Hey, one of the Suno founders/creators of Bark here. Thanks for all the comments, we love seeing how we can improve things in the future. At Suno we work on audio foundation models, creating speech, music, sounds effects etc….
Text to speech was a natural playground for us to share with the community and get some feedback. Given that this model is a full GPT model, the text input is merely a guidance and the model can technically create any audio from scratch even without input text, aka hallucinations or audio continuation.
When used as a TTS model, it’s very different from the awesome high quality TTS models already available. It produces a wider range of audio – that could be a high quality studio recording of an actor or the same text leading to two people shouting in an argument at a noisy bar. Excited to see what the community can build and what we can learn for future products.
Please let us know with any feedback, or if you’re interested in working on this: bark@suno.ai
This tech will be used by crooks to automate attacks. Generate the language using GPT-4 and the audio using Bark, and then start making phone calls. Because it’s open source, all you need is GPUs. This is not a criticism. I’m impressed and grateful for the openness. Everyone needs to wake up and recognize that these attacks are coming at us essentially right now.
As technology gets more powerful more quickly, and the rule of law becomes more and more unable to prevent societal damage, this response becomes woefully inadequate.
For reference, look at how the societal damage of social networks has been handled: too little, too late. Same goes for RentTech.
But, I don't know the solution. The common computer has become so powerful that we cannot simply rely on inaccessible materials to prevent the danger of overly-powerful tech spreading too fast, as we do with bioweapons or traditional WMDs.
Like kitchen knives, which are used to end unhappy relationships. Is this really an argument we should be talking about? Wouldn't you feel silly presenting such an argument, to a household knive maker?
That comparison holds up better if you imagine a world without knives or sharp objects of any kind. Now you can suddenly do tremendous harm by wielding a pointy stick. I don't think it's reaching to point out the dangers you just introduced.
Except that world used to exist, back in the stone age, and we're all far better off now because we didn't choose to live in fear of the misuse of powerful tools.
Recommend to encode sub-audible tracking symbols in synthesize speech patterns that includes GPS, IP, timestamp, country of origin. We do that already in bootleg movies so we can apply similar methods in synthesizes speech.
Hmm, I don't know about that. The data rate used by civilian GPS (L1 C/A) is only 50 bps. The symbols are normally spread over a couple MHz of bandwidth to make it possible to recover at levels below the thermal noise floor. I see no reason why the same thing couldn't be done at baseband, adding an imperceptible bit of extra noise to an audio signal.
Of course, you wouldn't encode real-time navigation data, but a small block of identifying text. Either way, though, someone without a copy of the spreading code isn't going to notice it or decode it. Given enough redundancy in both the time and frequency domains, removing it wouldn't be easy either.
The real problem is that bad actors would simply encode some other person's coordinates/metadata into the recordings they produce, and we'll have been trained by then to blindly accept the presence of these markers as strong evidence of guilt.
thanks, the model itself is a pretty vanilla gpt model based heavily on karpathy's nanogpt, so should not need too many bells and whistles to get it running on specific architectures. that said i have very little experience with platform specific development, so would looove some help from the community :)
Could you stick a torch.compile in the inference and training code, maybe gated behind a flag? This should help AMD/Nvidia performance (and probably other vendors soon) significantly.
A serious nod to Karpathy here. They could have chosen any other Transformer architecture, but chose perhaps the most reachable one - in the literal sense.
I like the emphasis tags, it's something that is not seen with a lot of these transformer models.... things like [laughs] makes alot of sense... i could see where hundreds or possibly thousands of emphasis style tags could be added to support a vast array of intonations in human speech. i.e. [yells], [shouts], [cries], [crying], [whispers], [sarcasasm], etc
Very cool. Side note: bark-gpt.com is already taken for a dog translator: "The world’s first AI powered, real-time communications tool between humans and their furry best friends."[0] I only know this because my law firm partner's name is Bark, and I wanted to automate some legal work and name the software "Bark GPT" after him.
