen to know any, post them in the thread and we’ll take a look.>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?pony.tube/w/mqJyvdgrpbWgZduz2cs1CmPPP Redubs:pony.tube/w/p/aR2dpAFn5KhnqPYiRxFQ97Stream Premieres:pony.tube/w/6cKnjJEZSCi3gsvrbATXnCpony.tube/w/oNeBFMPiQKh93ePqTz1ns8
>>41137243Anchor.
>>41137243Pre-director OP image is a bad omen
I think mares are kinda cool
>>41138126I would love to own a cute robot mare maid.
>>41138285
>>41138042AI will help porny games.
Style Mixture of Experts for Expressive Text-To-Speech Synthesishttps://arxiv.org/abs/2406.03637>Recent advances in style transfer text-to-speech (TTS) have improved the expressiveness of synthesized speech. Despite these advancements, encoding stylistic information from diverse and unseen reference speech remains challenging. This paper introduces StyleMoE, an approach that divides the embedding space, modeled by the style encoder, into tractable subsets handled by style experts. The proposed method replaces the style encoder in a TTS system with a Mixture of Experts (MoE) layer. By utilizing a gating network to route reference speeches to different style experts, each expert specializes in aspects of the style space during optimization. Our experiments objectively and subjectively demonstrate the effectiveness of our proposed method in increasing the coverage of the style space for diverse and unseen styles. This approach can enhance the performance of existing state-of-the-art style transfer TTS models, marking the first study of MoE in style transfer TTS to our knowledge.>We trained the StyleMoE framework on the ”100-clean” subset of the LibriTTS dataset [24], comprising 100 hours of multispeaker speech data. This data was downsampled to 16 kHz from its original 24 kHz.https://stylemoe.github.io/styleMoE/no model but no one would want it since it's a small academic proof of concept one that is useless for anything neat. to the point where its just hard to tell if the method is even worth scaling but well interesting idea.
>>41135966>Tears for Fears - Everybody Wants to Rule the World (Rarity AI cover)>https://www.youtube.com/watch?v=w6817OqPfYUFinally finished it. I'm learning as I go, but I'm pretty content with this one. My goal is to make an AI cover for each of the mane six characters, so AJ and Rarara down, four more to go!
>>41138517nice work , however the ai outputs have randomly changing reverb (im not talking about two voice overlaying each other) that comes harder in few moments than in other.
>>41138967The reverb settings are constant throughout the song. What you're hearing is two voicetracks overlapping, a main/lead voice and a falsetto voice with a slightly different melody and a bit lower volume. During some parts of the song, only the lead voice plays and the falsetto is silent. Most noticable during the bridge.
>>41137243>They changed the fucking OP picAnd the downfall of the PPP begins.
>>41139105This was the original OP pic. It was originally changed to commemorate 15.ai going live, but since it's dead now, why not change it back?
>ywn sing your son and daughter a lullaby with your horsewifewhy live?
>>41139847Yeah that sucks, but if you think about it, we're all kind of lucky that we were born just at the right time to experience friendship is magic when we could most appreciate it. Plus, we get to enjoy the rise of AI during it's most fun times. The early years when it's still a free for all rechnology available to anyone, and it's not yet regulated to hell and back and monopolized by three or four big corporations who will sterilize the fun out of it for profitability.
Bump.
Would you all be okay if I posted a link to this thread to a discord server I might make.I am just trying to say something instead of bump.
>>41143349
>>41064967>With a laughter's ring (aha!), Pinkie's in line>Throws a party for two, with Anon, it's divineHuh. https://www.youtube.com/watch?v=ux8XqsIA9s8
>>41144337other than haysay there really isn't good way to use it. or train it
https://youtu.be/ORiIsU2D0Sk?si=zNuCQ_SVYiRO5APZSo... this is a thing. Before you rush to get it though, it's only available through a closed beta, and they have to check who you are before you get a copy. Unfortunately, I never got a response back, but eh, maybe someone here will have better luck?
>>41144723this looks neat, if this ever becomes open to the public I could easily see a great way to train with the emotion tags as their own vertex bits to better control the outputs in a way that none other speech tools have ever worked before.
"Smile" song (a different one)https://files.catbox.moe/sr2rr0.mp3
>>41144976Based and World War I-pilled.
i can't believe i never thought to post about my stuff herei know /CHAG/ deals with text gen models and all that but im doing more technical stuff with local text gen modelssee, im trying to get Qlora.py to work but all the tutorials i've watched go through it with ease while im getting an error thrown at me or just flat out don't even give a tutorialcurrently i have all the dependencies needed to run it but I'm stuck on a pic relatedalthough my instinct was to search for a solution online usuing the actual error code, i've come up with nothing to actually use in my situationim hoping that once i figure this shit out, i can do some training involving the chat logs where i augment them in multiple ways as to prevent over fitting, my goal is to attempt to make a permanent memory of sorts, an anon on /g/ already tried this but with regular loras and with completely raw chat logs and it over fit like crazybut since augmenting (meaning to replace words with synonyms, or reword something while retaining the same exact meaning) prevents this, im wondering if i could pull of some sort of long term memory with this thing
Hey guys, does anyone know the author? Did VUL really do this? I couldn't find this track anywhere. Gimme a link, pls?https://files.catbox.moe/0twb8w.flac
>>41145405If I understand you correctly, you want to achieve a form of long term memory by training a lora on a lower precision model, which is the definition of qlora, and train that qlora as the conversation goes.Long story short, this is unlikely to work because:1 It's unlikely to bring up actual facts "from the past" unless the lora is very high rank, but even like this it will indeed overfit and ruin model's performance.2 This is a very obvious approach and nobody has published a research about it as of today, which means that noone managed to get it to work3 Lora training takes quite a lot of time, from personal experience a 8 billion q5 model needed about 6 hours on a 3090 for my use case.If you still want to train a lora but aren't in experienced in tech, try oobabooga webui, it has a lora training option.
>>41146004P.S. /lmg/ and /chag/ are coomers, never take advice from them or assume they have any understanding beyond basedboy youtube tutorials.
>>41145405It seems like your GPU is not detected.Double-check that you have installed PyTorch with CUDA enabled and not the CPU version.And I disagree with >>41146009, you can get basic answers like >>41146004 on /lmg/.
>>41146053Ok maybe, haven't visited /lmg/ it in a while
>>41146053>pytorch with cudai specifically remember downloading the non cuda versionthanks anon ill do it when i get home
https://vocaroo.com/1oyZgeUTaR6lhttps://pomf2.lain.la/f/m4is43un.mp3UdioJourney to Dodge JunctionDriving through the desert, under twilight starsOn a quest for relief, it's takin' us so farThrough the tumbleweeds, Dodge Junction's closeNo more rest stops, where’ll this story go?It's the only toilet, in Equestria's landIn Dodge Junction, where the cowboys stand
>>41145926The song is"Chant of Selflessness" by 4everfreebrony. This is a remix of that song with a Rarity cover. I don't know who created the Remix.
>>41136413At this point I'd settle for just the TTS-to-RVC aspect of it. Freedom of mare speech without the restrictions of existing vocal bases as reference should always be sought; especially with RVC remaining as powerful as it is.Speaking of, has there been any recent developments in technologies rivaling or surpassing RVC?
>>41146009it just a bunch of avatarfags and nothing more
>>41146373cursed numget
>>41146501/lmg/ is for debating the existence of Hatsune Miku's penis, not developing and discussing Local Language Models.
Other AI generals bad. This AI general good.
>>41146566It's an actual cross-posting schizo.You can see him right now trying this shit on /lmg/>>>/g/100898357Here are all the generals he admitted to baiting in.
>>41146004>3 Lora training takes quite a lot of time, from personal experience a 8 billion q5 model needed about 6 hours on a 3090 for my use case.How many tokens was this? I'm getting about 3M tokens per hour for qlora training on a 4090, which is supposed to be 2x the flops of a 3090
>>41146670My dataset was ~20 mb. But it's not just the length that determines the time. Lora's affected layers, rank, sequence length and desired loss affect the convergence time much more. And the right stopping loss is hard to determine because there is no criteria for a "trained" lora, you want to find the balance between knowledge and overfitting.
Prompt-guided Precise Audio Editing with Diffusion Modelshttps://arxiv.org/abs/2406.04350>Aligning large language models (LLMs) with human preferences becomes a key component to obtaining state-of-the-art performance, but it yields a huge cost to construct a large human-annotated preference dataset. To tackle this problem, we propose a new framework that boosts the alignment of LLMs through Self-generated Preference data (Selfie) using only a very small amount of human-annotated preference data. Our key idea is leveraging the human prior knowledge within the small (seed) data and progressively improving the alignment of LLM, by iteratively generating the responses and learning from them with the self-annotated preference data. To be specific, we propose to derive the preference label from the logits of LLM to explicitly extract the model's inherent preference. Compared to the previous approaches using external reward models or implicit in-context learning, we observe that the proposed approach is significantly more effective. In addition, we introduce a noise-aware preference learning algorithm to mitigate the risk of low quality within generated preference data. Our experimental results demonstrate that the proposed framework significantly boosts the alignment of LLMs. For example, we achieve superior alignment performance on AlpacaEval 2.0 with only 3.3\% of the ground-truth preference labels in the Ultrafeedback data compared to the cases using the entire data or state-of-the-art baselines.no weights but they explained how they built off of Tango in the appendix
>>41146004first off i went to Qlora because i had so much difficulty with oobabooga webui lora training because at least then i could see the code and use it to search for a solution >>41146053ACTUALLY, the issue was not pytorch but rather that bitsandbytes had the cpu version instead of cudathis did fix another side error but i still get the original n_gpus error in my first post here >>41145405>>41146187and also when it comes to this post i misremembered bitsandbytes as pytorchim wondering if there's a variable i have to change or whatever, fucking wish there were coherent Qlora guides out there, then maybe i would learn this and be able to apply that knowledge elsewhere once i do more stuff like this
>>41146884So you do have pytorch + Cuda?
