Thanks! Would "under 10G" also include 8 GB, by any chance? Although I do die inside a little every time I see "install Torch for your CUDA version", because I never managed to get that working in Linux.
It actually uses less than 3 GB of VRAM. One issue is that the research code is actually loading multiple models instead of one, which is why it was initially reported you need 8 GB if VRAM.
However, it cannot be used for the same use case because it’s currently very slow, so real time usage is not yet possible with the current release code, in spite of the 0.15 RTF claimed in the paper.
Thanks, but I don't think I'm going to reinstall my entire OS to run these. I'll see if I can get Docker working, it's been more reliable with CUDA for me.
The live test on https://play.ai/ didn't work for me in firefox. swapped to chrome and it worked quickly. I cloned my voice in 30s and was instantly talking to myself. This would easily fool most people who know me. Wild stuff.
I still use FF, for now anyway, so I’m not trying to be a dick here, but we’re talking less than 4% market share, so it’s hard to fault a small team for prioritizing the 82% they get with Chrome+Safari
What's SOTA for open source or on-device right now?
I tried building a babelfish with o1, but the transcription in languages other than English are useless. When it gets it correct, the translations are pretty perfect and the voice responses are super fast, but without good transcription it's kind of useless. So close!
I work at Cartesia, which operates a TTS API similar to Play [1]. I’d be willing to venture a guess and say that our TTS model, Sonic, is probably SoTA for on-device, but don't quote me on that claim. It's the same model that powers our API.
Sonic can be run on a MacBook Pro. Our API sounds better, of course, since that's running the model on GPUs without any special tricks like quantization. But subjectively the on-device version is good quality and real-time, and it possesses all the capabilities of the larger model, such as voice cloning.
Our co-founders did a demo of the on-device capabilities on the No Priors podcast [2], if you're interested in checking it out for yourself. (I will caveat that this sounds quite a bit worse than if you heard it in person today, since this was an early alpha + it's a recording of the output from a MacBook Pro speaker.)
Only thing missing (for me) is "emotion tokens" instead of forcing the entire generation to be with a specific emotion, as the generated voice is a bit too robotic otherwise.
> based on flow matching with Diffusion Transformer
Yeah that's not gonna be realtime. It's really odd that we currently have two options, ViTS/Piper that runs at a ludicrous speed on a CPU and is kinda ok, and these slightly more natural versions a la StyleTTS2 that take 2 minutes to generate a sentence with CUDA acceleration.
Like, is there a middle ground? Maybe inverting one of the smaller whispers or something.
To be clear, what I mean by realtime is full gen under at most 200ms so it can be sent to the sound card and start playing, not generating under the amount of time it would take to play it, which would add that as an unusably long delay in practice.
I suppose it might be possible to do it with streaming very short segments, but I haven't seen any implementation with it that would allow for that, and with diffusion based models it doesn't even work conceptually either.
I'm not sure what you mean fully, this is TTS, but it sounds like you're expecting an answer about transcription
So its both hard to know what category you'd like to hear about, as well as if you do mean transcription, what your baseline is.
Whisper is widely regarded the best in the free camp, but I wouldn't be surprised to see a paper of a model claiming better WER, or a much bigger model.
If you meant you tried realtime 4o from OpenAI, and not o1*, it uses whisper for transcription on server, so I don't think you'll see much gain from trying whisper. my next try would be the Google Cloud APIs, but they're paid and with regard to your question re: open source SOTA, the underlying model isn't open.
But also if you did mean 4o, the transcription shouldn't matter for output transcription quality, the model is taking in voice (I verified their claim by noticing when there's errors in the transcription, it answers correctly)
* I keep messing these two up when talking about it, and it seems unlikely you meant o1 because it has a long synchronous delay before any part of the answer is available, and doesn't take in audio.
If you did mean o1, then, I'd use realtime 4o for TTS, and have it natively do the translation, as it will be unaffected by errors in transcription like you're facing now
This is awesome! Over the summer I wrote API clients for both Go [1] and Rust [2] as we were using Play in my job at the time but there was only Python and Node SDK.
Has anyone done a comparison of combined speech to text and TTS vs speech-to-speech for create audio only interfaces? Particularly curious around latency, and quality of audio output.
Does anyone know of a TTS mod that could convey feeling? E.g. ebook reading for novels? Or can one request feeling in any of the models of this discussion?
First question, does it pronounce numbers > 9 correctly? At least OpenAI's model doesn't perform at all, marking garbage out of almost every number it finds. I actually dont remember if I checked with EleventLabs... But I was shocked enough that in 2024, someone could release a TTS model that doesn't do numbers correctly. As if the AI industry was approaching Xerox level of failings. However, the TTS models are way worse then the Xerox compression algo ever was.
I believe verifying numbers up to at least 100000 should be a requirement for new TTS models.
Is there any way to use the TTS on its own? I maintain an obsidian TTS plug-in, and am starting to add new TTS providers (its just been OpenAI thus far). From the documentation at https://docs.play.ai/documentation/get-started/introduction, it looks like their API seems to couple it to an LLM for building conversational agents. Seems like it might be nice to use standalone as just TTS.
Sounds quite good, but this prompt is NOT what I'd expect an automated system to feed into it:
“I’ve successfully processed your order and I’d like to confirm your product ID. It is A as in Alpha, 1, 2, 3, B as in Bravo, 5, 6, 7, Z as in Zulu, 8, 9, 0, X as in X-ray.“
Phone numbers and others were read nicely, but apparently a string of alphanumerics for an order number aren't handled well yet.
Sorry, Do you mean to the audio for this text is not good?
“I’ve successfully processed your order and I’d like to confirm your product ID. It is A as in Alpha, 1, 2, 3, B as in Bravo, 5, 6, 7, Z as in Zulu, 8, 9, 0, X as in X-ray.“
I thought this was included in the demo, it seemed okay!
The phone numbers were not naturally read at all. A human would have read a grouping of 123-456-789 like "123", "456", "789", but instead the model generated something like "123", "45", "6789". Listen to the RVSP example again and you'll know what I mean. The pacing is generally off for normal text too, but extra noticeable for the numbers.
My hunch would be that it's because of tokenization, but I wouldn't be able to say that's the issue for sure. Sounds like it though :)
I use OpenAI's voice models a lot and I have access to them all and I'm honestly more impressed with the ease at which one can conduct a conversation with this voice model.
Honestly, this feels like the first voice model I would pilot as a customer service rep in a hospitality setting.
Is there a way to train this on common AI voices from video games/movies, I'd very much like a voice assistant to sound like Father/Mother from Alien or Dead Space.
Very good voice cloning capability. Runs under 10G vram nvidia gpu.
reply