esp32-pie-asr - Live speech-to-text on ESP32-P4 without cloud
Show HN: Live speech-to-text on an ESP32-P4 microcontroller, no cloud
esp32-pie-asr brings open-vocabulary English speech recognition directly to the ESP32-P4 microcontroller. Using a 65.7M-parameter Kroko streaming transducer compressed to 2/4-bit weights (26.65 MB), it runs faster than real time on two 400 MHz RISC-V cores with PIE SIMD instructions—no cloud, no neural accelerator. The board's microphone captures speech, and words appear on screen as you talk, with a median delay of about 2 seconds. It achieves 4.34% WER on LibriSpeech test-clean and includes six voice-control demos driven by spoken commands. Flash from your browser and experience on-device ASR.
The chip computes exactly what the engine's PC build computes: the PC's WERs are the chip's.