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

The project demonstrates open-vocabulary, live English speech recognition running entirely on an ESP32-P4, without cloud access or a neural accelerator. It uses a 65.7-million-parameter streaming Zipformer2 transducer whose weights are compressed from 71 MB to 26.65 MB using a mix of 2- and 4-bit quantization. On two 400 MHz RISC-V cores, the system processes 1.28 seconds of audio in 0.90 seconds, faster than real time; words are typically finalized about two seconds after being spoken. On LibriSpeech, it reports word error rates of 4.34% on test-clean and 10.26% on test-other. The compressed model loses about half a percentage point of accuracy versus the original, while the chip’s integer arithmetic adds no further loss.

The firmware includes editable phrase lists and six voice-control demonstrations, including smart-home, kitchen, and robot-vacuum scenarios. On 140 synthetic speech clips, phrase biasing reduced word error from 19.5% to 3.4%, with no false alarms in 40 clips without the listed phrases. In tests of each voice-control demo, phrase lists raised exact command recognition to 94-99%, with no false actions among 88 near-miss examples; actions typically followed an order by about two seconds. The system’s specialized SIMD kernels, tests, and PC build are provided for verification. Running it requires an ESP32-P4 revision 3 board with 32 MB PSRAM, a microphone, display, and microSD card; the model loads from the card into memory at startup.