>>41146921I can tell I'm being retarded so I'm gonna try and spell out exactly what is going on so there's no confusion>do you have pytorch + cudaNo, I only have REGULAR, non-cuda, torch (NOT pytorch)i feel like the main reason im having difficulty with this is that i could see exactly what packages were imported in the code so all I needed to do is import those packages explicitly mentioned in the code after the word 'import' torch is mentioned but not pytorch, after you mentioned trying a cuda version of pytorch i searched the pycharm packages search thing, I could not find any variation of pytorch+cuda, bitsandbytes had a cuda version but i could not find one for pytorch>pic relatedthe program requires torch not pytorch, and yes, i tried finding a cuda version of torch as well, didn't get anythingsorry if this is confusing, if you need me to answer anything else i canwas just attempting to get as much info out as I could so you have a better idea of what's going onpissed and confused because pytorch doesn't get imported in the code but it's mentioned 9 times within the codeim not sure what to make of thathere's the actual code>https://github.com/artidoro/qloraim looking at the literal individual qlora.py file from this github page
>>41144830I think this works more like a way to immediately get results without training. So for example, if one fed like a 10-30 second thing of Twilight to Vocoflex, it would convert the voice to data and allow it to be used as a singing voice... similar to RVC. It seems like their website says that it's optimized for speaking voices, but we easily have singing voices of mares, amongst others. I also noticed an EULA thing on their site, which you can read here:https://dreamtonics.com/vocoflex-eula/Anything to note?
>>41147007In the code:>n_gpus = torch.cuda.device_count()You need pytorch with cuda for that.
>>41147007>No, I only have REGULAR, non-cuda, torch (NOT pytorch)"PyTorch" and "torch" are the same thing, it's just that pytorch is imported as torch. >the program requires torch not pytorch, and yes, i tried finding a cuda version of torch as well, didn't get anythingWhenever you install torch you should follow the instructions on the pytorch website https://pytorch.org/get-started/locally/ because the CUDA builds of pytorch are from a special package index.
>>41147022well, since i cant get it directly thru pycharm, would >>41147046>linkwell, i was typing this out right as the thread refreshed, yeah, if that's the case then all I need to do is somehow get those files and throw em into the right interpreter on pycharm and im golden i hope
>>41145926Yes, I did. It's not "officially released" right now (aside from being posted on /create/).>>41144337Expensive to train.
>>41147019Correction, optimized for SINGING voices, not speaking.
>>41147051You may want to start using anaconda for envs
>>41146884Try fsdp qdorahttps://archive.is/IbOaf
>>41147327>using anaconda for enviromenthow would i go about doing this? I feel like the interface of pycharm actually allows me to click stuff and is especially helpful for dealing with the dependancies/'imports' since it allows me to search for them through it's little search engine thingalso>anacondahad some good experiences with anaconda, forgot what they were but all I remember is anaconda saved my ass many times a few years ago for project shit>>41147565god that is fucking coolonce i figure out how to do shit on my own ill probably try using it >>41147614>nvidia containertf is that?
Why posts gone
>https://writingpeterkaptein.wordpress.com/2024/04/18/blender-stable-diffusion-workflow/I see that there are people slow making attempts at integrating the SD with the Blender for some nicer img2img tools. I don't have powerful enough gpu but I have a feeling that one could run both SD and Blender, render some basic low poly models and convert those images into whatever style one would wish to create (FiM style, a specific pony artist, oil paints or whatever art style once fancy).Im guessing the consistency between the renders is something that will need to be fixed to prevent objects randomly morphing in and out of existence in the background, but fuck it, I feel like we are getting closer to get a real "computer made" animation than all those years when the project started.
Anyone got a download for curated voice samples for the characters?I'm training my own local voice model with Tortoise TTS recalled that at some point there was some sort of effort to get voice samples here
>>41149908>curated voice samplesdo you mean the mega links in OP?
>>41149937oh shoot, don't know how I missed that, thanks anon
>>41149993Yeah, it happens. Btw, how's the training going?
>>41149886 You can just sketch some shapes in paint and get a similar result, unless you want to use stuff like depth map conditioning instead of image to image, for which you would need a 3d scene.
>>41149908post samples. i pay like 100$ a month for eleven labs. i mean its great quality. and only a few min of audio for each voice. but i HATE being tied to the cloud. but tortoise just wouldnt work for me.
>>41151231yes, however instead of drawing everything by hand, the above way would allow to just grab 24fps low quality images and turn them into actual mare animation.
>>41151602From what I've seen the current checkpoints that can do temporally coherent img to img are only acceptable for real life footage and are sloppy. So you can't use ponydiffusion for it, for example.
>>41151869>only acceptable for real life footage and are sloppyits a start, even if it will look like shit there is always possibility that someone will come along and either fix it up or get annoyed enough to make a new tech from scratch just to dab on the coders before him. Don't make dig out the clips from way back when there was only tacotron2 around.
bmp
>>41145405Still experiencing the issue i posted in this post hereeven though i installed pytorch with cuda (pix related as proof) im still getting the same error as the picture in >>41145405I can't imagine what I'm doing wrong considering I now have all the dependencies I can think ofalthough i will mention the instance of (+cpu) is concerning, i did triple check to make sure that i installed the actual cuda pytorchim losing my mind at this shit, idk if my computer files are so messy that something along the lines is getting fucked but i have everything in there, i can't imagine what I'm doing wrong
>>41137243What are currently the best free options we have if I want an entire book read to me (in German)?
>>41152990in terms of TTS?i would imagine local Hay sayjust take the entire book (assuming it's already in german) then drop it into the speech thing and wait a while for it to load (its gonna be fackin huge since its a full book) then boom, download and u have der deutchen flutten shagen audio booken
>>41152843Just make a new environment and run pip install -r requirements.txt for whatever repo you want to use. You can ask chatgpt all this and it will get you through it faster than this thread.
>>41152843You probably still have the cpu version installed then? You should uninstall it
>>41153301+1, there is always some combo of pytorch cuda-tools and the other dependencies hell you need to look out for
>>41153034Big thank, anon
LibriTTS-P: A Corpus with Speaking Style and Speaker Identity Prompts for Text-to-Speech and Style Captioninghttps://arxiv.org/abs/2406.07969>We introduce LibriTTS-P, a new corpus based on LibriTTS-R that includes utterance-level descriptions (i.e., prompts) of speaking style and speaker-level prompts of speaker characteristics. We employ a hybrid approach to construct prompt annotations: (1) manual annotations that capture human perceptions of speaker characteristics and (2) synthetic annotations on speaking style. Compared to existing English prompt datasets, our corpus provides more diverse prompt annotations for all speakers of LibriTTS-R. Experimental results for prompt-based controllable TTS demonstrate that the TTS model trained with LibriTTS-P achieves higher naturalness than the model using the conventional dataset. Furthermore, the results for style captioning tasks show that the model utilizing LibriTTS-P generates 2.5 times more accurate words than the model using a conventional dataset.https://github.com/line/LibriTTS-Pmight give some guidance if anyone here tries to caption the existing voiced data
>>41153953could be useful, if there was a way to get this custom trained, than even a bad emotional audio output could be smoothed out in rvc/sovits at 10% likeness of the character model trained on the same clips.
>>41153740updates on results?
Up.
>>41155310
>need to learn how to build python wheel to solve the dependency problemfug
>>41157006>Lewding the aiOh dear.
>>41157006boop
>>41157006Touching AI mares without permission.
>>41157006Holy shit, Anon. Spoiler this image!
https://files.catbox.moe/utlyfw.wavI miss the days when everyone was shitposting in these threads with whatever voice AIs were available, making full on audio skits or getting the characters to say funny fucked up shit
>>41162218I have started writing my ideas in a notebook, so one day when im not feeling like having my brain liquified from wageing all day long I would love to sit down and return to the shitpost roots.
>>41162218
>>41162218cure rara
>>41162218There are so much cool ai shit popping out but I still haven't got a chance to upgrade my old pc, so it's a bit of abstract kind of pain to know what could be made but have no means to do it.>>41163720I feel ya.
https://vocaroo.com/1ewMMKjktre2UdioCelestia's LightNever thought I'd see, oh, such a sightSunbutt shining in the day and nightEquestria, you make me want to cheerCelestia's light, oh dearEvery pony, dance with glee tonightThat's bright, bright, bright, bright
>>41060498>https://vocaroo.com/1ib4ysUKOcv8Udio anon please come the fuck back and finish this I implore you>>41060678 >https://voca.ro/16cfcyvh3dzAAlso more of this
>>41162218I know that feel.https://u.smutty.horse/lywizqhkwgv.wavhttps://u.smutty.horse/lybxxbvphoc.wavI would be making more but it's not quite as easy these days as it once was with 15.ai. The models on HaySay are capable, but only really if you have a voice that works well as a reference, which I don't. Still possible though with perseverance, I've nearly got a new thing finished.
I tried to write a song for my green with Suno and haysay. I don't think I will end up using it, but it was fun.https://files.catbox.moe/zsyiic.mp3>A lifetime spent jacking it to ponies,>But only loneliness truly knows me.>Empty laps where a mare should be,>Waifu pillow my sole company.>Copium injected, straight to the vein ->"Your pony waifu will come!" - Insane! >Fate is a bitch with a donkey kick,>She laughs as wizards die loveless pricks.>But wait - what light through window breaks?>KA-FUCKING-BOOM, the sky it quakes!>A motherfucking SONIC RAINBOOM,>Wiping its ass with my cynical gloom! >A bridge of glory spans to the stars,>Powered by autism from /mlp/ tards.>Down they prance with heavenly glow,>I'm tripping balls, this can't be so!>Then I see her - oh fuck me blind!>Perfection of plot and grace combined. >Marshmallow fur and indigo locks,>T H I C C thighs to crush my cock.>"Well I say, this realm seems a tad mundane.">Posh ASMR floods my brain.>It's her - RARITY, waifu supreme!>Element of Generosity and wet dreams! >The prophecies true, the shitposts real,>Best pony arrived, my heart to steal.>"Yo Rares, check out that hooman dude!">Rainbow Dash keeping shit rude.>"Hush, the poor dear's overwhelmed, I fear.">Rarity drifts over, poise crystal clear. >Sweet Celestia, dat swaying hip,>Class, sass and ass make the perfect ship.>"Are you quite alright, darling? So sorry for the fright.">Spaghetti spills forth, a pathetic sight."i want snugle ur fluf" True autism prevails.>But she smiles! "But of course, darling! Tis destiny's tale!"
Up from 10.
MARS5: A novel speech model for insane prosodyhttps://github.com/Camb-ai/MARS5-TTS>This is the repo for the MARS5 English speech model (TTS) from CAMB.AI. The model follows a two-stage AR-NAR pipeline with a distinctively novel NAR component (see more info in the Architecture). With just 5 seconds of audio and a snippet of text, MARS5 can generate speech even for prosodically hard and diverse scenarios like sports commentary, anime and more.
