PiVo ships 1.4M-param TTS in a 5.8MB Rust binary, 4x realtime on one CPU thread
solyarisoftware · x · 2026-11-03
Empero released PiVo Studio, a pure-Rust, zero-native-dependency local streaming TTS stack shipping with PiVo-1.4M: a 1.4M-parameter, 5.8MB English TTS model with two voices (Pi and Vo) and the vocoder included.
- No PyTorch, no ONNX Runtime: a hand-written f32 kernel parses and executes ONNX graphs directly, auto-vectorized to AVX2/NEON
- 4x faster than real time on a single CPU thread, 30ms time-to-first-audio
- Ships as one binary with CLI, HTTP server, Web Studio GUI, and an OpenAI-compatible API; runs on desktop, iOS, Android, and browser
- Weights available on Hugging Face; build takes about a minute
A notable example of ultra-small, CPU-only speech generation.
Related event: PiVo: Open-Source 1.4M-Parameter TTS Runs 4x Real-Time on CPU(2 posts)→
More from Infra
- Anthropic paid $2.4B to data vendors this quarter — just 6% of its compute spend — himanshustwts · 2026-11-03
- Anthropic S-1: $17B quarterly revenue vs $895B in signed compute obligations — FinanceYF5 · 2026-11-03
- TensorFold Dev Crowdfunds £24k for M5 Ultra Mac Studio to Push Local Inference — AIFlow_ML · 2026-11-03
- O'Reilly Releases Hands-On LLM Serving and Optimization, a Guide to Hosting LLMs at Scale — kalyan_kpl · 2026-11-03
- Hot Aisle rents 8×MI300X VMs at $2.99/hr amid compute price surge talk — AccBalanced · 2026-11-03
- GPU lead times hit 30–45 weeks as TrendForce flags tightening AI supply chain — AccBalanced · 2026-11-03