I'm psyched to see so much interest in my post about Kyutai's latest model! I'm working on part of a related team in Paris that's building off Kutai's research to provide enterprise-grade voice solutions. If anyone building in this space I'd love to chat and share some our upcoming models and capabilities that I am told are SOTA. Please don't hesitate to ping me via the address in my profile.
Woah, I'm impressed! The voice cloning also worked much better than expected! Will there be separate models for other languages? I know the National Library in Norway has done a good job curating speech datasets with many different dialects [1][2].
You can do the same thing with Firefox' Reader Mode. On Linux you have to set up speech-dispatcher to use your favorite TTS as a backend.Once it is set up, there will be an option to listen the page.
Firefox should integrate that in their Reader Mode (the default System Voices are often very un-listable). Would seems like an easy win, and it's a non-AI feature so not polarising.
Not sure about macOS or Windows, but on Linux Firefox uses speech-dispatcher, which is a server, and Firefox is the client. Speech-dispatcher then delegates the text to the correct TTS backend. It basically runs a shell command, either sending the text to a TTS HTTP server using curl, or piping it to the standard input of a TTS binary.
Speech-dispatcher commonly uses espeak-ng, which sounds robotic but is reportedly better for visually impaired users, because at higher speeds it is still intelligible. This allows visually impaired users to hear UI labels more quickly. For non visually impaired users, we generally want natural sounding voices and to use TTS in the same way we would listen to podcasts or a bedtime story.
With this system, users are in full control and can swap TTS models easily. If a model is shipped and, two weeks later, a smaller, newer, or better one appears, their work would become obsolete very quickly.
macOS already has some great intrinsic TTS capability as the OS seems to include a naturally sounding voice. I recently built a similar tool to just run the "say" command as a background process. Had to wrap it in a Deno server. It works, but with Tahoe it's difficult to consistently configure using that one natural voice, and not the subpar voices downloadable in the settings. The good voice seems to be hidden somehow.
My mistake, seems like I was refering to the Siri voice, which seems to be the default. It sounds good. It is selectable and to my surprise - even configurable in speed, pitch and volume - in the OS Accessibility settings -> System Voice -> Click on the (i) symbol. (macOS Tahoe)
Oh this is sweet, thanks for sharing! I've been a huge fan of Kokoro and event setup my own fully-local voice assistant [1]. Will definitely give Pocket TTS a go!
It seems like Kokoro is the smaller model, also runs on CPU in real time, and is more open and fine tunable. More scripts and extensions, etc., whereas this is new and doesn't have any fine tuning code yet.
Kokoro is fine tunable? Speaking as someone who went down the rabbit hole... it's really not. There's no (as of last time I checked) training code available so you need to reverse engineer everything. Beyond that the model is not good at doing voices outside the existing voicepacks: simply put, it isn't a foundation model trained on internet scale data. It is made from a relatively small set of focused, synthetic voice data. So, a very narrow distribution to work with. Going OOD immediately tanks perceptual quality.
There's a bunch of inference stuff though, which is cool I guess. And it really is a quite nice little model in its niche. But let's not pretend there aren't huge tradeoffs in the design: synthetic data, phonemization, lack of train code, sharp boundary effects, etc.
Zero shot voice clones have never been very good. Fine tuned models hit natural speaker similarity and prosody in a way zero shot models can't emulate.
If it were a big model and was trained on a diverse set of speakers and could remember how to replicate them all, then zero shot is a potentially bigger deal. But this is a tiny model.
I'll try out the zero shot functionality of Pocket TTS and report back.
Less licensing headache, it seems. Kokoro says its Apache licensed. But it has eSpeak-NG as a dependency, which is GPL, which brings into question whether or not Kokoro is actually GPL. PocketTTS doesn't have eSpeak-NG as a dependency so you don't need to worry about all that BS.
Btw, I would love to hear from someone (who knows what they're talking about) to clear this up for me. Dealing with potential GPL contamination is a nightmare.
If it depends on espeak NG code, the complete product is 100% GPL. That said, if you are able to change the code to take off the espeak dependency then the rest would revert to non-GPL (or even if it's a build time option that you can disable like FFMPEG with --enable-gpl)
Kokoro only uses Espeak for text-to-phoneme (AKA G2P) conversion.
If you could find another compatible converter, you could probably replace eSpeak with it. The data could be a bit OOD, so you may need to fiddle with it, but it should work.