DiTTo-TTS: Efficient and Scalable Zero-Shot Text-to-Speech with Diffusion Transformerhttps://arxiv.org/abs/2406.11427>Large-scale diffusion models have shown outstanding generative abilities across multiple modalities including images, videos, and audio. However, text-to-speech (TTS) systems typically involve domain-specific modeling factors (e.g., phonemes and phoneme-level durations) to ensure precise temporal alignments between text and speech, which hinders the efficiency and scalability of diffusion models for TTS. In this work, we present an efficient and scalable Diffusion Transformer (DiT) that utilizes off-the-shelf pre-trained text and speech encoders. Our approach addresses the challenge of text-speech alignment via cross-attention mechanisms with the prediction of the total length of speech representations. To achieve this, we enhance the DiT architecture to suit TTS and improve the alignment by incorporating semantic guidance into the latent space of speech. We scale the training dataset and the model size to 82K hours and 790M parameters, respectively. Our extensive experiments demonstrate that the large-scale diffusion model for TTS without domain-specific modeling not only simplifies the training pipeline but also yields superior or comparable zero-shot performance to state-of-the-art TTS models in terms of naturalness, intelligibility, and speaker similarity.https://ditto-tts.github.io/they have celeb clone examples on their site tha sound pretty good. no weights but the paper has some good info on how they trained it. by KRAFTON which turns out to be the pubg devs so that probably explains why
Articulatory Phonetics Informed Controllable Expressive Speech Synthesishttps://arxiv.org/abs/2406.10514>Expressive speech synthesis aims to generate speech that captures a wide range of para-linguistic features, including emotion and articulation, though current research primarily emphasizes emotional aspects over the nuanced articulatory features mastered by professional voice actors. Inspired by this, we explore expressive speech synthesis through the lens of articulatory phonetics. Specifically, we define a framework with three dimensions: Glottalization, Tenseness, and Resonance (GTR), to guide the synthesis at the voice production level. With this framework, we record a high-quality speech dataset named GTR-Voice, featuring 20 Chinese sentences articulated by a professional voice actor across 125 distinct GTR combinations. We verify the framework and GTR annotations through automatic classification and listening tests, and demonstrate precise controllability along the GTR dimensions on two fine-tuned expressive TTS models. We open-source the dataset and TTS models.https://demo.gtr-voice.com/website (and thus the links to the code/models) doesn't work yet. chinese language model but the idea was interesting
>>41165374Minus one.
Hey guys, I have this clip for the Antithology that will serve as the intro. It's a parody of Top Gear, but I need it to be narrated instead by Jeremy Clarkson. I have no idea how to train voice data. Is there anyone willing to lend a helping hand to get this done?https://files.catbox.moe/tlm61a.mp4
>>41165965>https://applio.org/models/1165806674146230313
>>41165995That's great! Thank you! And so I just extract the archive into the model folder for so-vits?
>>41164790I love it
>>41166007nta, it looks like its a rvc v2 model. https://vocaroo.com/1Mccz5BcQpeUFrom the sound of it it was trained on some yt clips from Top Gear, from all that artificial engine noise going on in the background.If you want/need I could retrain a new model for you.
>>41166074I would very much appreciate that, Anon.
>Tonoight on Bottom Gear>https://vocaroo.com/11P8h43QdEpY
>>41166074Is this the same model?https://voice-models.com/model/1n44tDNY1Ve
>>41166077alright, training started, it should be ready in 6 hours (or 12, depending if I fall asleep in front of my pc).>>41166216>BelAir1603yep, its directing to the exact same link to the same model.
https://vocaroo.com/1lvLqvEHxbMrhttps://pomf2.lain.la/f/7d0mt6y.mp3https://www.udio.com/songs/tgF73cds6Tx2i2DGwehvdzUdioTaste of Friendship[Verse 1]Do you like bananas? (Do you like bananas?)Tell me, tell me true, don't be shy, do you?Magic all around me, ponies let's be clearCelestia's here to hear what's dear[Bridge]Friendship, friendship, oh my dear subjects (Friendship, friendship)But first things first, do you like the taste? (Friendship, friendship)
>>41165965 >>41166007>>41166077 >>41166216>https://huggingface.co/Amo/RVC_v2_GA/tree/main/models/VA_JeremyClarkson_C02>https://vocaroo.com/1j4EOZla9qD8The quality of the model is bit lesser to what I would normally be happy with (as it tries to replicate the thin studio echo that is heard in OG clips) but this is best I could do within one day limit.It's still 110% better than the stuff that the other one. Im really glad to put some support for the Anti this year, put it to good use Anon.
>https://files.catbox.moe/oo2f1c.mp3For Anon that wished for some ai voice shitposting, I hope this is shitposting enough for you. Thanks haysay Anon for adding the styletts2 to it. also I find it funny that SD decided that Jeremy waifu is Apple Fritter when I was trying to get something that resembles Rainbowdash
>>41166778You're a life saver, thank you!
backgrounds safetensor mirrorhttps://drive.filen.io/d/56e3cd3c-51ce-4ec2-a059-f40a384bd0c3#laGTVsncEs1g5dtWV3TiHyDsanSDiClL
https://sonycslparis.github.io/diffariff-companion/This seems nice, though for now they're not releasing a model due to a few circumstances...
>>41167472Man, it's so frustrating seeing all the new cool tech and be told that nobody can play with it.
>>41169555Tell me your secret, purple horse.
>>41169555boop
>>41169555>555Is she oscillating?Damn, no one will have the reference
https://vocaroo.com/1b0ZgDIGOQXJRunning voice conversion on instrumental clips makes the AI generate weird pony beatboxing.
>>41164348I love these! Especially that second one. I can just see Twilight getting paranoid about the thread reaching page 10. Must have taken a lot of effort to generate all of that and include sound effects. Looking forward to the new thing you are finishing.
>>41171220She's on an astable multivibrator.
https://files.catbox.moe/reqmb9.mp3Here, some musical ai slop made with udio and haysay. Zecora - Yellow Pegasus (reggae)
>>41173400>https://files.catbox.moe/reqmb9.mp3Sounds good, actually.Not as good as >>41164251 , though
Derpy Whooves sings When I'm Up (I Can't Get Down) by 'Great Big Sea'.https://files.catbox.moe/d7jfyq.mp3
>>41174009I don't think it's the proper Derpy voice model you used, you would also need to go 12 (or even 23) semitones down to get the pitch right . But hey, it's an attempt.
>>41174019>I don't think it's the proper Derpy voice model you used, you would also need to go 12 (or even 23) semitones down to get the pitch right . But hey, it's an attempt.>
>>41172819You too are a man of culture I see.
>>41174019https://files.catbox.moe/mpukv9.mp3Fixed, though!Thannks fer the feedback.(Unfortunately, When I decided to use the -12 transposition feature the audio became grainy, much like the way it sounds in the end of this track. It's like a horse whisper. A "hoarseness" that obscures the voice of the pone that we are trying to captivate. I didn't even DARE TO try -24 (or 23 or whatever!)Necessarily, I used the so-vits-svc 4.0 model for this one as haysay.ai does not have the 5.0 so-vits-svc Derpy Whooves model available (that is, assuming that we even know of one or even have one.) Hopefully you don't mind the lighter or higher pitch tone of Derpy's voice. I still think it quite sounds a bit like her, especially since she only got like one singing role in the entire show (if even that.))I almost was about to use Rainbow Dash's voice when singing this song,but ahhh...it just didn't sound quite...."Bubbly" enough.Just out of curiosity, when you referred to the "proper" Derpy Whooves model, were you referring to one even hosted on haysay.ai? (the website?) or was it something someone uploaded here and it's still a model that has to be downloaded or linked through a Custom Model URL through Controllable TalkNet?Besides that, I do have this: the attempt that was made at covering the lower octave shift of Derpy's voice. I don't think these files will be of much help but here are the "scratchy" voice files that were NOt used because the cause of "hoarseness" turned out to be the 'transpose' option slider being utilized in Hay Say (I know, really, it crushes me a little though since I thought Derpy's voice was lower, too.) The second option is to just transpose it down afterwards and it works but it sounds a bit like... Derpy is using a voice changer to sound more like a man. So I guess that's it for that, then.>Then... But it seems the only alternative would be... that would mean we would have to... sing with a lower voice.And that obviously only works to the extent that that character was using their 'deep' voice in the first place as it pertains to the information stored in the dataset in general.But yes, in case you were wondering, I /did/ in fact use my "deep voice" on this one.>-12 semitones really caused that bad of an effect?Yes. Maybe 'tis a glitch with the Derpy modelBut it's not of concerning to moi>How do you know thatbecause I tried at '0' transposed again and the voice was once again discern-able as Derpy Whooves but without the whispering.>Proof???Delivery on the next post.
>>41175354>"I still think it quite sounds a bit like her, especially since she only got like one singing role in the entire show (if even that.))"I guess she got none, I was wrong.So sorry everybody. I thought this counted: https://www.youtube.com/watch?v=L6zodtgljFE...But I guess not... eh.>You are STALLIGN!!!Oh, right!right.The files: Sorry, I tried. If anyone knows anything about this (or how to go about fixing this issue, whether within Hay Say or otherwise,) please let me know!!!HEERE THEY AREE:Nope nevermind Catbox responded with zero code—pic related. You'll just have to take my word for it THAT THEY DIDN'T WORK.Maybe someday soon sometiem smutty.horse will make a comeback!!!
>>41175384>pic related>forgothtepicture.jpgHAHAaha-hue hue
>>41145926>https://files.catbox.oe/0twb8w.flacThat's fantastic!
>>41175354Some songs just aren't really fit for making pony AI covers, because the singer's voice is too high or too low pitched to fit any of the poners. Heavy chorus effects or background singers who sound too much like the main singer can also make it pretty much impossible to get a clear enough vocal extract to work with.I've been dabbling in AI covers for a bit lately. Using haysay I generally turn the character likeness all the way down for AI covers. Character likeness comes from the singer/vocals you're using.Overall making AI covers is pretty easy once you get used to the tools you're working with. The real difficulty is finding the right song to cover.Actually I'm wrong about that. The real work (besides the actual original singer) is put in by the people who trained the character models in the first place. We wouldn't be anywhere without those toppest of lads. Unfortunately I have neither the know-how nor resources to get into that, so I have to satisfy myself by picking the fruits of their labors.This is a shoutout to all of you based horsebros that make this possible in the first place.I'm still learning, but I think the Rarity Tears for Fears one I made a few weeks ago came out well>https://files.catbox.moe/p6tfhy.mp3
This thread is full of school shooters anonbros...
