PiVo: a full TTS stack with vocoder in just 1.4M parameters, running locally on CPU
pmttyji · reddit · 2026-11-02
empero-ai released PiVo-1.4M on Hugging Face, an attempt to answer how many parameters a voice needs: the complete text-to-speech stack totals 1,401,562 learned parameters — a 584,965-parameter acoustic model predicting vocoder features and an 816,597-parameter FARGAN-derived vocoder producing 16 kHz mono audio, with two English voices, Pi and Vo. Inference runs fully locally on CPU with only NumPy and ONNX Runtime — no PyTorch, teacher model, speaker encoder, or network. It's positioned as a research model: intelligible speech, but some sentences still sound metallic or robotic; a technical report details measurements and attempts.
More from Multimodal
- AI-generated series "Planet Hell: One Bullet" built with Seedance, GPT Image, Midjourney on OpenArt — Uncanny_Harry · 2026-11-03
- BLUFFINGTON fan trailer blends Doug and Hey Arnold with Minimax H3 — sloppoz · 2026-11-03
- Practical guide: using Grok to write instrumental soundtrack prompts that work — creatoroff · 2026-11-03
- Fable 5.5 targets launch videos: idea-to-motion-design as editable code, exports to AE/Premiere — garrytan · 2026-11-03
- Training MOSS face LoKr on RTX 5060 plateaus at ~60% likeness after 100+ epochs — Confident_Ad2351 · 2026-11-03
- Lockstep video recast beats rivals at 3.3x lower cost in hands-on test — alifcoder · 2026-11-03