Because the GPL is outdated and doesn't really consider modern gen AI, what you could also do is to generate a bunch of text-to-phoneme pairs with Espeak and train your own transformer on them,. This would free you from the GPL license completely, and the task is easy enough that even a very small model should be able to do it.
So, on my M1 mac, did `uvx pocket-tts serve`. Plugged in
> It was the best of times, it was the worst of times, it was the age of
wisdom, it was the age of foolishness, it was the epoch of belief, it was the epoch of incredulity, it was the season of Light, it was the season of Darkness, it was the spring of hope, it was the winter of despair, we had everything before us, we had nothing before us, we were all going direct to Heaven, we were all going direct the other way—in short, the period was so far like the present period, that some of its noisiest authorities insisted on its being received, for good or for evil, in the superlative degree of comparison only
(Beginning of Tale of Two Cities)
but the problem is Javert skips over parts of sentences! Eg, it starts:
> "It was the best of times, it was the worst of times, it was the age of
wisdom, it was the epoch of belief, it was the epoch of incredulity, it was the season of Light, it was the spring of hope, it was the winter of despair, we had everything before us, ..."
Notice how it skips over "it was the age of foolishness,", "it was the winter of despair,"
Which... Doesn't exactly inspire faith in a TTS system.
All the models I tried have similar problems. When trying to batch a whole audiobook, the only way is to run it, then run a model to transcribe and check you get the same text.
Václav from Kyutai here. Thanks for the bug report! A workaround for now is to chunk the text into smaller parts where the model is more reliable. We already do some chunking in the Python package. There is also a more fancy way to do this chunking in a way that ensures that the stitched-together parts continue well (teacher-forcing), but we haven't implemented that yet.
It says MIT license but then readme has a separate section on prohibited use that maybe adds restrictions to make it nonfree? Not sure the legal implications here.
For reference, the MIT license contains this text: "Permission is hereby granted... to deal in the Software without restriction, including without limitation the rights to use". So the README containing a "Prohibited Use" section definitely creates a conflicting statement.
I think the only restriction that seems problematic is not being able to clone someone’s voice without permission. I think there’s probably a valid case for using it for satire.
You might use it for something illegal in one country, and then leave for another country with no extradition… but you’ve lost the license to sue the software and can be sued for copyright infringement.
From my understanding, the code is MIT, but the model isn't? What consitutes a "Software" anyway? Aren't resources like images, sounds and the likes exempt from it (hence, covered by usual copyright unless separately licensed)? If so, in the same vein, an ML model is not part of "Software". By the way, the same prohibition is repeated on the huggingface model card.
If memory serves, the license is the ultimate source of truth on what is allowed or not. You cannot add some section that isn't in the text of the license (at least in the US and other countries that use similar legal systems) on some website and expect it to hold up in court because the license doesn't include that text. I know of a few other bigger-name projects that try to pull these kinds of stunts because they don't believe anyone is going to actually read the text of the license.
The copyright holder can set whatever license they want, including writing their own.
In this case, I'd interpret it as they made up a new licence based on MIT, but their addendum makes it non-MIT, but something else. I agree with what others said; this "new" license has internal conflicts.
I think if they took you to court for cloning someone's voice without permission they would probably lose because this conflict makes the terms unclear.
Perhaps I have been not talking to voice models that much or the chatgpt voice always felt weird and off because I was thinking it goes to a cloud server and everything but from Pocket TTS I discovered unmute.sh which is open source and I think is from the same company as Pocket TTS/can I think use Pocket TTS as well
I saw some agentic models at 4B or similar which can punch above its weights or even some basic models. I can definitely see them in the context of home lab without costing too much money.
I think atleast unmute.sh is similar/competed with chatgpt's voice model. It's crazy how good and (effective) open source models are from top to bottom. There's basically just about anything for almost everyone.
I feel like the only true moat might exist in coding models. Some are pretty good but its the only industry where people might pay 10x-20x more for the best (minimax/z.ai subscription fees vs claude code)
It will be interesting to see if we will see another deepseek moment in AI which might beat claude sonnet or similar. I think Deepseek has deepseek 4 so it will be interesting to see how/if it can beat sonnet
Yes, apart from voice cloning nothing really new. Kokoro is out since a long time and it supports at least a few languages other than english. Also there is supertonic TTS and there is Soprano TTS. The latter is developed by a single guy while Kyutai is funded with 150M€.