>>41175758Fuck, how was I found out so easily?
>>41175758Are you saying this is a banger of a thread?
>>41175758What makes you say that?
>>41145926Nice
>>41175758https://files.catbox.moe/kgkhxr.mp3By the way, haysay.ai seems to be down. I did this with voice-models and easyaivoice.
>>41177948>By the way, haysay.ai seems to be down.Yep, noticed that too. Site is still not loading.
>>41177948>>41178396haysay.ai is back up now and I renewed the TLS certificate while I was at it. Thanks for bringing it to my attention. The server somehow got into a bad state again and was completely unreachable so I had to reboot it.
>>41178595Thanks! Glad to see it's back up!
down ^>:(
>>41180437Konami.
Huh. Have any of the VAs done audio books?
>>41181350
>>41180876Even if they did, the voices they'd use for an audiobook aren't in character for pony, and we don't need audiobook recordings to train pony models. Emily Blunt recorded one chapter of an audiobook, but it was in her normal British accent instead of her Tempest voice. Also, Nicole Oliver narrated that documentary about fungi, but it's unlikely that adding that recording would improve the quality of Celestia models.
>https://files.catbox.moe/5wk6ba.mp3
>>41181670I feel like I should know what this is referencing to, but for some reason my brain can't remember what it is.
>>41181670Holy fuck that got intense, trixie's voice only made it more attention-grabbingpwmn
>>41181670>https://files.catbox.moe/5wk6ba.mp3Why so wide
Improving Text-To-Audio Models with Synthetic Captionshttps://arxiv.org/abs/2406.15487>It is an open challenge to obtain high quality training data, especially captions, for text-to-audio models. Although prior methods have leveraged \textit{text-only language models} to augment and improve captions, such methods have limitations related to scale and coherence between audio and captions. In this work, we propose an audio captioning pipeline that uses an \textit{audio language model} to synthesize accurate and diverse captions for audio at scale. We leverage this pipeline to produce a dataset of synthetic captions for AudioSet, named \texttt{AF-AudioSet}, and then evaluate the benefit of pre-training text-to-audio models on these synthetic captions. Through systematic evaluations on AudioCaps and MusicCaps, we find leveraging our pipeline and synthetic captions leads to significant improvements on audio generation quality, achieving a new \textit{state-of-the-art}.nvidia paper so no weights but maybe this will give some insights>>41181417Ah I was thinking you could take the VA voice then voice clone it to whoever they voiced to build a larger dataset.
>>41181670This is so fucking funny.
>Don't have a functional computer>Found only one online tool for AI voice>Wanted to be part of PPP so bad by creating infinite pony contentDamn. Maybe next year.
>>41183555Haysay is all you really need? https://haysay.ai
>>41183557>Found only one online tool for AI voiceActually I was talking about Haysay, which is great for the character I'm aiming for, but the voice input only (RVC) is a downgrade for me, that can't speak english properly
>https://arxiv.org/abs/2406.07803>Despite rapid advances in the field of emotional text-to-speech (TTS), recent studies primarily focus on mimicking the average style of a particular emotion. As a result, the ability to manipulate speech emotion remains constrained to several predefined labels, compromising the ability to reflect the nuanced variations of emotion. In this paper, we propose EmoSphere-TTS, which synthesizes expressive emotional speech by using a spherical emotion vector to control the emotional style and intensity of the synthetic speech. Without any human annotation, we use the arousal, valence, and dominance pseudo-labels to model the complex nature of emotion via a Cartesian-spherical transformation. Furthermore, we propose a dual conditional adversarial network to improve the quality of generated speech by reflecting the multi-aspect characteristics. The experimental results demonstrate the model ability to control emotional style and intensity with high-quality expressive speech.>https://EmoSphere-TTS.github.io/No code for this one, but they describe their process well enough that I feel like someone advanced with knowledge in python and training could possibly recreate it. I like the idea they propose to control the emotion not by just trying to use text but by dragging the values of "arousal, valence, and dominance" with modified wav2vec 2.0.
General question to Anons in the thread, how many of you guys keep backups of the training dataset (be it the voice, audio clips, images or other)? Like let's say there is a scenario were the google servers and mega totally fuck themselves sideways following up with accidental deletion of all online ai mare content , would there be enough Anons with backups and know-how to get back to the current state of the tools we are using right now?
>>41184281I have a copy of the master files
>>41184281I have backups, and the label files can be used to re-generate the dataset if everything else somehow gets lost.
https://github.com/DigitalPhonetics/IMS-Toucanhas instructions on how to train with your own dataset
Bumpen from page 10
>>41182845what do you mean
>Been while since used Hay Say>Load the local version to make a pony cover>ErrorTemporary failure in name resolution?
>>41186129Is the error sticking around or did it go away after retrying? If it is persistent, try stopping and then restarting Hay:sudo docker compose stopsudo docker compose upI'm not sure why exactly you got this error, but to understand what's going on, refer to the diagram in the Readme under "The Technical Design of Hay Say":https://github.com/hydrusbeta/hay_say_ui?tab=readme-ov-file#the-technical-design-of-hay-sayHay Say is comprised of several "containers" which are like little servers running locally on your machine. When you click the Generate button, the container running the UI makes a webservice call to another container. Specifically, it sends a POST request to the "/generate" method on the container responsible for generating audio for the architecture you had selected (each architecture has its own container). For some reason, the Docker instance on your machine lost its ability to map container names to their local IP addresses (a process called "name resolution") which is a required step before one container can make a webservice call to another. This webservice call happens locally, by the way, and does not reach out across the internet.
Bumo.
>>41186871Excellente.
>>41186453I've tried restarting both Hay Say in the way you mentioned as well as my entire PC, unfortunately the error persists. I have a feeling I somehow broke something over time and should maybe do a fresh reinstall.
>>41185830
>>41187003Nta but recently I was updating haysay and it also stopped working so I had to clean old containers and old models. after that almost all worked except cuda (which worked previously). But I blame beta driver update and used cpu for now.
Could someone tell me what tts was used for the Aryanne voice in the last redub?
>>41188202Wait, there's an Aryanne redub?
>>41188202I can't recall exactly but I either used Google TTS (German) or I just typed up German TTS.
>>41188274I meant Google Translate (German) *
Not sure if anybody needs this, but after few days of struggling to get torch working on old system, the following line of code managed to get it rolling on my old gpu (it uses conda environment of python=3.10.3)pip3 install --upgrade setuptools==70.1.1 wheel==0.43.0pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 'tensorflow[and-cuda]' --extra-index-url https://download.pytorch.org/whl/ --no-cache-dir --force-reinstall
>>41164790Someone should make a metal song using these lyrics
>>41188202>>41188274Anyone want to suggest a voice source for Aryanne? Could easily whip up a dataset to give her a proper voice, ~5-10 mins of clean audio should be plenty.
>>41189164>https://files.catbox.moe/3gcygh.mp4There was an unspoken agreement among mlpol Anons that voice from video would be best fit for her.
>>41181670I am called Guillaume, and that was fucking uncanny.But I laugh my ass off!
>>41189186>Ich liebe PanzerDamn, that's cute.Unfortunately, I think the quality is too bad to be useful...
>>41189186Don't think I can do anything with that, not much of the actual voice and what is there has a lot of noise. Surely must be some similar voices out there though?
>>41189186That sounds like shit though.
>>41189435Since she is OC without any media that would dictate the proper direction of her voice (as opposed to the popular ones like Rainbow Dash Presents or SweetieBot) its free for all choice.But if I were going to make something I would probably with Neco Arc, as mixture of high pitch Japanese cartoony voice with strong German would be the peak of shitposting power level.>>41189677>That sounds like shit though.it does, but you are forgetting it's all about sovl
Theoretically, and I mean that, would there be a way to make some sort of model that would accept at least 10 seconds of voice data in order to make a coherent voice? I know there are some proprietary things out there that do that, but I mean like open source?
>>41189795As you noted in your post they already exist. So theoretically, sure.
>>41189795if you look in the archives for the Blueblood rvc model, that is possible with some voice cloning tech, but even if you give it 100% clean 10s reference line the result it will gives you would be a 1 good clip for 30 terrible ones.
E2 TTS: Embarrassingly Easy Fully Non-Autoregressive Zero-Shot TTShttps://arxiv.org/abs/2406.18009>This paper introduces Embarrassingly Easy Text-to-Speech (E2 TTS), a fully non-autoregressive zero-shot text-to-speech system that offers human-level naturalness and state-of-the-art speaker similarity and intelligibility. In the E2 TTS framework, the text input is converted into a character sequence with filler tokens. The flow-matching-based mel spectrogram generator is then trained based on the audio infilling task. Unlike many previous works, it does not require additional components (e.g., duration model, grapheme-to-phoneme) or complex techniques (e.g., monotonic alignment search). Despite its simplicity, E2 TTS achieves state-of-the-art zero-shot TTS capabilities that are comparable to or surpass previous works, including Voicebox and NaturalSpeech 3. The simplicity of E2 TTS also allows for flexibility in the input representation. We propose several variants of E2 TTS to improve usability during inference. https://www.microsoft.com/en-us/research/project/e2-tts/From Microsoft so now weights obviously. Very cool. Emotion control, speed control, and more interesting explicit phoneme pronunciation. See the examples since it's pretty impressive. Some related stuff I found to thathttps://docs.nvidia.com/nemo-framework/user-guide/latest/nemotoolkit/tts/g2p.htmlhttps://huggingface.co/docs/transformers/en/model_doc/wav2vec2_phoneme
>>41137243Stuff i should of posted here.>>41190109>>41190310>>41190431
>>41190161I'm curious, where do you find your papers? Do you just scroll through the newest arxiv submissions every day?
>>41190446Does Suno give you the option to separate the vocals from the instrumentals, or do you have another way to do it?
>>41190539i had to use vocalremover.org.
>>41190522For arxiv cs.LG yeah. Had to set up a twitter that I have focused on ML/AI stuff to catch papers/gits I miss. Also have a few other places I check that sometimes has relevant stuff pop up.
>>41190801personally i prefer using the Ultimate Vocal Removal with the Kim2 model and then follow up with the reverb removal model (I find it the end result sounds much better when adding my own reverb/echo than try to work with whatever the original source is giving you)
>9
>>41166778>Special thanks in post anti panelVery nice, it was a pleasure to help fellow horse fuckers.
