I read the comments praising these voices as very life like, and went to the page primed to hear very convincing voices. That is not at all what I heard though.
The voices are decent, but the intonation is off on almost every phrase, and there is a very clear robotic-sounding modulation. It's generally very impressive compared to many text-to-speech solutions from a few years ago, but for today, I find it very uninspiring. The AI generated voice you hear all over YouTube shorts is at least as good as most of the samples on this page.
The only part that seemed impressive to me was the English + (Mandarin?) Chinese sample, that one seemed to switch very seamlessly between the two. But this may well be simply because (1) I'm not familiar with any Chinese language, so I couldn't really judge the pronunciation of that, and (2) the different character systems make it extremely clear that the model needs to switch between different languages. Peut-être que cela n'aurait pas été si simple if it had been switching between two languages using the same writing system - I'm particularly curious how it would have read "simple" in the phrase above (I think it should be read with the French pronunication, for example).
And, of course, the singing part is painfully bad, I am very curious why they even included it.
Their comments about the singing and background music are odd. It’s been a while since I’ve done academic research, but something about those comments gave me a strong “we couldn’t figure out how to make background music go away in time for our paper submission, so we’re calling it a feature” vibe as opposed to a “we genuinely like this and think its a differentiator” vibe.
> In fact, we intentionally decided not to denoise our training data because we think it's an interesting feature for BGM to show up at just the right moment. You can think of it as a little easter egg we left for you.
Is there any better model you can point at? I would be interested in having a listen.
There are people – and it does not matter what it's about – that will overstate the progress made (and others will understate it, case in point). Neither should put a damper on progress. This is the best I personally have heard so far, but I certainly might have missed something.
It’s tough to name the best local TTS since they all seem to trade off on quality and features and none of them are as good as ElevenLabs’ closed-source offering.
However Kokoro-82M is an absolute triumph in the small model space. It curbstomps models 10-20x its size in terms of quality while also being runnable on like, a Raspberry Pi. It’s the kind of thing I’m surprised even exists. Its downside is that it isn’t super expressive, but the af_heart voice is extremely clean, and Kokoro is way more reliable than other TTS models: It doesn’t have the common failure mode where you occasionally have a couple extra syllables thrown in because you picked a bad seed.
If you want something that can do convincing voice acting, either pay for ElevenLabs or keep waiting. If you’re trying to build a local AI assistant, Kokoro is perfect, just use that and check the space again in like 6 months to see if something’s beaten it. https://huggingface.co/hexgrad/Kokoro-82M
I recently implemented Fish for a project and found it adequate for TTS but wildly impressive in voice cloning. My POC originally required 3-10 audio samples but I removed the minimum because it could usually one shot it.
The model is good, but I will say their inference code leaves a lot to be desired. I had to rewrite large portions of it for simple things like correct chunking and streaming. The advertised expressive keywords are very much hit and miss, and the devs have gone dark unfortunately.
There's a certain know-nothing feeling I get that makes me worried if we start at the link (which has data showing it > ElevenLabs quality), jump to eh it's actually worse than anything I've heard then last 2 years, and end up at "none are as good as ElevenLabs" - the recommendation and commentary on it, of course, has nothing to do with my feeling, cheers
One of the things this model is actually quite good at is voice cloning. Drop a recorded sample of your voice into the voices folder, and it just works.
I agree. For some reason the female voices are waaay more convincing than the male ones too, which sound barely better than speech synthesis from a decade ago.
Results correlate to investment, and there’s more in synthesizing female coded voices. As for the why female coded voices gets more investments, we all know, only difference is in attitude towards that (the correct answer, of course, is “it sucks”)
There's a lot of money and effort spent in satisfying the sexual desires of (predominantly straight) men. There's not typically quite as much interest in doing the same for women.
For example I've been looking at models and loras for generating images, and the boards are _full_ of ones that will generate women well or in some particular style. Quite often at least a couple of the preview images for each are hidden behind a button because they contain nudity. Clearly the intent is that they are at least able to generate porn containing women. There's a small handful that are focused on men and they're very aware of it, they all have notes lampshading how oddball they are to even exist.
I would expect that this is not as pronounced an effect in the world generating speech, but it must still exist.