I echo this. For a TTS system to be in any way useful outside the tiny population of the world that speaks exclusively English, it must be multilingual and dynamically switch between languages pretty much per word.
That's a pretty crazy requirement for something to be "useful" especially something that runs so efficiently on cpu. Many content creators from non-english speaking countries can benefit from this type of release by translating transcripts of their content to english and then running it through a model like this to dub their videos in a language that can reach many more people.
Uh, no? This is not at all an absurd requirement? Screen readers literally do this all the time, with voices that are the classic way of making a speech synthesizer, no AI required. ESpeak is an example, or MS OneCore. The NVDA screen reader has an option for automatic language switching as does pretty much every other modern screen reader in existence. And absolutely none of these use AI models to do that switching, either.
That doesn't really change what I said though. It isn't crazy to call it useless without some form of ALS either. Given that old school synthesis has been able to do it for like 20 years or so.
You mean youtubers? And have to (manually) synchronise the text to their video, and especially when youtube apparently offers voice-voice translation out of the box to my and many others' annoyance?
YouTube's voice to voice is absolutely horrible though. Having the ability for the youtubers to clone their own voice would make it much, much more appealing.
But it wouldn't be for those who "speak exclusively English", rather, for those who speak English. Not only that but it's also common to have system language set to English, even if one's language is different.
There's about 1.5B English speakers in the planet.
Let's indeed limit the use case to the system language, let's say of a mobile phone.
You pull up a map and start navigation. All the street names are in the local language, and no, transliterating the local names to the English alphabet does not make them understandable when spoken by TTS. And not to mention localised foreign names which then are completely mangled by transliterating them to English.
You pull up a browser, open up an news article in your local language to read during your commute. You now have to reach for a translation model first before passing the data to the English-only TTS software.
You're driving, one of your friends Signals you. Your phone UI is in English, you get a notification (interrupting your Spotify) saying 'Signal message', followed by 5 minutes of gibberish.
But let's say you have a TTS model that supports your local language natively. Well due to the fact that '1.5B English speakers' apparently exist in the planet, many texts in other languages include English or Latin names and words. Now you have the opposite issue -- your TTS software needs to switch to English to pronounce these correctly...
And mind you, these are just very simple use cases for TTS. If you delve into use cases for people with limited sight that experience the entire Internet, and all mobile and desktop applications (often having poor localisation) via TTS you see how mono-lingual TTS is mostly useless and would be switched for a robotic old-school TTS in a flash...
> only that but it's also common to have system language set to English
Ask a German whether their system language is English. Ask a French person. I can go on.
If you don't speak the local language anyway, you can't decode pronounced spoken local language names anyway. Your speech sub-systems can't lock and sync to the audio track containing languages you don't speak. Let alone transliterate or pronounce.
Multilingual doesn't mean language agnostic. We humans are always monolingual, just multi-language hot-swappable if trained. It's more like you can make;make install docker, after which you can attach/detach into/out of alternate environments while on terminal to do things or take in/out notes.
People sometimes picture multilingualism as owning a single joined-together super-language in the brain. That usually doesn't happen. Attempting this especially at young age could lead a person into a "semi-lingual" or "double-limited" state where they are not so fluent or intelligent in any particular languages.
And so, trying to make an omnilingual TTS for criticizing someone not devoting significant resources at it, don't make much sense.
> If you don't speak the local language anyway, you can't decode pronounced spoken local language names anyway
This is plainly not true.
> Multilingual doesn't mean language agnostic. We humans are always monolingual, just multi-language hot-swappable if trained
This and the analogy make no sense to me. Mind you I am trilingual.
I also did not imply that the model itself needs to be multilingual. I implied that the software that uses the model to generate speech must be multilingual and support language change detection and switching mid-sentence.
> it must be multilingual and dynamically switch between languages pretty much per word
Not abundantly obviously a satire and so interjecting: humans, including professional "simultaneous" interpreters, can't do this. This is not how languages work.
But that's my point. You'll stop, switch, speak, stop, switch, resume. You're not going to be "I was in 東京 yesterday" as a single continuous sentence. It'll have to be broken up to three separate sentences spoken back to back, even for humans.
I think it's the wrong example, because this is actually very common if you're a Chinese speaker.
Actually, people tend to say the name of the cities in their own countries in their native language.