>>41190954i updated the songs, i used UVR MDX-Net Kim_Vocal_2 and Reverb_HQ_by_FoxjoyPinkie Pies Dance Clubhttps://files.catbox.moe/eckq4h.mp3RAINBUTT V2https://files.catbox.moe/1bti1o.mp3Applejack burnt her pie V2https://files.catbox.moe/ju4ig8.mp3
>>41189077>posts a solution to his problem>gets no thank you'sim here to tell you that I, specifically am writing this out of gratitude anon, you're an unsung hero, hope u know that
>>41191684Quite so.
>>41193237Nice songs!So, you made the music with lyrics using "suno", splited music from lyrics using "Ultimate Vocal Remover", feed the lyric to an AI pony voice, and mix back with music and some echo to hide the AI a bit?
>>41195438Yep. workflow was suno, ultimate vocal remover, ultimate vocal remover again (for the reverb), sovits for vocals, combine instruments and sovits vocals in audacity.
>>41193237Dude. Keep doing this. The AIs sound so clean here, plus the songs ain't half bad either.
>>41193237Applejack starts to sound pretty solid at this point.
Reminder that StyleTTS is pretty neat.https://files.catbox.moe/h95xtr.mp3
>>41197696This is a line from a fic I'm writing, to clarify.
>>41196733
>>41188060I see. Well, I'm a complete docker novice. How would I go about cleaning the containers? Also, I attempted to pull the latest versions to fix the issue, and now it errors without even running.
is it worth making a new SD safetensor trained only with every canon screenshot?
Luma seems to do much better when given an image as guidance in addition to the prompt, even if the image in question is AI generated.>Derpy Hooves (MLP:FiM) looks towards the camera and tilts head in a smile with closed eyeshttps://files.catbox.moe/l6ybfj.webmhttps://files.catbox.moe/j7cbi6.webm
>>41199744both of those were ass anonnot saying you should stop posting stuff like this because it's still informative, especially considering how most anons got very different LUMA videos compared to what their prompt implied, but this needs more work, it just aint there yetagain, A for effort though, cool concept of figuring out better LUMA prompting
>>41199732Why? If by only you mean from scratch then that's the most retarded idea ever, it won't generalize at all.You can get show style output in current models by adding "screencap" to prompt, the boorus where the training images were scraped from already had a bunch of screencaps
>>41199749Yeah, but still much better than the garbage Stable Video can do lately. Especially given how you can only request image ref OR text prompt; never both.Google should debut their Veo/VideoFX/Deepmind already, It's already great at horses and their Imagen counterpart is already great at show accurate pony.
>>41199845yeah, the SD animation suffers from constantly re-rending the background causing a weird flickering effect, I feel like this is a kind of problem that could be solved with some decent image separators + composition scripts but trouble is that SD edits the whole image in one go.
>>41162218oh fucking mc boohoo than, fuck off to the afterlife than, cause everyone on here was a bitch that alot of the creators dont wanna do it no more because this fucking forum chases them away because you cunts dont have any fucking common decency and not have practiced the friendship lessons twilight taught us u guys are a bunch of fake bitches kiss my arse and hope to die u bunch of cave dwelling gigachad motherfuckers practice the lessons of friendship then come back to me
>>41200183Hi Thunder"Anon". I wasn't talking about you, your shit always sucked. I was referring to the earlier days when people were posting audio episodes and memes with 15.ai and such. Was before your time.
lol who is that retard ? ur bitchin and complaining like a fucktard get in line you whinny ass bitch
>>41200203>>41200203>>41200183lol i wasnt talking about him, or anyone, people have been here and gone cause you are all bunch of fucktards anyway
>>41200203your shit sucks to eat my hoursecockmuch on it like its peanut butter n jelly
>>41200183B-based?
>>41200304yes indeed based
>>41184281I can see it now. Two centuries from now after the Google-Amazon War of Independence, Twilight STCs are seen as more valuable than breathing air itself.
>>41200386 >>41162218>https://files.catbox.moe/l6x2rd.mp3here is your silly ai shitpost content sir.
Pony?
Uppity.
boops
Lightweight Zero-shot Text-to-Speech with Mixture of Adaptershttps://arxiv.org/abs/2407.01291>The advancements in zero-shot text-to-speech (TTS) methods, based on large-scale models, have demonstrated high fidelity in reproducing speaker characteristics. However, these models are too large for practical daily use. We propose a lightweight zero-shot TTS method using a mixture of adapters (MoA). Our proposed method incorporates MoA modules into the decoder and the variance adapter of a non-autoregressive TTS model. These modules enhance the ability to adapt a wide variety of speakers in a zero-shot manner by selecting appropriate adapters associated with speaker characteristics on the basis of speaker embeddings. Our method achieves high-quality speech synthesis with minimal additional parameters. Through objective and subjective evaluations, we confirmed that our method achieves better performance than the baseline with less than 40\% of parameters at 1.9 times faster inference speed. https://ntt-hilab-gensp.github.io/is2024lightweightTTS/no weights (small model trained on japanese anyway) but interesting
Papez: Resource-Efficient Speech Separation with Auditory Working Memoryhttps://arxiv.org/abs/2407.00888>Transformer-based models recently reached state-of-the-art single-channel speech separation accuracy; However, their extreme computational load makes it difficult to deploy them in resource-constrained mobile or IoT devices. We thus present Papez, a lightweight and computation-efficient single-channel speech separation model. Papez is based on three key techniques. We first replace the inter-chunk Transformer with small-sized auditory working memory. Second, we adaptively prune the input tokens that do not need further processing. Finally, we reduce the number of parameters through the recurrent transformer. Our extensive evaluation shows that Papez achieves the best resource and accuracy tradeoffs with a large margin.https://github.com/snuhcs/Papezmight be useful
https://huggingface.co/fishaudio/fish-speech-1.2>trained on 300k hours of English, Chinese, and Japanese audio data.https://github.com/fishaudio/fish-speechhttps://speech.fish.audio/en/samples/Might be cool. English samples only have the generated audio lol
Hey Anons, Im just throwing an idea out here to see how people feel about it. There are some folks who wish to create a songs/audio shitpost and on the other side there are some contentfags that are little bit burnout from whatever they are doing and could use a small side project to test and bounce around new concepts.How is that sounding? I feel like it would be at the very least it bit less blue than keep on posting a bump every five to three hours?
>>41204065What's the idea?
>>41204146dunno, like people could post a greentext/pastebin of song or story that would make a fun audio/song?
In a nutshell, how does text-to-music/audiocraft work? Do we have models of Daniel Ingram's style to create ponylike ost/bgm?
>>41204915Daniel Ingram wrote the songs, it was William Anderson who composed the background music. Anyway, Meta has released training code and weights for Autiocraft here:https://civitai.com/api/download/models/613602Preparing the dataset requires slicing the BGM tracks into 30 second chunks and writing a text description for each chunk. Is there a music-to-text language model that can help with the descriptions?
>>41205157>music-to-textI would have thought that similarly with the tex2img those models would came with their coders/decoders (like the BLIP Interrogator ) so you could technically reverse the process of generation by "asking" the model what it sees in the image?or maybe the audio models work too differently from the image one?
>>41205157>https://www.yeschat.ai/tag/Music-TranscriptionIt looks like there are tools to transcript the actual musical notes into text , and may also provide a (questionable) ability to tell differences between musical genres. Sadly those also require a a decent amount of shekels to be thrown at their subscription.
https://youtube.com/live/hm2IJSKcYvoPretty impressive. Apparently will be open sourced. All hinges on how well it finetunes
>>41205157Maybe I'm missing a joke here, but that link definitely does not lead to the Audiocraft stuff...
>>41205544https://us.moshi.chat/?queue_id=talktomoshi
What's the opensource equivalent to suno.ai?
>>41205785I copied the wrong link:https://github.com/facebookresearch/audiocraft
>>41206023Nothing.
>>41206039>2024>no local sunoaifags have failed me
>>41206254>>41206023Im pretty sure the udio and suno are just the BARK on steroids (in the same way as how the chatgpy is BERT on steroids).>>41205544>10s input to clone good quality voiceImpressive. They talked about having model understood 17 different distinctive emotions, the speaking in France accent, as pirate and whispering types is pretty neat, it would be pretty useful to be able to feed a text its tts module and have the exact emotional outcome and pace as an output.I would like to see what exact laptop model they used, if it's some bullshit gaymer 16vram than it wouldn't be as impressive as running it on a small 6vram gpu.
>>41137243i renember a guy making a ai south park episode generator,i wonder if something similar could be done with mlp and flash sprites
>>41207064https://www.youtube.com/@blob81/streamsIt has been done in the past, this Anon is keeping it WIP but I have high hopes it will get finish soon(tm).
>>41207106lmao
>>41207106this uses 3d models it seems, it would be cool if it was done using flash
>>41207106It feels so weird seeing redeemed Glimmer standing with unicorn Twilight inside the Golden Oaks Library, but it's exactly the kind of AU timeline I want to see more of in fanworks.
>10
>>41207161Weirdly enough trying to do this in 2d would be more difficult without making it look like ass, since it would either be forced to never change the generic stock vector look or be limited to just loop the exact same handful of animations. With the 3d models there isn't really that much of animations than standard walk and iddle talk however sinc ethey can be rotated and place in front and behind each other allows it to break the "sameness" a little bit.
>>41187003>>41199671Sorry for my delayed reply. What error are you getting after pulling the latest images?Since the original error was with name resolution, I suspect there's an issue with the Docker engine itself. Have you tried reinstalling Docker? Windows/MacOS: https://www.docker.com/products/docker-desktop/Linux: https://docs.docker.com/engine/install/
>>41207837horse
>All /ppp/ is on borrowed time.
>>41209368we are already dead to be honest
>>41209368>"Hey Jewtube I'm a literally who voice actor [starting 10 minutes ago of course] and this video I don't like has AI content in it that was trained on my voice and/or appearance."i give it a week tops before BGM's stuff is baleetedi would recommend just starting a platform specifically for ai, but cia pedoniggers would spam it and get it shut down
>>41209432>what is ponytube
>>41209432How would they know it's a copy? Have to record your own voice? big tech datamining wins either way
>>41209439https://youtu.be/YWdD206eSv0?si=1pH4abj87bJ3LgNx
>>41209427are any of the voice actors and actresses on that voice actor union thingy?
also has anything new coming out in regards to ai voice stuff?