I think this is a very lazy kind of cultural analysis. The reason female voices are being chosen over male ones is a little more multifaceted than just SEX. Heterosexual women also tend to prefer female voices over male ones.
Female voices are often rated as being clearer, easier to understand, "warmer", etc.
Why this is the case is still an open question, but it's definitely more complex than just SEX.
That you consider it sex (rather than gender), is exactly why there’s a preference for female coded voices. Consider where we do hear male recorded voices used as default.
The English/Mandarin section was VERY impressive. The accents of both the woman speaking English and the man speaking Chinese were spot on. Both sound very convincingly like they are speaking a second language, which anyone here can hear from the Chinese woman speaking English voice. I'd like to add that the foreigner speaking Chinese was also spot on.
The male Chinese speakers had THICK American accents. Nothing really wrong with the language, but think the stereotype German speaking English. That was kind of strange to me.
I think it's because it was using the American voice for it. Conversely the female voice in the Mandarin conversation spoke English with a Chinese accent.
The Chinese is good. The Mandarin to English example she sounds native. The English to Mandarin sounds good too but he does have an English speaker's accent, which I think is intentional.
This is close to SOTA emotional performance, at least the female voices.
I trust the human scores in the paper. At least my ear aligns with that figure.
With stuff like this coming out in the open, I wonder if ElevenLabs will maintain its huge ARR lead in the field. I really don't see how they can continue to maintain a lead when their offering is getting trounced by open models.
I really hope someone within Microsoft is naming their open source coding agent Microsoft VibeCode. Let this be a thing. Its either that or "Lo" then you can have Lo work with Phi, so you can Vibe code with Lo Phi.
Knowing the history of Microsoft marketing, it will either be called something like "Microsoft Copilot Code Generator for VSCode" or something like "Zunega"...
Any insight on my the code and the large model were removed? Some copies are floating around and are MIT licensed. In cases like this I do not know why the projects are yanked. If the project was mistakenly released under MIT, copied elsewhere, is any damage control possible by yanking the copies you have control over? Mostly seems like bad PR, if minor.
This is clearly high quality but there's something about the voices, the male voices in particular, which immediately register as computer generated. My audio vocabulary is not rich enough to articulate what it is.
I'm no audio engineer either, but those computer voice sound "saw-tooth"y to me.
From what I understand, it's more basic models/techniques that are undersampling, so there is a series of audio pulses which give it that buzzy quality. Better models are produced smoother output.
I would describe it as blockly, as if we visualise the sound wave it seems to be without peaks and cut upwards and downwards producing a metallic boxy echo.
The male voices seem much worse than the female voices, borderline robotic. Every sample of their website starts with a female voice. They clearly are aware of the issue.
Generally if a model is trending on that page, there’s enough juice for it to be worth a try. There’s a lot of subjective-opinion-having in this space, so beyond “is it trending on HF” the best eval is your own ears. But if something is not trending on HF it is unlikely to be much good.
Unfortunately it's not usable if you're GPU-poor. Couldn't figure out how to run this with an old 1080. I tried VibeVoice-1.5B on my old CPU with torch.float32 and it took 832 seconds to generate a 66 second audio clip. Switching from torch.bfloat16 also introduced some weird sound artifacts in the audio output. If you're GPU-poor the best TTS model I've tried so far is Kokoro.
Someone else mentioned in this thread that you cannot add annotations to the text to control the output. I think for these models to really level up there will have to be an intermediate step that takes your regular text as input and it generates an annotated output, which can be passed to the TTS model. That would give users way more control over the final output, since they would be able to inspect and tweak any details instead of expecting the model to get everything correctly in a single pass.
This is ludicrous. macOS has had text-to-speech for ages with acceptable quality, and they never needed energy- and compute-expensive models for it. And it reacts instantly, not after ridiculous delays. I cannot believe this hype about “AI”, it’s just too absurd.
I tried listening to audiobooks generated with tts. It takes me out of it most of the time, and I lose focus. That podcast thing from google was the first time I felt like I could listen to an entire thing without feeling the uncanny valley thing. And I knew it was genAI. So I'm looking for that, but for my content. Grab a bunch of articles (long form, deeply researched) and "podcast" them but with natural voices, sans hype. Or books. Have them ready when I'm out and about.