> I went to Nantes [0], to eat some kouign-amann [1].
As a French, both [0] and [1] will be spoken the French way on the fly in the sentence, while the other words are in English. Switching happens without any pause whatsoever (because there is really only one single way to pronounce those names in my mind, no thinking required).
Note that with Speech Recognition, it is fairly common to have models understanding language switches within a sentence like with Parakeet.
I think this is totally wrong. When you have both parties speaking multiple languages this happens all the time. You see this more with English being the loaner more often than it is the borrower, due to the reach that the language has. Listen to an Indian or Filipino speak for a while, it's interspersed with English words ALL the time. It happens less in English as there is not the universal knowledge base of one specific other language, but it does happen sometimes when searching for a certain, je ne sais pas.
Not really, most multilinguals switch between languages so seamlessly that you wouldn't even notice it! It even has given birth to new "languages", take for example Hinglish!!
The speed of improvement of tts models reminds me of early days of Stable Diffusion. Can't wait until I can generate audiobooks without infinite pain. If I was an investor I'd short Audible.
Isn’t it more like an art gallery of prints of paintings? The primary art is the text of the book (like the painting in the gallery), TTS (and printing a copy) are just methods of making the art available.
I think it can be argued that audiobook's add to the art by adding tone and inflection by the reader.
To me, what you're saying is the same as saying the art of a movie is in the script, the video is just the method of making it available. And I don't think that's a valid take
No, that's an incorrect analogy. The script of a movie is an intermediate step in the production process of a movie. It's generally not meant to be seen by any audiences. The script for example doesn't contain any cinematography or any soundtrack or any performances by actors. Meanwhile, a written work is a complete expressive work ready for consumption. It doesn't contain a voice, but that's because the intention is for the reader to interpret the voice into it. A voice actor can do that, but that's just an interpretation of the work. It's not one-to-one, but it's not unlike someone sitting next to you in the theater and telling you what they think a scene means.
So yes, I mostly agree with GP. An audiobook is a different rendering of the same subject. The content is in the text, regardless of whether it's delivered in written or oral form.
It's not perfect, but I already have a setup for doing this on my phone. Add SherpaTTS and Librera Reader to your phone. (both available free on fdroid).
Set up SherpaTTS as the voice model for your phone (I like the en_GB-jenny_dioco-medium voice option, but there are several to choose from). Add a ebook to librera reader and open it. There's an icon with a little person wearing headphones, which lets you send the text continuously to your phone's tts, using just local processing on the phone. I don't have the latest phone but mine is able to process it faster than the audio is read, so the audio doesn't stop and start.
The voice isn't totally human sounding, but it's a lot better than the microsoft sam days, and once you get used to it the roboticness fades into the background and I can just listen to the story. You may get better results with kokoro (I couldn't get it running on my phone) or similar tts engines and a more powerful phone.
One thing I like about this setup is that if you want to swap back and forth between audio and text, you can. The reader scrolls automatically as it makes the audio, and you can pause it, read in silence for a while yourself and later set it going from a new point.
I feel like TTS is one of the areas that as evolved the least. Small TTS models have been around for like 5+ years and they've only gotten incrementally better. Giants like ElevenLabs make good sounding TTS but it's not quite human yet and the improvements get less and less each iteration.
Chatterbox does something like that. For example, if the input is
"so and so," he <verb>
and the verb is not just "said", but "chuckled", or "whispered", or "said shakily", the output is modified accordingly, or if there's an indication that it's a woman speaking it may pitch up during the quotation. It also tries to guess emotive content from textual content, such if a passage reads angry it may try to make it sound angry. That's more hit-and-miss, but when it hits, it hits really well. A very common failure case is, imagine someone is trying to psych themselves up and they say internally "come on, Steve, stand up and keep going", it'll read it in a deeper voice like it was being spoken by a WW2 sergeant to a soldier.
But there seems to be a bug maybe? Just for fun, I had asked it to play the Real Slim Shady lyrics. It always seems to add 1 extra "please stand-up" in the chorus. Anyone see that?
Just added it to my codex plugin that reads summary of what it finishes after each turn and I am spooked! runs well on my macbook, much better than Samantha!
Gabriel from Kyutai here, we do support outputting wav to stdout. We don't support reading text from stdin but that should be easy enough. Feel free to drop a pull request!
This is amazing. The audio feels very natural and it's fairly good at handling complext text to speech tasks.