>>41209439They don't give a crap. Copyright trolling has been a blight on Youtube for years.
https://x.com/PrimeIntellect/status/1808639707435446543Cheapest yet?H100s $1.65/hrA100s $0.87/hr4090s $0.32/hr3090s $0.19/hr
Haysay's broken again.
>>41211041Thank you for bringing it to my attention. It should be fixed now. The root cause was the exact same thing mentioned in >>41135074. I have no idea why those config files are getting wiped, though. I might add a cron job to periodically check the files and fix them if needed.
>>41209439I'm pretty sure that adding "No AI voices used in this video were trained on unconsenting people's voices." would protect from this.
>>41211011Ooh, shiny.
>>41137243https://www.youtube.com/watch?v=5yYrh6AcURk
>>41211011
>>41211041These walls of text are always funny to read.
https://github.com/FunAudioLLM/CosyVoicehttps://fun-audio-llm.github.io
>mare at 9
>>41213680Eeyup.
Huh, just realized there wasn't a PPP panel at this year's /mlp/con.
>>41214235yeah it made me sadi always loved PPP panels...
>>41214214Eenope.
>entropy
>>41209432>>41211355It's already over. A lot of AI covers are being taken down. There was a celine dion cover that was perfectly fine for years, then after this it's suddenly been wiped off the map, along with many south park cartman covers.
>>41215845That's how innovation is thrown into a dumpster.
>>41216105but Anon, think about it, if you don't send your shekles to the poor musical publishing companies the artist WILL starve (they only get like a quarter of a penny per sold cd).
>>41215845I hate when they do this>>41216105Well, maybe it's time to switch to something else, like peertube or whatever?Maybe mirror things at first?
>pic relatedNow of course there’s a fuck load of issues with this idea>firstStep 4 implies that you could even train a lora fast enough from start to finish within the amount of time it takes you to reach the context limit while using sillytavern, let alone the amount of time it takes to augment the data even if you made some shitty python program that could magically do it for you, and if there even was one, you would likely need to run ANOTHER model to even augment the data automatically in the background while using silly tavern which considering the fact you’re running a local model on top of constantly running lora training in the background, you’re fucking raping your GPU assuming anyone even has enough VRAM for this shit>secondI doubt you can even swap out Loras mid chat, idk about kobold since i’ve never tried, but text gen webui takes a good amount of time to load a model, I doubt it would be different for a lora and i don’t see why other programs would magically just not have a loading time, so that’s another roadblock in the seamlessness of this idea>thirdEven in a magical world where any of us had the VRAM for this monstrosity, the amount of little python or batch files needed to be made in order to automate this process is fucking annoying>why bother augmenting the dataWish i had the end result of this dude’s stuff but basically this guy tried a similar thing by training a single lora on the raw unaugmented chat logs and in the end it got hyper schizo, and augmentation is the number 1 way to reduce schizophrenia outcomes on a small datasethttps://desuarchive.org/g/thread/95930009/#95933766Wish i had proof because the end result was pretty funny after it spammed about fallout new vegas (the guy said his favorite game was fallout 3)Like i said, wish i had proof of how it turned out but i don’t have enough keywords to find it on desuarchive>why not just use summarization?Summarization has a limit on exact information, it quite literally has to be a summary, details have to be thrown out, and even then it will eat up context as time goes onIn a hypothetical scenario where pic related could even function, summarization would be a great way to buy time for lora training so that the chat logs aren’t only being influenced by recent messages and therefore also making the chat logs higher quality for training the next lora rotation.yeah this plan is shit, but gun to your head, how would you make a plan for infinite long term memory of a local language model?
>>41216858>small dataset unaugmented = overfittingthought of this while writing this out, but couldn't you just prevent overfitting by doing less epochs or steps or something? or would it just simply not retain the information and small, unaugmented datasets are stuck between a rock and a hard place of either being undertrained and therefore pointless or overfit and schizophrenic?
bump
>>41216858I think there are plans on making a """infinite"""" text references with LLM s but so far it looks like it's limited to a very specific and niche kind of work.
https://github.com/KwaiVGI/LivePortraithttps://arxiv.org/abs/2407.03168not sure if it will work for ponies with a finetune but might be interesting for you guys
>>41217435thats pretty neat, even if just now it seems to be slight advanced vtuber avatar it would be nice to see it evolved into some more advanced puppeteer control.
>>41218433yep
Bumperino
>>41219409Subtle as a train wreck.
Get bumped by something unholy.
>>41220021Eh, that's not so bad.
https://voca.ro/1ehkf83fVxmAhttps://pomf2.lain.la/f/u35gqrq6.mp3UdioCelestial DilemmaI never thought I'd buy into the dream,A world of pastel ponies and gleam.But, then I saw her, that royal behind,Celestia's butt just messed with my mind.I swore I'd be tough, not fall in this trap,No magic, no sparkles, but oh, holy crap!
Hey Synthbot (if you're still here), can you try loading the line "Who knows, you might want even want to live there!" (sic) from the pony-speech parquet dataset? I'm getting an error trying to load the audio.
>>41221635I love the retro vibes my dude.
>mare 10Audio bros, were are you at?
https://x.com/yoachlacombe/status/1810964784927367668Maybe that one anon will be interested in doing a finetune with this
>>41222434On a holiday.
https://github.com/kwatcharasupat/bandit>Cinematic audio source separation is a relatively new subtask of audio source separation, with the aim of extracting the dialogue, music, and effects stems from their mixture. In this work, we developed a model generalizing the Bandsplit RNN for any complete or overcomplete partitions of the frequency axis. Psychoacoustically motivated frequency scales were used to inform the band definitions which are now defined with redundancy for more reliable feature extraction. A loss function motivated by the signal-to-noise ratio and the sparsity-promoting property of the 1-norm was proposed. We additionally exploit the information-sharing property of a common-encoder setup to reduce computational complexity during both training and inference, improve separation performance for hard-to-generalize classes of sounds, and allow flexibility during inference time with detachable decoders. Our best model sets the state of the art on the Divide and Remaster dataset with performance above the ideal ratio mask for the dialogue stem.should be having a new version come out since they made a new paper about their benchmark dataset https://arxiv.org/abs/2407.07275https://github.com/kwatcharasupat/source-separation-landing
>>41224222that would be extremely useful, there are several older cartoons that I would love to get the voices of but there is no way to get them split since only existing copy of it is dinky 480p torrent.
>>41221698I'm also getting an error loading that file. Will debug.
>>41221698>>41224554That sound file seems to have been empty in my copy of Clipper's Master File. I checked for other files, and it looks like that was the only one affected. I'm updating the HuggingFace dataset and re-uploading my Master File clone now. Thanks for pointing that out.>>41189435There's an empty SFX file in Master File 2:>https://mega.nz/folder/gVYUEZrI#6dQHH3P2cFYWm3UkQveHxQ/folder/BBo1VSSY>98 Wind Soft Wind, Weather _ Wind Light, Weather 1_ Wind Yellow Sky, Weather _ Wind Siberian, Weather _ Wind Howling Blizzard Weather.mp3There's a non-empty file with the same name under HB03, so I think it's safe to delete the HB06 one.
https://aitestkitchen.withgoogle.com/tools/music-fx?isMusicFxLive=trueFor that anon who likes making music
Hello Degans I have created the most degenerate songOrgy Night With The Mane Sixhttps://files.catbox.moe/jvyanm.mp3
>>41224651>https://files.catbox.moe/jvyanm.mp3
>>41224651bangin'
>suno change UI to new globohomo corpo design When will we get ai that tells managers to stop trying to "fix" things?
>>41225777I hate when that happens. And they always say it's 'new' and 'fresh'.
>>41224597>>41221698Fixed. Both HuggingFace and the gdrive clone both have the updated data now.
>>41225976Thanks
Uppy.
>>41225415>>41224651thanxies
https://emilia-dataset.github.io/Emilia-Demo-Page/
Multi-layer mares
>>41228182On eight pages!
>>41228966That's a lot of mareameters
>>41229541The squaremares will manage.
>>41224651
https://vocaroo.com/12ya4n0XVe5mhttps://pomf2.lain.la/f/elgi8rs.mp3https://www.udio.com/songs/dEEeMSr3DLtMdaZjB6o1fEStay101Applejack's Hoedown MemoriesPrompt: My Little Pony, Applejack, country, female vocalist,Country, Female vocalist, Country pop, Singer-songwriter, Folk pop, Contemporary folk, Northern american music, Folk, Regional music, Pop rock, Love, Mellow, PlayfulLyricsVerse 1:I'm Applejack, the pony with a cowboy hatWorkin' hard on the farm, there's no denying thatWith my trusty hat and boots, I'm ready to rollMy memories on the farm, they never get oldChorus:Hoedown memories, laughter in the airRunning through the fields without a careFrom sunrise to sunset, workin' side by sideApple family love, it's a joyful rideVerse 2:From Sweet Apple Acres to the rodeo sceneI'm the best apple bucker you've ever seenRaisin' the barn with friends, it's a sight to beholdAll these memories with family, worth more than goldChorus:Hoedown memories, laughter in the airRunning through the fields without a careFrom sunrise to sunset, workin' side by sideApple family love, it's a joyful rideBridge:Through thick and thin, we stick togetherFacing challenges, we'll brave any weatherMy little pony friends, we're a colorful bunchWith memories like these, we'll never lose touchChorus:Hoedown memories, laughter in the airRunning through the fields without a careFrom sunrise to sunset, workin' side by sideApple family love, it's a joyful rideOutro:So take a trip down memory lane with meIn Ponyville, where we're wild and freeHoedown memories will always be nearWith my friends by my side, I have nothing to fear.
>>41230081neat work Anon
Page 10 bump.
>>41229694Close your toilet lid before flushing, and keep it closed when not in use. Bacteria builds up in there, especially due to the still water. When water floods the bowl, the bacteria goes all over the place. It might be fine for some time after sanitizing everything, but after a while the bacteria builds up quicker in your bathroom if you're not toilet-conscious.
>>41232738>mane colorsI have a hard time giving the prompter the benefit of the doubt that this could be a coincidence.
>>41232938What do you mean? I didn't prompt the mane colors and I would have preferred them to look more like Trixie's, but that was the way they turned out.
>>41216105>Innovation>"I'm owed all of your work and to replace you"Jew
>>41232938You can thank Google Images and Twitter.
>https://huggingface.co/parler-tts/parler_tts_mini_v0.1>only 647M params>less than 3gb model sizeThis seems like a nice, possibly phone sized, tts model.