Audiobooks and other material you want to listen to (articles, blog posts, etc.).
There's a lot of stuff I don't have time to sit down and read, but want to listen to while I cook/laundry/shower/drive/etc.
Often recordings don't exist. Or when they do, an audiobook just has a bad voiceover artist, or one that just rubs you the wrong way.
The more human text-to-speech sounds, the easier and less distracting it is to listen to. There's real value in it, it's not "because I can".
You know how it's nicer to read in 300 dpi instead of 72 dpi? Or in Garamond rather than Courier? Or in Helvetica rather than Comic Sans? It's like that, only for speech.
> Question is why would you need to have the computer sound more human
I think translation would be a big use - maybe translating your voice to another language while maintaining emotion and intonation, or dubbing content (videos, movies, podcasts, ...) that isn't otherwise available in your native language.
Traditional non-ML TTS for longer content like podcasts or audiobooks seems like it'd become grating to the point of being unlistenable, or at least a significantly worse experience. Stands to benefit from more natural sounding voices that can place emphasis in the right places.
Since Stephen Hawking was brought up, there are likely also people with voice-impairing illnesses who would like to speak in their own voice again (in addition to those who are fine with a robotic voice). Or alternatively, people who are uncomfortable with their natural voice and want to communicate closer to how they wish to be perceived.
Could also potentially be used for new forms of interactive media that aren't currently feasible - customised movies, audio dramas where the listener plays a role, videogame NPCs that react with more than just prerecorded lines, etc.
What an odd name to me, becaus "Vibe" is, in my mind, equal to somewhat poor quality. Like "Vibe Coding". But that's probably just some bias from my side.
Vibe coding just became a term this spring.
I doubt that that the substantial part, like giving it a project code name and getting company approval of this research project started after that. It's not libe vibe has a negative connotation in general yet.
Vibe always meant "specific feel" and makes sense related to AI coding "by touch" vs. understanding what's actually happening. It's just the results have now made the word pejorative.
A lot of code has multiple FOSS licenses that are not contaminating like GPL. GPL violations do occur on code, but have nothing to do with the training Data.
For example, many academic data sets are not public domain, and can't be used in a commercial context. A GPL claim on that data is often an argument of which thief showed up first.
Rule #24: A lawyers Strategic Truth is to never lie, but also avoid voluntarily disclosing information that may help opponents.
Thus, a business will never disclose they paid a fool to break laws for them... =3
VibeVoice-Large is the first local TTS that can produce convincing Finnish speech with little to no accent. I tinkered with it yesterday and was pleasantly surprised at how good the voice cloning is and how it "clones" the emotion in the speech as well.
There are 2 "best" TTS models out right now: HiggsAudio and VibeVoice. I found that Higgs is both faster and much higher fidelity than Vibe. Can't speak to expressiveness, but don't sleep on it.
Here is AI being as close as possible to the most animated person I know and here I am sounding robotic in every conversation I have, despite my best efforts to sound otherwise. Sometimes, I just wish I could have an AI speak for me
Very good and I could see how I might believe they are real people if I let my guard down. The male voice sounded a little sedated though and there was a smoothness to it that could be samey over long stretches.
Still not at the astonishing level of Google Notebook text to speech which has been out for a while now. I still can't believe how good that one is.
Ok, this is nit-picking, but it's very obvious that the sample voices these were trained with were captured in different audio environments. There's noticeable reverb on the male voice that's not there on the other.
So that's a useful next step: for multi-voice TTS models, make them sound like they're in the same room.
To me this is like early generative AI art, where the images came out very "smooth" and visually buttery, but instead there's no timbre to the voices. Intonation issues aside, these models could use a touch of vocal fry and some body to be more believable
I thought the name sounded familiar, I'm guessing its no relation to this project which has been around for 7 months? https://github.com/mpaepper/vibevoice
I tried the colab notebook that they link to and couldn't replicate the quality for whatever reason. I just swapped out the text and let it run on the introduction paragraph of Metamorphosis by Franz Kafka and it seemingly could not handle the intricacies.
I’m just a yank, but a lot of the AI-voiced videos on YouTube that I’ve been listening to while I’m falling asleep lately have British voices that sound quite nice to me.