I've been working on WithAudio (https://with.audio). Currently it only uses Kokoros. I need to test this a bit more but I might actually add it to the app. It's too good to be ignored.
I love that everyone is making their own TTS model as they are not as expensive as many other models to train. Also there are plenty of different architecture.
Václav from Kyutai here. Yes the original naming scheme was from Les Miserables, glad you noticed! We just stuck to Alba because that's the real name of the voice actor that provided the voice sample to us (see https://huggingface.co/kyutai/tts-voices), the other ones are either from pre-existing datasets or given anonymously.
It's very impressive!
I'm mean, it's better than other <200M TTS models I encounter.
In English, it's perfect and it's so funny in others languages. It sounds exactly like someone who actually doesn't speak the language, but got it anyway.
I don't know why Fantine is just better than the others in others languages. Javer seems to be the worst.
Try Jean in Spanish « ¡Es lo suficientemente pequeño como para caber en tu bolsillo! » sound a lot like they don't understand the language.
Or Azelma in French « C'est suffisament petit pour tenir dans ta poche. » is very good.I mean half of the words are from a Québécois accent, half French one but hey, it's correct French.
[3] (2021 https://arxiv.org/abs/2106.07889) UnivNet: A Neural Vocoder with Multi-Resolution Spectrogram Discriminators for High-Fidelity Waveform Generation
>If you want access to the model with voice cloning, go to https://huggingface.co/kyutai/pocket-tts and accept the terms, then make sure you're logged in locally with `uvx hf auth login`
lol
I’ve tried the voice clinking and it works great. I added a 9s clip and it captured the speaker pretty well.
But don’t do the fake mistake I did and use a hf token that doesn’t have access to read from repos! The error message said that I had to request access to the repo, but I’ve had already done that, so I couldn’t figure out what was wrong. Turns out my HF token only had access to inference.
[1] https://data.norge.no/en/datasets/220ef03e-70e1-3465-a4af-ed...
[2] https://ai.nb.no/datasets/
Ok, who knows where I can get those high-quality recordings of Majel Barrett' voice that she made before she died?
https://github.com/lukasmwerner/pocket-reader
Speech-dispatcher commonly uses espeak-ng, which sounds robotic but is reportedly better for visually impaired users, because at higher speeds it is still intelligible. This allows visually impaired users to hear UI labels more quickly. For non visually impaired users, we generally want natural sounding voices and to use TTS in the same way we would listen to podcasts or a bedtime story.
With this system, users are in full control and can swap TTS models easily. If a model is shipped and, two weeks later, a smaller, newer, or better one appears, their work would become obsolete very quickly.
Just made it an MCP server so claude can tell me when it's done with something :)
https://github.com/Marviel/speak_when_done
How am I supposed to enable this?
[1] https://github.com/acatovic/ova
For voice cloning, pocket tts is walled so I can't tell
It seems like Kokoro is the smaller model, also runs on CPU in real time, and is more open and fine tunable. More scripts and extensions, etc., whereas this is new and doesn't have any fine tuning code yet.
I couldn't tell an audio quality difference.
There's a bunch of inference stuff though, which is cool I guess. And it really is a quite nice little model in its niche. But let's not pretend there aren't huge tradeoffs in the design: synthetic data, phonemization, lack of train code, sharp boundary effects, etc.
If it were a big model and was trained on a diverse set of speakers and could remember how to replicate them all, then zero shot is a potentially bigger deal. But this is a tiny model.
I'll try out the zero shot functionality of Pocket TTS and report back.
Btw, I would love to hear from someone (who knows what they're talking about) to clear this up for me. Dealing with potential GPL contamination is a nightmare.
If you could find another compatible converter, you could probably replace eSpeak with it. The data could be a bit OOD, so you may need to fiddle with it, but it should work.
Because the GPL is outdated and doesn't really consider modern gen AI, what you could also do is to generate a bunch of text-to-phoneme pairs with Espeak and train your own transformer on them,. This would free you from the GPL license completely, and the task is easy enough that even a very small model should be able to do it.
