Show HN: Live speech-to-text on an ESP32-P4 microcontroller, no cloud

Category: ai-ml

Tags: embedded, speech-recognition, quantization, simd, risc-v, edge-computing

Score: 8.5/10 (Innovation: 9, Technical: 9, Documentation: 9, Utility: 7)

This project runs a 65.7M-parameter streaming Zipformer2 ASR model entirely on an ESP32-P4 microcontroller, using 2/4-bit quantized weights and hand-written PIE SIMD assembly kernels to achieve 0.70x real-time speech recognition with no cloud dependency. It includes bit-exact integer arithmetic validation against a PC reference, phrase biasing, and six voice-control demos, with thorough WER benchmarks and reproducible evaluation tooling.

Target audience: embedded engineers, edge ML practitioners, speech recognition researchers

Repository: https://github.com/BoumedineBillal/esp32-pie-asr · C++ · Apache-2.0

View on Hacker News