Fair enough... though it would be possible to generate that and edit to overlay the speech, introducing stuttering/pauses at the beginning and end of statements then edit the output to overlay the steps.
Would probably want to do similar to balance crossfade anyway... having each speaker's input offset from center instead of straight mono.
Yeah, a lot of the TTS has gotten really impressive in general. Definitely a clear leap from the TTS stuff I worked with for training simulations a bit over a decade ago. Aside: Installing a sound card (unused) on a windows server just to be able to generate TTS was interesting. It was required by the platform, even if it wasn't used for it.
I generally don't like a lot of the AI generated slop that's starting to pop up on YouTube these days... I do enjoy some of the reddit story channels, but have completely stopped with it all now. With the AI stuff, it really becomes apparent with dates/ages and when numbers are spoken. Dates/ages/timelines are just off as far as story generation, and really should be human tweaked. As to the voice gen, saying a year or measurement is just not how English speakers (US or otherwise) speak.
does anyone know of recent TTS options that let you specify IPA rather than written words? Azure lets you do this, but something local (and better than existing OS voices) would be great for my project.
I'm using Kokoro via https://github.com/remsky/Kokoro-FastAPI. It has a `generate_audio_from_phonemes()` endpoint that I'm sure maps to the Kokoro library if you want to use it directly.
My usage is for Chinese, but the phonemes it generated looked very much like IPA.
The most popular song of the year from one of the most popular movie franchises that had been in the global news due to the death of its star. Probably the most memorable song from a soundtrack of the century so far.
I'm Just Ken (Barbie), Skyfall, Let it Go (Frozen), Remember Me (Coco), Happy (from Despicable Me 2), a Star is Born (Shallow), are all arguably wayyyyy more memorable and these are just off the top of my head. We've had quite a few memorable songs in soundtracks this millennium.
edit: I had forgotten about Jai Ho (Slumdog Millionaire) and Lose Yourself (8 mile)
Have you tried Soniox for speech recognition? It supports Croatian. Or are you just looking for self-hosted open-source models? Soniox is very cheap ($0.1/h for async, $0.12/h for real-time) and you get $200 free credits on signup.
I meant in general purpose tools from Google and Apple. Most of this assistant and "AI" stuff is practically useless for me because I refuse to talk to my devices in English.
In Android Auto / CarPlay I can't even get voice guidance that works properly, much less reading notifications, or composing a reply using STT
It is an unfortunate recycling of an existing regime that no doubt offends Stallman to his very core, but I wouldn't call it meaningless.
If you're in a company and need a model which one do you think you're getting past compliance & legal - the one that says MIT or the one that says "non-commercial use only"?
what does that mean in this context? it seems to depend on an LLM. so can i run this completely offline? if i have to sign up and pay for an LLM to make it work, then it's not really more useful than any other non-free system
I tried some TTS models a while ago, but I noticed that none of them allowed to put markup statements in the text. For example, it would be nice to do something like:
Hey look! [enthusiastic] Should we tell the others? Maybe not ... [giggles]
etc.
In fact, I think this kind of thing is absolutely necessary if you want to use this to replace a voice actor.
I feel like this is a step in the right direction, but a lot of emotive text-to-speech models are only changing the duration and loudness of each word, the timing/pauses are better too.
I would love to have a model that can make sense of things like stressing particular syllables or phonemes to make a point.
The voices are decent, but the intonation is off on almost every phrase, and there is a very clear robotic-sounding modulation. It's generally very impressive compared to many text-to-speech solutions from a few years ago, but for today, I find it very uninspiring. The AI generated voice you hear all over YouTube shorts is at least as good as most of the samples on this page.
The only part that seemed impressive to me was the English + (Mandarin?) Chinese sample, that one seemed to switch very seamlessly between the two. But this may well be simply because (1) I'm not familiar with any Chinese language, so I couldn't really judge the pronunciation of that, and (2) the different character systems make it extremely clear that the model needs to switch between different languages. Peut-être que cela n'aurait pas été si simple if it had been switching between two languages using the same writing system - I'm particularly curious how it would have read "simple" in the phrase above (I think it should be read with the French pronunication, for example).
And, of course, the singing part is painfully bad, I am very curious why they even included it.
