tiny-voca: 4.5M-param TTS trained on synthetic speech, 70x realtime on CPU

aigclink · x · 2026-11-16

Developer @n4ze3m trained tiny-voca, a 4.5M-parameter TTS model built entirely on synthetic data: 17.6 hours / 16,018 English utterances generated by Qwen3-TTS. It runs on a regular CPU at 70x realtime (4 threads), with total GPU training cost of $9.56. Instead of a neural vocoder it uses a Source-Filter architecture — predicting pitch, energy, and voicing, then synthesizing waveforms via a source signal and dynamic filter — which is what keeps the model so small. A custom phoneme converter handles context-aware pronunciation of numbers, dates, and amounts. English-only for now, no voice cloning or streaming, but it demonstrates distilling lightweight on-device models from high-quality synthetic speech.

Related event: tiny-voca: 4.5M-Parameter TTS Trained Fully on Synthetic Speech(4 posts)→

Original post →

More from Multimodal

Multimodal channel →