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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
46 comments
[ 312 ms ] story [ 1700 ms ] threadMost that claim to do a British accent end up sounding like Kelsey Grammer - sort of an American accent pretending to be British.
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.
https://techcommunity.microsoft.com/blog/azure-ai-foundry-bl...
So that's a useful next step: for multi-voice TTS models, make them sound like they're in the same room.
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.
I would love to have a model that can make sense of things like stressing particular syllables or phonemes to make a point.
I'm actually more interested in STT (ASR) but the choices there are rather limited.
In fact, I think this kind of thing is absolutely necessary if you want to use this to replace a voice actor.
It seems that it's only variants of English, Spanish and Chinese which are somewhat working.
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.
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.