So, on my M1 mac, did `uvx pocket-tts serve`. Plugged in
> It was the best of times, it was the worst of times, it was the age of wisdom, it was the age of foolishness, it was the epoch of belief, it was the epoch of incredulity, it was the season of Light, it was the season of Darkness, it was the spring of hope, it was the winter of despair, we had everything before us, we had nothing before us, we were all going direct to Heaven, we were all going direct the other way—in short, the period was so far like the present period, that some of its noisiest authorities insisted on its being received, for good or for evil, in the superlative degree of comparison only
(Beginning of Tale of Two Cities)
but the problem is Javert skips over parts of sentences! Eg, it starts:
> "It was the best of times, it was the worst of times, it was the age of wisdom, it was the epoch of belief, it was the epoch of incredulity, it was the season of Light, it was the spring of hope, it was the winter of despair, we had everything before us, ..."
Notice how it skips over "it was the age of foolishness,", "it was the winter of despair,"
Which... Doesn't exactly inspire faith in a TTS system.
(Marius seems better; posted https://github.com/kyutai-labs/pocket-tts/issues/38)
- "its noisiest superlative insisted on its being received"
Win10 RTX 5070 Ti
I wonder what's going wrong in there
It says MIT license but then readme has a separate section on prohibited use that maybe adds restrictions to make it nonfree? Not sure the legal implications here.
If a license says "you may use this, you are prohibited from using this", and I use it, did I break the license?
In this case, I'd interpret it as they made up a new licence based on MIT, but their addendum makes it non-MIT, but something else. I agree with what others said; this "new" license has internal conflicts.
Simply, it's MIT licensed. If they want to change that, they have to remove that license file OR clearly update it to be a modified version of MIT.
All too often, new models' codebases are just a dump of code that installs half the universe in dependencies for no reason, etc.
I saw some agentic models at 4B or similar which can punch above its weights or even some basic models. I can definitely see them in the context of home lab without costing too much money.
I think atleast unmute.sh is similar/competed with chatgpt's voice model. It's crazy how good and (effective) open source models are from top to bottom. There's basically just about anything for almost everyone.
I feel like the only true moat might exist in coding models. Some are pretty good but its the only industry where people might pay 10x-20x more for the best (minimax/z.ai subscription fees vs claude code)
It will be interesting to see if we will see another deepseek moment in AI which might beat claude sonnet or similar. I think Deepseek has deepseek 4 so it will be interesting to see how/if it can beat sonnet
(Sorry for going offtopic)
I think they should have added the fact that it's English only in the title at the very least.
Cool tech demo though!
There's about 1.5B English speakers in the planet.
You pull up a map and start navigation. All the street names are in the local language, and no, transliterating the local names to the English alphabet does not make them understandable when spoken by TTS. And not to mention localised foreign names which then are completely mangled by transliterating them to English.
You pull up a browser, open up an news article in your local language to read during your commute. You now have to reach for a translation model first before passing the data to the English-only TTS software.
You're driving, one of your friends Signals you. Your phone UI is in English, you get a notification (interrupting your Spotify) saying 'Signal message', followed by 5 minutes of gibberish.
But let's say you have a TTS model that supports your local language natively. Well due to the fact that '1.5B English speakers' apparently exist in the planet, many texts in other languages include English or Latin names and words. Now you have the opposite issue -- your TTS software needs to switch to English to pronounce these correctly...
And mind you, these are just very simple use cases for TTS. If you delve into use cases for people with limited sight that experience the entire Internet, and all mobile and desktop applications (often having poor localisation) via TTS you see how mono-lingual TTS is mostly useless and would be switched for a robotic old-school TTS in a flash...
> only that but it's also common to have system language set to English
Ask a German whether their system language is English. Ask a French person. I can go on.
Multilingual doesn't mean language agnostic. We humans are always monolingual, just multi-language hot-swappable if trained. It's more like you can make;make install docker, after which you can attach/detach into/out of alternate environments while on terminal to do things or take in/out notes.
People sometimes picture multilingualism as owning a single joined-together super-language in the brain. That usually doesn't happen. Attempting this especially at young age could lead a person into a "semi-lingual" or "double-limited" state where they are not so fluent or intelligent in any particular languages.
And so, trying to make an omnilingual TTS for criticizing someone not devoting significant resources at it, don't make much sense.
This is plainly not true.
> Multilingual doesn't mean language agnostic. We humans are always monolingual, just multi-language hot-swappable if trained
This and the analogy make no sense to me. Mind you I am trilingual.
I also did not imply that the model itself needs to be multilingual. I implied that the software that uses the model to generate speech must be multilingual and support language change detection and switching mid-sentence.