> In fact, we intentionally decided not to denoise our training data because we think it's an interesting feature for BGM to show up at just the right moment. You can think of it as a little easter egg we left for you.
It's not a bug, it's a feature! Okaaaaay
There are people – and it does not matter what it's about – that will overstate the progress made (and others will understate it, case in point). Neither should put a damper on progress. This is the best I personally have heard so far, but I certainly might have missed something.
However Kokoro-82M is an absolute triumph in the small model space. It curbstomps models 10-20x its size in terms of quality while also being runnable on like, a Raspberry Pi. It’s the kind of thing I’m surprised even exists. Its downside is that it isn’t super expressive, but the af_heart voice is extremely clean, and Kokoro is way more reliable than other TTS models: It doesn’t have the common failure mode where you occasionally have a couple extra syllables thrown in because you picked a bad seed.
If you want something that can do convincing voice acting, either pay for ElevenLabs or keep waiting. If you’re trying to build a local AI assistant, Kokoro is perfect, just use that and check the space again in like 6 months to see if something’s beaten it. https://huggingface.co/hexgrad/Kokoro-82M
The model is good, but I will say their inference code leaves a lot to be desired. I had to rewrite large portions of it for simple things like correct chunking and streaming. The advertised expressive keywords are very much hit and miss, and the devs have gone dark unfortunately.
- http://dia-tts.com/
- https://github.com/canopyai/Orpheus-TTS
https://github.com/mlang/llm-tts
Strictly speaking, even music generation fits the usage pattern: text in, audio out.
llm-tts is far from complete, but it makes it relatively "easy" to try a few models in an uniform way.
For example I've been looking at models and loras for generating images, and the boards are _full_ of ones that will generate women well or in some particular style. Quite often at least a couple of the preview images for each are hidden behind a button because they contain nudity. Clearly the intent is that they are at least able to generate porn containing women. There's a small handful that are focused on men and they're very aware of it, they all have notes lampshading how oddball they are to even exist.
I would expect that this is not as pronounced an effect in the world generating speech, but it must still exist.
Female voices are often rated as being clearer, easier to understand, "warmer", etc.
Why this is the case is still an open question, but it's definitely more complex than just SEX.
Women also prefer female voices.
> satisfying the sexual desires of
So, "sex" as a reference to "sexual desires". In English, it just so happens that "sex" has other meanings, but those weren't in play at the time.
I trust the human scores in the paper. At least my ear aligns with that figure.
With stuff like this coming out in the open, I wonder if ElevenLabs will maintain its huge ARR lead in the field. I really don't see how they can continue to maintain a lead when their offering is getting trounced by open models.
https://yummy-fir-7a4.notion.site/dia
I am not sure why but I find the pacing of the parakeet based models (like Dia) to be much more realistic.
https://en.wikipedia.org/wiki/Gell-Mann_amnesia_effect
https://techcommunity.microsoft.com/blog/azure-ai-foundry-bl...
From what I understand, it's more basic models/techniques that are undersampling, so there is a series of audio pulses which give it that buzzy quality. Better models are produced smoother output.
https://www.perfectcircuit.com/signal/difference-between-wav...
I'm actually more interested in STT (ASR) but the choices there are rather limited.
Generally if a model is trending on that page, there’s enough juice for it to be worth a try. There’s a lot of subjective-opinion-having in this space, so beyond “is it trending on HF” the best eval is your own ears. But if something is not trending on HF it is unlikely to be much good.
edit: Ah, there's a lock icon next to the name of each proprietary model.
Someone else mentioned in this thread that you cannot add annotations to the text to control the output. I think for these models to really level up there will have to be an intermediate step that takes your regular text as input and it generates an annotated output, which can be passed to the TTS model. That would give users way more control over the final output, since they would be able to inspect and tweak any details instead of expecting the model to get everything correctly in a single pass.
Compared to IBMs Steven Hawking's chair, maybe. But apple tts is not acceptable quality in any modern understanding of SotA, IMO.
If you need a not-visual output of text, SoyA is a waste of electrons.
If you want to try and mimic a human speaker, then it ain’t.
Question is why would you need to have the computer sound more human, except for “because I can”.