Honestly you need to annotate datasets for barks and we're building this up with our users. Otherwise it's unclear what's happening even with video (we tried to build a YouTube/Instagram scraper and it was very unreliable even with vet tech annotations).
I already know when my dogs needs to eat, drink, shit or pee or go for a walk or play because he usually will tell me.
Choosing to ignore your dog won't change because some magical AI can now translate it to -Im fine, Im only barking because you're an awesome being, keep your subscription humaaan-
But can your dog (translator) say "I love you?" In a doggy-voice? Replika proves people will pay for this and convince themselves it's real, because they want it to be.
Am I hallucinating or didn't several of the examples have background audio artifacts, like it's been trained on speech with noisy backgrounds, I'm guessing audio from movies paired with subtitles? Having random background audio can make it quite hard to use in production.
>Am I hallucinating or didn't several of the examples have background audio artifacts, like it's been trained on speech with noisy backgrounds, I'm guessing audio from movies paired with subtitles? Having random background audio can make it quite hard to use in production.
The other side of that problem is an opportunity. That's why the same model can also generate music, background noise and sound effects. And it's just because the prompt specifies those things explicitly. The input is truly semantic, so the output is rich and reflects that context. Is your input text sounds like it came from a speech, then there's a high chance your output audio will sound like a megaphone in a public space with crowd reactions and maybe even applause.
yeah sometimes there are definitely artifacts. technically they can be removed pretty easily with another model (like denoiser from FB) but for now we wanted to keep it simple to learn to control these things better through prompt engineering. Like when using a high quality input prompt it generally continues with high quality
At least in the last example, with the man and woman and the expensive oat milk, the background noise seemed to fit a likely public conversation scenario. I wasn't sure if it was accidental or not.
Man I know this is HN, and I know we have a certain decorum we should be maintaining, but with the recent activity in this field the most appropriate response to these posts is "4bit when?" or "f16 when?". Not sure which one is applicable. I am having no luck running it on a 6GB vram gpu, so I guess its the 16 bit floating point one.
related to this - to those releasing models, it would be great if you could share how much VRAM is required (seems very common for this key piece of info to be missing).
I'm successfully running it on a 12GB GPU (while it downloads some 12.1GB of model data on first run, the highest GPU memory usage was ~6.5GB, settling back down to around 5GB), however the results are nothing like the samples given on the github page. Using the exact code given and in the runs I've tried the results are rather terrible.
I'm not being negative -- some of the samples are really neat on their page -- and I know there is some idiosyncrasy of my setup that is causing issues, though it is a pretty typical conda + pytorch with CUDA 11.8.
Playing with the text and waveform temp from their defaults 0.7 is yielding some semi-decent results, but it feels essentially random.
A few years a back someone had the genius idea of making a robo-answering phone program that would give vague but encouraging replies when it received some unsolicited sales pitch. Although it was a fixed sequence of responses, it fooled some callers for a surprisingly long time.
Someone needs to hook create the plumbing to capture speech to text, feed it to a GPT script that has been told how to reply to such call center calls, then send that back through a TTS generator like this one.
To overcome any latency issues, it could build in a ploy to buy time like the old script did, eg, make the robo-answerer sound like a somewhat addled old man who has to think before each reply, perhaps prefixing responses with "hmm, ahh, ..." to buy time to generate the response.
It seems like a lot of the entries in TTS are either close sourced saas apps or something like this with limitations on customizing it. It seems clearly inevitable and likely only months away that a high quality unrestricted open source option for things like voice cloning will emerge so i'm not sure why these projects are even really bothering trying to stop it. I think in order for TTS to have its StableDiffusion moment it will just be a matter of an unrestricted easily trainable open source model.
No, the power and money to run high quality servers like ChatGPT4 and MidJourney or DallE. Sure there's local-able alternatives, but they're lower quality and lower bandwidth and not being used as a business proposition.