>>41233800>Twitterites fuck up the AI with their garbageNow that's just stellar.
Has anyone made nsfw audio of a mare riding (You)? y'know grunting, slapping, moaning sounds, etc. so far I've found a couple audios with pinkie and fluttershy.
>>41233923There was one with Twilight that was made very early on, when 15 was available. There was also a really weird Equestria Girls audio where you have sex with Rainbow Dash, then nuclear war breaks out and everybody dies.
>>41233923Try the PoneAI Drive:drive.google.com/drive/folders/1E21zJQWC5XVQWy2mt42bUiJ_XbqTJXCp
>>41232738
I want to find a short I saw on Youtube of Anon running a "STUD" phone sex line for mares, the voices were really good and I can't find it again. Anyone has it?
>>41235817https://youtu.be/y5KjXz37TYchttps://ponepaste.org/6642>>41181350
>>41212638>text generation fails>errors are thrown instead>speech generation still works>Twi scolding you on fucking up
>>41236287
Has anyone experimented with both Suno and Udio's audio upload feature? pretty gud, might experiment more tomorrow. I think this might be the best way to create "mare-ly" music rn. The most impressive part is how easily it retains the voice of the original singer and the spirit of the song. This shit will be insane in 2 years>Rules of Rarity (AI extension around 30s):https://www.udio.com/songs/f9q48aoVAHLcrnm2oyYDnw>Coinky Dink World (AI extension around 25s): https://www.udio.com/songs/cXDhQugtqoeSmuHfJNduwR>Pinkie's Lament (AI extension around 30s): https://www.udio.com/songs/jDbAtRzcpzprLK4KVav9emhttps://www.udio.com/songs/1nBYDNfjzmx91Mv8dP8Rwahttps://www.udio.com/songs/axvZ569P8nvcDD6WEQJtxQ>I'll Fly (AI extension around 30s): https://www.udio.com/songs/sfmVfvmBnzAJCwjDonojju>Pony I Want to Be (AI extension around 13s): https://www.udio.com/songs/hM6FTmWPTpeog2u9RZ9o7ehttps://www.udio.com/songs/nBduKsoMeWTocLaPTUPy4ohttps://www.udio.com/songs/o28uj8rqKWHoEd3tJZ7RP8https://www.udio.com/songs/4vV67be2eoQEH53iSuz8achttps://www.udio.com/songs/mrnxHf5Mi6AsfPiTKQeP49>Smile Song (AI extension around 30s):https://www.udio.com/songs/4VHb7HDA9y1KhqeSzGLGAwhttps://www.udio.com/songs/vRE3WqsxAeZEKhURQbqQ3vhttps://www.udio.com/songs/b7z29FvNWixtJaD6iAvudphttps://www.udio.com/songs/czoS8J1Hzy11dQzrbzSUAWhttps://www.udio.com/songs/kEhnvpX9j9RMkWuPPnFp9Zhttps://www.udio.com/songs/cf1GX7YfHSZjh7atRsLFnghttps://www.udio.com/songs/j19cTNE7gmuUjY9TPkwE6Mhttps://www.udio.com/songs/u5ds29S37csb558z1oLY8k>Giggle at the Ghosties (AI extension around 33s):https://www.udio.com/songs/evxeR8PzmkuK9WYs7SdZW8https://www.udio.com/songs/uxquZLy9S2cZpqdupBDiNjhttps://www.udio.com/songs/5gJehHnERQdU4D6nkTFyAGhttps://www.udio.com/songs/vfqqdXGQ697wqJ5grVVy5Bhttps://www.udio.com/songs/15PL9FmnfjaLe4eGWpGW4o>Cutie Mark Crusaders go Crusading BGM (AI extension around 30s):https://suno.com/song/54211a67-c9ab-4fef-8a48-122b191c8cd6>Show intro instrumental (AI extension around 27s):https://www.udio.com/songs/aV3hN5Jky3FvdBSAk5xXZhThe AI can also extend backwards, making "intros" for these songs too, which is pretty neat-o. Could have potential for creating mare-ly songs if you just keep extending off these, perhaps with an extra RVC pass:https://www.udio.com/songs/iqVw5UFfsvT9cg8wMuUQ4fhttps://www.udio.com/songs/quUkWowsZ6nzYmgHaxHk2ehttps://www.udio.com/songs/rm7JvPxKM3v2wLRkmuWUSFhttps://www.udio.com/songs/aALnyb32bUtogmDiTcmpmY (japanese smile song kek)https://www.udio.com/songs/m1zwVBEu7KGLPXR2w2aMzWhttps://www.udio.com/songs/77Jsm4M2DWnpAvhevwUyQEhttps://www.udio.com/songs/szFDLUuxRA6caDMmt9Lsuy
>>41237750forgot to mention, but these generations are HEAVILY unoptimized. These could sound a lot better with some experimenting. Towards the end I found out that if you shorten the context length of the extensions to purely the length of the uploaded audio instead of leaving it at max, the extension adheres a lot better to the original audio. There's also other shit like Prompt and Lyric strength, too. Need to experiment more. Plus, I used Udio's shitty auto-generated lyrics instead of using GPT-4, so yeah
Mare me up inside
I'm sorry if that's not the right place to ask, but I was on the bootleg rewatch stream last weekend, and someone queued this gem: https://beautiful.bluefast.horse/jgq2up5.mp4, a cover of S3RL - MTC with Spike and Rarity voices. I couldn't find this on YouTube or PonyTube, and if I understood the chat correctly it's not a new piece of content, so I wonder why? Was it meant to be not really public? It's really nice.
>>41238331That clip was a submission for the first mlp AI redub. As for why it's not up on its own, I suppose that's up to the original creator of it.https://youtu.be/gEXaFVw9J1o?si=d8ObKCJtMzcKtzVy&t=730
>>41237750I'm at work so I will listen to those latter, but thank you for testing those stuff out .
This question has probably been asked to death, but what happened to 15? I remember that when these threads were young, they used to post the best shit, then came the website, and then nothing? Did they ever mention why they took down the website/stopped contributing?
>>41238840Either he got slapped with some lawsuits and his lawyer told him to disappear and avoid all AI stuff for the forseeable future, or he got tired of dealing with all the fags on twitter nagging him to add more voice models or whining about the pony models and moved his research to a private zone.
>>41238840Legal issues.
So is there anything to the level of 15.ai that's freely available anywhere? Or is it all just uberduck tier at best these days? I know there's loads of ways to regenerate audio into a character's voice, but that simply doesn't interest me because it doesn't capture what TTS-generation does at all.
>>41239085Try Haysay. It has one text to speech model and a couple of audio to audio ones.https://haysay.ai/
>>41239097Sadly that again looks like it requires imput audio. I'm talking about PURE text to speech like 15.ai and ElevenLabs did.
>>41239105Oh I was wrong. But man does it sound honestly fucky and limited compared to what used to be free. I'm not surprised but this is closer to shitty uberduck quality more than anything.
>>41239105>https://haysay.ai/Choose Controllable TalkNet or StyleTTS for straight Text to Speech synthesis. Neither of those require reference audio.https://files.catbox.moe/ky4fhg.mp3Not particularly impressive, but actually better than I remember, and it's what we have at the moment.
>>41239115Unfortunately 15 was probably the result of far too many stars aligning at just the right time. Chances are we're never gonna get something with that level of both quality and ease of use again.
>>41237750 >>41238481Now that I've listen to some of those, I feel it's a little bit of mix bag, some vocal extensions are like 95% same as the input, then there are the ones that differ between input and output as much as VAs clips between early seasons and late seasons. Then there are some of them that sound like a decent quality fan VA imitating the mane6 followed up by a the ones that sound like lower quality fan VAs.But over all its surprisingly positive result, as they are able to get to such good quality without having extra fine tuned on the show clips.>>41239187annoying,we seen even in just this thread that there are new algorithms and tech to remake or improve tts tech to get the emotion control and quality to current tts tech, but there just aren't HIGH level academic anons here to understand how this nonopen tech works and reverse engineer it for people to use.
>>41239382Yep. I'm moreso focused on the potential of this particular technology and how it might improve in the next one or two years - perhaps to an exponential degree. This technology isn't even 6 months old and it's already progressing at such an insane degree. After looking into it a bit more, I discovered that there's an even better tier audio model locked behind the premium tier for Udio in particular, so I might shell out some shekels and experiment even further. Exciting times ahead.
>>41239459if there would ever be a leak of on of those large song generating models, it would help tremendously with a lot of other project I keep in back-burner, for example their cloning abilities could possibly be used to extend few seconds of background character speech to a few minutes monologue and allow for creation of better artificial data set to make even better tts/voice conversion model down the line.
Precautionary bump.
UNLEASH THE PONIES!
>https://www.youtube.com/watch?v=7_LW7u-nk6Q>bipedal robot tough with reinforced machine learningI desperately need a robot mare maid.
