Introducing Kitten TTS V0.8: open-source TTS that fits in 25MB. Three variants: 80M | 40M | 14M (<25MB) Highly expressive. Runs on CPU. Built for edge. No GPU? No problem. Ship voice anywhere. Check it out:

Github: Discord:

The real breakthrough isn't the file size. It's the death of the cloud dependency tax.

exactly!

finally, my fridge can talk back to me without needing a 4090. 14m on edge is actually insane. the local agent revolution is just getting started.

Yes! We’re gonna unlock ondevice voice agents in the next 90 days

@ron_joshi got it running on iOS with 1:1 quality from the python version check it out:

yoo thats sickk af

This is awesome ! just added Kitten TTS as a model in our Claude Code TTS plugin. Runs fully local which is nice. Can't wait to test the new voices

great, thanks!

I seriously expected this to just be a model trained on cats and it just meows.

now that would be real innovation

Let's go!!

can it do 🇯🇵 japanese ?

Not yet, but we have a multilingual model releasing in the coming weeks. Do you have a prod usecase for japanese?

yes. Quite a process actually. I have a Japanese Language learning app & a browser extension. The app has on device whisper for react native -> local llm (qwen 2.5/liquid ai) -> stt. but the stt part is the sketchy bit when it comes to being ondevice. So far found good result with onnx model/supertonic (not on device) but looking for something robust that can do Japanese better

@ron_joshi Oh bro, this is actually a sick idea for a language learning app. Very low friction. I'm building out a project along a similar vein. Curious to know where you plan to deploy the TTS part, as opposed to the STT part?

@ron_joshi STT is on device. & I tried with onnx models for TTS part on device using sherpa-onnx. Worked well for english. Fell apart for Japanese. then used supertonic, which is just a node.js server that can be hosted anywhere like hetzner / digital ocean

@anilkeshwani japanese is planned to be supported. in case you're interested: after the multilingual launch, we'll have an stt model out

@anilkeshwani I am quite interested thank you very much

😻

wow pretty cool Rohan! are you accepting donations via github? would love to donate fees to you bro

@kingslayer067 thanks a lot, glad you liked the model! havent set the donations up. we are providing commercial support for custom deployments, voices and languages tho.

no worries dude! so you dont need to setup anything, i can just redirect fees directly to ur github and you'd be able to claim it. lmk!!

oh wow this is awesome. how can i get setup?

thanks a lot, there are two ways: 1. Collison installation. or 2. pip install more instructions in the readme:

LOL option 1 is crazy - gotta pip install this

hell yeahh, lmk what you think.

Excellent work. Can it generate TTS for long text like articles, whatnot?

Yes, we’re gonna be putting out lots of demos with these kind of usecases ;) Streaming will be pushed to the repo by this weekend

I’ll try it out on my RPi. Last year I had tried out piper tts

great lmk what you think

Finally something faster than kokoro

And less than a fifth of the parameters ;)

Thanks, I’ve been looking for something like this. Perfect fit for my hardware and low latency tasks

yoo thanks a lot, what do you wanna run it for? lmk if i can help get it running.

i have built an openclaw like agent with google adk. i want to now add a tool so it can send me voice notes fast. kokoro is nice but still slower for me

🔥🚀🚀 😸

Thanks bro! runanywhere integration soon? 👀

#tinymodels ❤️ Nice voices.. very clear segway to companion sort of stuff. Lots of personality.

thanks, more voices and voice cloning coming v v soon ;)

Damn this is insane , have been playing with some voice models lately, will give it a spin !

@Hamzeml Amazing, have to try this soon!

@Hamzeml thanks a lot, would love to hear your thoughts.

Can it run on iOS and macOS?

Yes! But to make it easy, ios sdk is in the works. Will be out in 2-3ish weeks!

Awesome Job. Mac not supported? ONNX Runtime only *guarantees* support for models stamped with official released onnx opset versions. Opset 20 is under development and support for this is limited. The operator schemas and or other functionality may change before next ONNX release and in this case ONNX Runtime will not guarantee backward compatibility. Current official support for domain ai.onnx is till opset 19.

custom inference engine for arm and metal support is coming asap which is way faster than onnx. wanna use it in production? would love to provide custom support for that. lets chat

Guys i love it! Please get it to work with Spanish! I really love this, it makas for awesome toys and other interesting ideas

Spanish will be out in 3-ish weeks!

Does it support Chinese?

we are providing commercial support for chinese. The multilingual model- which will be out in 3 weeks will support chinese.

Thanks a lot. Will it be in Taiwanese Traditional Chinese, not Mainland Chinese?

@hnshah Any webgpu build yet? Would love to add this in

@hnshah not built yet. feel free to contribute. or let us know if you want us to build it. our roadmap is on the github readme and includes mobile and browser sdk. Would love to support any production apps so pls lmk how i can help.

@ldenoue @hnshah attempt to get this working on WebGPU

@ldenoue @hnshah The Micro and Mini models use int8 quantized ONNX ops (MatMulInteger, ConvInteger, DynamicQuantizeLSTM) which aren't supported by WebGPU in ONNX Runtime Web. Exporting FP32 (or FP16) versions of those models would fix it.

@ldenoue @hnshah yes great point, will be out by tomorrow or so!

Does it support French?

it will in the next release eta: 2-3 weeks.

purr

It's in Pinokio

Small models enable edge deployment. This democratizes AI by removing GPU dependencies. The future is on-device.

absolutely. we're gonna be releasing many more cloud-quality models for edge use-cases.

Does it support multiple languages? Like Chinese?

yes, we will be supporting multiple languages, within chinese, within 3-ish weeks. this release is english only.

Cheering for you, bro!

thanks m8

Nice! Spanish support?

Spanish and many more languages coming in 3ish weeks! ;)

Perfect!

@ron_joshi bro, it’s crazy. It’s amazing. I tried it on my edge Android device with Sherpa ONNX, but bro why it felt too slow compare to Android’s own on device TTS. I tried kitten tiny.

must be a bug or some env issue. We’re launching an inference engine + sdk soon that’ll be blazing fast. Stay tuned ;)

@llmgram

@hnshah interesting… is there a way to stitch it directly with Chrome? Like “Read Aloud” from Edge?

@hnshah yes we're launching a browser sdk and im sure someone will build an edge chrome extension for reading screens. if no one does, we'll just do it lol. we're all about edge processing ;)

Does it work on mobile devices for local inference?

yeah , we're launching a mobile sdk(eta: ~2-3 weeks) to make this a lot easier. join our discord to join the beta testing group for this .

@hnshah Amazing stuff Rohan!

@hnshah thanks a lot mate! glad you liked it

its surprising how you need just 25 MBs to produce human like audio but LLMs need to have a LOT of knowledge to produce coherent text.