¡Oído! Speech recognition that runs faster than speech on a $5 chip
This page runs the same C engine and the same int4 model that run on an ESP32-S3 microcontroller, compiled to WebAssembly. English, streaming: words appear while you talk. Nothing leaves your browser.
The transcript appears here.
What you are running
- A 13 M-parameter Conformer-CTC with int4 weights (7.7 MB, or 5.9 MB for B), 80 ms frames, and a 1.3 M-parameter GRU language model (beam search), exactly as on the chip.
- On one 240 MHz ESP32-S3 core the same code runs at 0.72× the audio duration (faster than speech) with the final text about 430 ms after you stop talking, measured on a physical board.
- LibriSpeech test-clean / test-other, streaming with 0.64 s chunks: 7.1 / 16.4 % word error rate (A), 9.8 / 21.0 % (B). It is English only, and noisy speech is still hard.
- The model files (≈ 9 MB) are downloaded once from GitHub when you pick a variant; audio never leaves this page.