Hey HN! I've been working on Kitten TTS for the past 6 months - a tiny text-to-speech model that runs entirely on CPU. At just 25MB, it's perfect for embedded applications and privacy-focused projects where you don't want to send text to external services.
The model is trained on a diverse dataset and supports multiple languages. Performance is surprisingly good for the size - would love to hear your thoughts!
The model is trained on a diverse dataset and supports multiple languages. Performance is surprisingly good for the size - would love to hear your thoughts!
This is incredible! I've been looking for something exactly like this for my IoT project. The fact that it's only 25MB and runs on CPU is a game changer. How does the voice quality compare to larger models?
reply | flag
Thanks! The voice quality is obviously not as good as larger models like Tacotron2 or the commercial APIs, but it's surprisingly decent for most use cases. I'd say it's about 75% of the quality at 1% of the size. Perfect for things like reading notifications aloud or simple voice interfaces.
Great question! It's a combination approach. I started with a modified Tacotron architecture but with much smaller hidden dimensions, then used knowledge distillation from WaveNet to get the vocoder down to size. Also heavily quantized the weights and used some pruning techniques. Happy to share more technical details if you're interested!
[5 more comments]