Supertonic WebGPU: blazingly fast text-to-speech running 100% locally in your browser.
Posted by xenovatech@reddit | LocalLLaMA | View on Reddit | 10 comments
Last week, the Supertone team released Supertonic, an extremely fast and high-quality text-to-speech model. So, I created a demo for it that uses Transformers.js and ONNX Runtime Web to run the model 100% locally in the browser on WebGPU. The original authors made a web demo too, and I did my best to optimize the model as much as possible (up to \~40% faster in my tests, see below).
I was even able to generate a \~5 hour audiobook in under 3 minutes. Amazing, right?!
Link to demo (+ source code): https://huggingface.co/spaces/webml-community/Supertonic-TTS-WebGPU
* From my testing, for the same 226-character paragraph (on the same device): the newly-optimized model ran at \~1750.6 characters per second, while the original ran at \~1255.6 characters per second.
VeterinarianNo5972@reddit
that speed jump is pretty wild especially for something running fully in the browser because most local tts setups choke on longer passages. if you end up exporting those long audiobook files and need them trimmed or converted for players that still struggle with newer codecs uniconverter handles that cleanup step without touching the actual voice quality.
Due-Function-4877@reddit
No thanks. It's a neat toy, but Openrail licensing is a trainwreck. People that break the rules don't care about what the law says. Why would bad actors and criminals care what the license says?
In practical use, alll that openrail license does is create more headaches for people that play by the rules. It reminds me of invasive DRM. It doesn't stop the bad guys and it makes the application more difficult to use for everyone else.
Jagerius@reddit
It's outputting gibberish for me no matter what language/text.
Randommaggy@reddit
Worked fine on Chrome for me, gibberish on firefox.
Steuern_Runter@reddit
Does it support other languages?
Gear5th@reddit
It is ultra fast! But also not really useable because of the errors it makes.
Keeps fumbling simple words. Every now and then the next sentence starts before the current has ended, so the voice overlaps.
Seems like the demo here uses a different model which is much better? https://huggingface.co/spaces/Supertone/supertonic#interactive-demo
Mrdifi@reddit
I want a chatbot with speech not Text to Speech!!!! VOICE-VOICE
ZealousidealShoe7998@reddit
get a model like llama , qwen3 or gemma and connect to this text to speech.
tell it in the system prompt its using a text to speech so avoid using certain behavior like html tags etc.
ogden9133@reddit
Then get that?
dumh3t3r@reddit
Neat!
Saw that there were two more voice files, I made a fork that shows those as well: https://github.com/dumheter/Supertonic-TTS-WebGPU Although not a fancy hugging face page, you would have to run it locally.