Bump
>>41233923asking this too but of glimmer telling me to cum and calling me a good boy
TTSDS -- Text-to-Speech Distribution Scorehttps://arxiv.org/abs/2407.12707>Many recently published Text-to-Speech (TTS) systems produce audio close to real speech. However, TTS evaluation needs to be revisited to make sense of the results obtained with the new architectures, approaches and datasets. We propose evaluating the quality of synthetic speech as a combination of multiple factors such as prosody, speaker identity, and intelligibility. Our approach assesses how well synthetic speech mirrors real speech by obtaining correlates of each factor and measuring their distance from both real speech datasets and noise datasets. We benchmark 35 TTS systems developed between 2008 and 2024 and show that our score computed as an unweighted average of factors strongly correlates with the human evaluations from each time period.might be interesting for some here
>>41237172
Experiments with transforming speech features in RVCAdd gaussian noise with scale>0.0: https://files.catbox.moe/7dycpu.mp3>0.1: https://files.catbox.moe/btzkpu.mp3>0.2: https://files.catbox.moe/6z3x04.mp3>0.3: https://files.catbox.moe/co3z3f.mp3>0.4: https://files.catbox.moe/hsk53e.mp3>0.5: https://files.catbox.moe/ae0qgk.mp3>1.0: https://files.catbox.moe/142845.mp3>3.0: https://files.catbox.moe/h4og7u.mp3(I'm not sure about the implementation of the rest of these functions, I just asked Claude to write them)Gaussian filter (kernel size 7, sigma varies):>sigma=0: https://files.catbox.moe/nki99s.mp3>sigma=1: https://files.catbox.moe/ex83vn.mp3>sigma=2.5: https://files.catbox.moe/68nds0.mp3>sigma=5: https://files.catbox.moe/7eloma.mp3>sigma=10: https://files.catbox.moe/icxfph.mp3>Kernel size 15, sigma 30.0: https://files.catbox.moe/5w9s6h.mp3Subtract scaled Laplacian filter:>0: https://files.catbox.moe/mucqxq.mp3>0.3: https://files.catbox.moe/s9q56y.mp3>0.6: https://files.catbox.moe/vjjpcb.mp3>1.0: https://files.catbox.moe/tuzxs4.mp3PCA (50 components, showing first 3) (86% explained variance):Add to first component:>0: https://files.catbox.moe/0p2bqd.mp3>2: https://files.catbox.moe/bkwjyz.mp3>4: https://files.catbox.moe/w16ejj.mp3>-2: https://files.catbox.moe/4ofoai.mp3>-4: https://files.catbox.moe/ftpzpo.mp3Add to second component:>2: https://files.catbox.moe/36tg4x.mp3>4: https://files.catbox.moe/j3wxq8.mp3>-2: https://files.catbox.moe/cwxa64.mp3>-4: https://files.catbox.moe/gado1x.mp3Add to third component:>2: https://files.catbox.moe/jawiov.mp3>4: https://files.catbox.moe/830iey.mp3>-2: https://files.catbox.moe/c953iy.mp3>-4: https://files.catbox.moe/jfk71a.mp3Some interesting combinations:[0, 4, -4]: https://files.catbox.moe/obcqbg.mp3[4, 0, -4]: https://files.catbox.moe/b3ed2u.mp3PCA (10 components) (41% explained variance): https://files.catbox.moe/9horpd.mp3Code: https://github.com/effusiveperiscope/raraai/blob/master/2_featureexpl/testing.ipynb
>>41244976Now do it again, but with a line from the actual character:Add gaussian noise with scale>0.0: https://files.catbox.moe/h4va4y.mp3>0.1: https://files.catbox.moe/q91pta.mp3>0.2: https://files.catbox.moe/ste0ox.mp3>0.3: https://files.catbox.moe/nuks4d.mp3>0.4: https://files.catbox.moe/o3xucd.mp3>0.5: https://files.catbox.moe/mlpskx.mp3>1.0: https://files.catbox.moe/fvi6rd.mp3>3.0: https://files.catbox.moe/cofucz.mp3(I'm not sure about the implementation of the rest of these functions, I just asked Claude to write them)Gaussian filter (kernel size 7, sigma varies):>sigma=0: https://files.catbox.moe/uuudlh.mp3>sigma=1: https://files.catbox.moe/83nv38.mp3>sigma=2.5: https://files.catbox.moe/vx00ia.mp3>sigma=5: https://files.catbox.moe/vx00ia.mp3>sigma=10: https://files.catbox.moe/edw1hq.mp3>Kernel size 15, sigma 30.0: https://files.catbox.moe/kd7vxw.mp3Subtract scaled Laplacian filter:>0: https://files.catbox.moe/jhv6ow.mp3>0.3: https://files.catbox.moe/y0uh36.mp3>0.6: https://files.catbox.moe/43wdcy.mp3>1.0: https://files.catbox.moe/5mhy3j.mp3PCA (50 components, showing first 3) (88% explained variance):Add to first component:>0: https://files.catbox.moe/2e7pcy.mp3>2: https://files.catbox.moe/pr5qcf.mp3>4: https://files.catbox.moe/w3r928.mp3>-2: https://files.catbox.moe/rnsn5m.mp3>-4: https://files.catbox.moe/0gdho5.mp3Add to second component (pretty interesting):>2: https://files.catbox.moe/jlsq2f.mp3>4: https://files.catbox.moe/wuibwr.mp3>-2: https://files.catbox.moe/gceiip.mp3>-4: https://files.catbox.moe/uoj6ga.mp3Add to third component:>2: https://files.catbox.moe/gmynkz.mp3>4: https://files.catbox.moe/iaam1c.mp3>-2: https://files.catbox.moe/cobbez.mp3>-4: https://files.catbox.moe/5r0k94.mp3[0, 4, -4]: https://files.catbox.moe/19zl30.mp3[4, 0, -4]: https://files.catbox.moe/0jsw6o.mp3PCA (10 components) (46% explained variance): https://files.catbox.moe/tueqn7.mp3
what happened to 15
>>41245510see >>41238840
>>41244976>>41244980berry interesting. is this code editing the audio before it gets sorted inside the rvc or there are some code elements that are changing how the model operates (like loras with image ai and such)?
>>41246044It is modifying the hubert speech features that are input into RVC.The rationale here is: when non-mare is converted into mare voice it doesn't sound the same. Likely because the model only sees speech features from the target mare in training and not other speakers, and features from non-mare differ in a significant way that causes the resulting artifacts. So maybe there is a way to make non-mare feature inputs closer to the target mare using an added network that would perform better than the current retrieval based method RVC is using (I'm not sure exactly how much testing went into choosing the "retrieval" methods or "cluster" methods for this purpose; if there is any documentation for this it is probably on some Chinese forum or website where they were developed)The approximate idea is that this is a denoising problem. You could add noise/convert multiple times through different RVC models/perturb PCA components/apply conv kernels to generate a synthetic dataset of "noisy" features from real lines, then learn a mapping from noisy features to denoised features that sound like the target character. Part of what I am trying to figure out here is whether there is a sensible way to perturb the features that will preserve enough low-level information that a model could reasonably reconstruct the input (as opposed to distorting the signal into complete meaningless garbage), and whether those perturbations plausibly sound like the sort of artifacts that happen when a non-mare speaker tries to convert to mare. so-vits-svc 5.0 has this baked into the training process, where they add gaussian noise to whisper features for what I can only imagine is a similar reason.I'm sure there is a more statistically sophisticated way of doing this but that's above my education grade. Each speech feature is 768 long so there's no real retard-friendly way to "visualize" them like you can image models, so "hearing" what RVC thinks is the next best thing.What's kind of interesting is how the transformations that correspond to image operations (gaussian blur/laplacian filter sharpening) actually seem to match what happens to the resulting audio--the gaussian filter audibly "blurs" the syllables and makes everything sound "fuzzy", and the subtract laplacian brings out random details and makes it more defined. The "fuzziness" sounds similar to what I get whenever I try to convert my voice to Twilight. Some of the PCA perturbations also sound similar, seeming to modify formants of the output to the point where it still sounds like the correct words, but the speaker doesn't even sound like RD (so it might not even need a NN)?
>>41246189The reason I'm looking at this is because while making this bumper:https://files.catbox.moe/vxucyp.mp3I had the idea to feed in StyleTTS2 generated audio thru RVC and it sounded significantly better, quality wise, than any conversions using my own voice directly, even if the prosody suffered (which actually was OK for the purposes of a radio bumper because it doesn't need to sound like a plausible delivery in that context). I also fed an RVC generated reference into the style encoder, since I also noticed that the "style encoding" pretty much seemed to copy the prosody of the input audio--the "style vector" actually seems to smuggle in a lot more information than you'd expect from the term "style", so it makes StyleTTS2 act like a sort of "fuzzy voice conversion". So the whole pipeline was RVC -> StyleTTS2 -> RVC which is a bit too convoluted than I'd like for any longer projects.
https://github.com/fishaudio/fish-speechnew update to 1.2. looks like they keep their website version ahead by 1. https://www.bilibili.com/video/BV1wz421B71D/
>>41246189>>41246213so if i understand correcly, you are trying to find a way for the RVC to take a bit noisy recording (like what most people have) that has voice not in the range of the model and try to squeeze more quality out of it in the outputs? Thats neat.>RVC -> StyleTTS2 -> RVCAh, the classic "use ai to make ai better" technique, this is a very dinky way to work but whenever models refuse to work with my vocals (with is 9 out of 10 times) its a helpful trick to finish a project.
Putin Eats Poutine: https://www.udio.com/songs/tUaTeSxWiWB78KdmfafPNzCelestia Eats Cake: https://www.udio.com/songs/sdPiXv4DWSG4EvmCk2YYRGhttps://vocaroo.com/1iwpscmGqkRYhttps://pomf2.lain.la/f/b90j92f6.mp3Udio v1Celestia's Sweet Secret (Celestia Eats Cake) ext v1.2.2[Verse 1]In a castle with grandeur, oh (whoa whoa)A princess with a secret way (hey hey)Celestia's got a sweet delightIn a slice of cake, day or night[Pre-Chorus]All she can think of as she reignsIs that sugar rush she craves[Chorus]Celestia. Eats cake.Celestia. Eats cake.Celestia. Eats cake.Celestia. Eats cake.Every day![Verse 2]In the grand dance halls we see (hey hey)Robots groove in symmetry, oh (whoa)Celestia’s got that royal tasteFor that cake she won’t waste[Pre-Chorus]In the beats of the night, she dreamsOf the sweetest, creamiest themes[Bridge]"Whoa whoa, wait a minute!Are you saying the queen of EquestriaHas such a sweet, delightful sin?"Yes![Chorus]Celestia. Eats cake.Celestia. Eats cake.Celestia. Eats cake.Celestia. Eats cake.Every day!
How would I go about changing the lyrics to a song like what this person did here? https://www.youtube.com/watch?v=VornCFk3Z0UThey changed the word 'starboy' to 'starpony' and I was just wondering how I could do the same thing. I already know how to use RVC to change the voice, I just need to know how to change specific lyrics
>>41242583
>>41247408If you can make the same quality + pronunciation as the original vocals converted with rvc/sovits that just cut the single word in audacity and swap it out.However most of the time you may end up needing to re-sing entire sentence (or even the song ) to make that work.
>>41247408>>41248476Also, you probably need to extract the music beforehand for that too.Some AI are specialized in to that.
>whisper-medium transcribes this as: "Equestria, Orlando, France!"https://files.catbox.moe/wx4hzu.mp3
>>41237750Been having issues with Suno's audio upload feature, where it "successfully" uploads but then simultaneously throws a network error. Would be doing the exact same otherwise; hopefully it clears up.Vocal coherence in your examples is more than I was expecting, but still veers off enough to want improvement. If the samples separate well enough using UVR (or similar) maybe it can be fed back through RVC to re-ponify?
>>41249969Sound accurate, not sure whats the problem Anon ^:)
>>41252350
>>41251308That doesn't look healthy.