tiny-voca: an open 4.5M-param TTS that runs 70x real time on a CPU
nato_nob · reddit · 2026-11-15
A developer released tiny-voca, a fully open English TTS with code, weights, training recipe and dataset public.
Key specs
- 4,537,291 parameters, 18 MB total including the waveform stage
- Two voices (male/female), 24 kHz
- 70x real time on 4 CPU threads; runs with PyTorch, ONNX, and fully in-browser
- Trained from scratch in 7.7 hours on one RTX 4090
Design choices
- Custom phonemizer (dizione)
- No neural vocoder — a small source-filter synth generates the audio, hence the speed
Limitations
- English only, two voices, no cloning; misreads foreign names and odd spellings; sounds somewhat synthetic since training audio was itself synthesized with Qwen3-TTS
The author asks users to stress-test hard texts and report failures.
More from Multimodal
- mark_k: Killing Grok Imagine would be a huge mistake, ceding image/video to Chinese models — mark_k · 2026-11-15
- Omni Video Generation Space Trends on Hugging Face — CumaSystem · 2026-11-15
- AI glitches become a Seinfeld script: a hobbyist builds his own ComfyUI front-end — AxonkaiLab · 2026-11-15
- A 5,000-word Claude prompt wrangles 9 reference characters into a 20-second Subaru ad — Nimblecloud13 · 2026-11-15
- UChicago study: artists can engineer visual memorability, but computational models capture it better — plopesresearch · 2026-11-15
- Creator finds MiniMax H3 base R2V beats bf variants for voice cloning — Foreforks · 2026-11-15