Not abundantly obviously a satire and so interjecting: humans, including professional "simultaneous" interpreters, can't do this. This is not how languages work.
I think it's the wrong example, because this is actually very common if you're a Chinese speaker.
Actually, people tend to say the name of the cities in their own countries in their native language.
> I went to Nantes [0], to eat some kouign-amann [1].
As a French, both [0] and [1] will be spoken the French way on the fly in the sentence, while the other words are in English. Switching happens without any pause whatsoever (because there is really only one single way to pronounce those names in my mind, no thinking required).
Note that with Speech Recognition, it is fairly common to have models understanding language switches within a sentence like with Parakeet.
To me, what you're saying is the same as saying the art of a movie is in the script, the video is just the method of making it available. And I don't think that's a valid take
So yes, I mostly agree with GP. An audiobook is a different rendering of the same subject. The content is in the text, regardless of whether it's delivered in written or oral form.
Set up SherpaTTS as the voice model for your phone (I like the en_GB-jenny_dioco-medium voice option, but there are several to choose from). Add a ebook to librera reader and open it. There's an icon with a little person wearing headphones, which lets you send the text continuously to your phone's tts, using just local processing on the phone. I don't have the latest phone but mine is able to process it faster than the audio is read, so the audio doesn't stop and start.
The voice isn't totally human sounding, but it's a lot better than the microsoft sam days, and once you get used to it the roboticness fades into the background and I can just listen to the story. You may get better results with kokoro (I couldn't get it running on my phone) or similar tts engines and a more powerful phone.
One thing I like about this setup is that if you want to swap back and forth between audio and text, you can. The reader scrolls automatically as it makes the audio, and you can pause it, read in silence for a while yourself and later set it going from a new point.
"so and so," he <verb>
and the verb is not just "said", but "chuckled", or "whispered", or "said shakily", the output is modified accordingly, or if there's an indication that it's a woman speaking it may pitch up during the quotation. It also tries to guess emotive content from textual content, such if a passage reads angry it may try to make it sound angry. That's more hit-and-miss, but when it hits, it hits really well. A very common failure case is, imagine someone is trying to psych themselves up and they say internally "come on, Steve, stand up and keep going", it'll read it in a deeper voice like it was being spoken by a WW2 sergeant to a soldier.
I just tried some sample verses, sounds natural.
But there seems to be a bug maybe? Just for fun, I had asked it to play the Real Slim Shady lyrics. It always seems to add 1 extra "please stand-up" in the chorus. Anyone see that?
https://gist.github.com/britannio/481aca8cb81a70e8fd5b7dfa2f...
https://github.com/agentify-sh/speak/
Of course, English only
Another recent example: https://github.com/supertone-inc/supertonic
https://huggingface.co/spaces/Supertone/supertonic-2
It seems like it is being trained by one person, and it is surprisingly natural for such a small model.
I remember when TTS always meant the most robotic, barely comprehensible voices.
https://www.reddit.com/r/LocalLLaMA/comments/1qcusnt/soprano...
https://huggingface.co/ekwek/Soprano-1.1-80M
In English, it's perfect and it's so funny in others languages. It sounds exactly like someone who actually doesn't speak the language, but got it anyway.
I don't know why Fantine is just better than the others in others languages. Javer seems to be the worst.
Try Jean in Spanish « ¡Es lo suficientemente pequeño como para caber en tu bolsillo! » sound a lot like they don't understand the language.
Or Azelma in French « C'est suffisament petit pour tenir dans ta poche. » is very good.I mean half of the words are from a Québécois accent, half French one but hey, it's correct French.
Però non capisce l'italiano.
[1] (2016 https://arxiv.org/abs/1609.03499) WaveNet: A Generative Model for Raw Audio
[2] (2017 https://arxiv.org/abs/1711.10433) Parallel WaveNet: Fast High-Fidelity Speech Synthesis
[3] (2021 https://arxiv.org/abs/2106.07889) UnivNet: A Neural Vocoder with Multi-Resolution Spectrogram Discriminators for High-Fidelity Waveform Generation
[4] (2022 https://arxiv.org/abs/2203.14941) Neural Vocoder is All You Need for Speech Super-resolution
But don’t do the fake mistake I did and use a hf token that doesn’t have access to read from repos! The error message said that I had to request access to the repo, but I’ve had already done that, so I couldn’t figure out what was wrong. Turns out my HF token only had access to inference.