We can't be friends
There's a lot of stuff I don't have time to sit down and read, but want to listen to while I cook/laundry/shower/drive/etc.
Often recordings don't exist. Or when they do, an audiobook just has a bad voiceover artist, or one that just rubs you the wrong way.
The more human text-to-speech sounds, the easier and less distracting it is to listen to. There's real value in it, it's not "because I can".
You know how it's nicer to read in 300 dpi instead of 72 dpi? Or in Garamond rather than Courier? Or in Helvetica rather than Comic Sans? It's like that, only for speech.
I think translation would be a big use - maybe translating your voice to another language while maintaining emotion and intonation, or dubbing content (videos, movies, podcasts, ...) that isn't otherwise available in your native language.
Traditional non-ML TTS for longer content like podcasts or audiobooks seems like it'd become grating to the point of being unlistenable, or at least a significantly worse experience. Stands to benefit from more natural sounding voices that can place emphasis in the right places.
Since Stephen Hawking was brought up, there are likely also people with voice-impairing illnesses who would like to speak in their own voice again (in addition to those who are fine with a robotic voice). Or alternatively, people who are uncomfortable with their natural voice and want to communicate closer to how they wish to be perceived.
Could also potentially be used for new forms of interactive media that aren't currently feasible - customised movies, audio dramas where the listener plays a role, videogame NPCs that react with more than just prerecorded lines, etc.
But I do agree with you in that generally there's probably no negative connotation (yet).
Making it "open" would be unwise for a commercial entity. =3
For example, many academic data sets are not public domain, and can't be used in a commercial context. A GPL claim on that data is often an argument of which thief showed up first.
Rule #24: A lawyers Strategic Truth is to never lie, but also avoid voluntarily disclosing information that may help opponents.
Thus, a business will never disclose they paid a fool to break laws for them... =3
Indeed, these adversarial behaviors do not follow the spirit of FOSS community standards. If a project started as FOSS, than FOSS it should remain. =3
They could have skipped the singing part, it would be better if the model did not try to do that :)
1. https://music.youtube.com/watch?v=xl8thVrlvjI&si=dU6aIJIPWSs...
https://github.com/microsoft/VibeVoice
I was trying to get this working on strix halo.
https://github.com/mpaepper/vibevoice
Still not at the astonishing level of Google Notebook text to speech which has been out for a while now. I still can't believe how good that one is.
So that's a useful next step: for multi-voice TTS models, make them sound like they're in the same room.
Most that claim to do a British accent end up sounding like Kelsey Grammer - sort of an American accent pretending to be British.
A 100M podcast model
https://huggingface.co/spaces/fluxions/vui-space
Would probably want to do similar to balance crossfade anyway... having each speaker's input offset from center instead of straight mono.
I generally don't like a lot of the AI generated slop that's starting to pop up on YouTube these days... I do enjoy some of the reddit story channels, but have completely stopped with it all now. With the AI stuff, it really becomes apparent with dates/ages and when numbers are spoken. Dates/ages/timelines are just off as far as story generation, and really should be human tweaked. As to the voice gen, saying a year or measurement is just not how English speakers (US or otherwise) speak.
My usage is for Chinese, but the phonemes it generated looked very much like IPA.
https://github.com/WhisperSpeech/WhisperSpeech
Or is there some OpenAI official Whisper TTS?
edit: I had forgotten about Jai Ho (Slumdog Millionaire) and Lose Yourself (8 mile)
Nothing on that list - movies or songs - had the cultural impact of Furious 7 or See You Again.
It seems that it's only variants of English, Spanish and Chinese which are somewhat working.
https://soniox.com/
Disclaimer: I used to work for Soniox
In Android Auto / CarPlay I can't even get voice guidance that works properly, much less reading notifications, or composing a reply using STT
If you're in a company and need a model which one do you think you're getting past compliance & legal - the one that says MIT or the one that says "non-commercial use only"?
In fact, I think this kind of thing is absolutely necessary if you want to use this to replace a voice actor.
https://elevenlabs.io/blog/v3-audiotags
Some of them have tone wobbles which iirc was more common in early TTS models. Looks like the huge context window is really helping out here.
I would love to have a model that can make sense of things like stressing particular syllables or phonemes to make a point.