tiny-voca: a 4.5M-parameter English TTS that runs locally in your browser at 21MB

thisdudelikesAI · x · 2026-11-17

tiny-voca is an English text-to-speech model trained from scratch with just 4.5M parameters — 21MB total including the waveform generator (ONNX float32), 24kHz mono output, two American English voices (F2/M2), trained on only 17.6 hours of synthetic speech, and Apache-2.0 licensed. The demo page runs inference entirely in the browser: nothing you type leaves your device, and it shows real-time audio, generation time, speed, and phoneme conversion.

The author is upfront about limits: English only, two fixed voices, no voice cloning, foreign names and odd spellings often mispronounced, and no streaming yet. It's positioned not as a studio tool replacement but as a tiny local English TTS you can actually run for free. Browser demo, GitHub repo, and Hugging Face weights are all available.

Related event: tiny-voca: 18MB TTS model runs in browser at 70x real-time on CPU(4 posts)→

Original post →

More from Multimodal

Multimodal channel →