Just like a private individual can own lock smith tools and play around with locks... don't go basing your business of supplying them wholesale to the general public worldwide for free.
Local LLMs are in production as the backbone of highly valued businesses. That will increasingly be true with the advent of commercially licensed foundation models currently in training such as StableLM and RedPajamas.
Is there likely to be a way to stream this audio in the future? As in, here's an incoming stream of text, generate the audio on-the-fly instead of all at once.
Great news! It's astounding how quickly technology is advancing. Only yesterday, I was wondering about when a new model for text-to-speech would be developed, and today a game-changing model has been released! This new model is simply incredible!
Any idea what the training data for this is? Looking at the model, it looks like it is literally just copy-paste from Karpathy's nanoGPT, so the training data is what's most interesting. Pretty amazing anyway.
I found a secret demo page that shows in real time how they assess any sound file's mood swings along with number of detected laughs, coughs, etc. Guessing that ability is involved somehow.
Imagine Torvalds saying the same in the context of linux- 'to mitigate misuse of this technology, we limit the audio history prompts to a limited set of Suno-provided...'
On the other hand, Russian was disappointing. It put a stress in one word incorrectly (it confused the grammatical form, used the genitive case instead of the accusative) and in general sounded strange.
The fact that this is open source and can generate more thann just speech is really nice, but for speech itself, it's much lower quality than what Eleven Labs provides.
All the open source models I've seen so far have this weird kind of neural fuzziness to them. I don't know what Eleven does better, but there's definitely a big difference.
"However, to mitigate misuse of this technology, we limit the audio history prompts to a limited set of Suno-provided, fully synthetic options to choose from for each language."
Isn't this open source and can be easily removed or am I missing something?
Have you seen any information on how to create the .npy files for custom voices, though? That's where I hit a wall. There must be some process that takes a .wav or similar audio file and creates the (relatively small) history prompt data.
Some of it is very impressive although some of it seems about equal to the TTS built into my phone. How long until someone can package this up and make a program that takes in epubs and spits out mp3s?
Does it sound fairly robotic/static-y to anyone else or just me? Doesn't sound any better than any other TTS software I've tried and in fact sounds a bit worse, like it's noisy.
>Bark has the capability to fully clone voices - including tone, pitch, emotion and prosody. The model also attempts to preserve music, ambient noise, etc. from input audio. However, to mitigate misuse of this technology, we limit the audio history prompts to a limited set of Suno-provided, fully synthetic options to choose from for each language.
It's not immediately clear how the audio history prompts are created.
I don't know how they're made exactly, but one can just edit the code a bit and delete the restriction to just the given audio history prompts. It's literally just enforced, affect, with a simple "assert" command.
history prompts are just unconditionally generated TTS from the same model. any of those can be used as history, but for convenience 10 are provided for each language (to generate things with consistent voices)
There appear to be several ways to do it, but what worked for me was:
from scipy.io.wavfile import write as write_wav
write_wav("c:\\wherever\\whatever.wav", SAMPLE_RATE, audio_array)
If you want to play it right away, you can use:
from playsound import playsound
playsound("c:\\wherever\\whatever.wav")
What I wasn't able to do was make torchaudio play it directly, without creating an intermediate .wav file. Apparently there is a live-playback backend on the Mac, but not on on any other OS.
Text to speech was a natural playground for us to share with the community and get some feedback. Given that this model is a full GPT model, the text input is merely a guidance and the model can technically create any audio from scratch even without input text, aka hallucinations or audio continuation.
When used as a TTS model, it’s very different from the awesome high quality TTS models already available. It produces a wider range of audio – that could be a high quality studio recording of an actor or the same text leading to two people shouting in an argument at a noisy bar. Excited to see what the community can build and what we can learn for future products.
Please let us know with any feedback, or if you’re interested in working on this: bark@suno.ai