Oído: Open-Source Speech Recognition on a $5 Chip
Описание и применение
Oído обеспечивает распознавание произнесённых предложений на английском языке на микроконтроллере ESP32‑S3 с WER 3,7 % на тест‑clean LibriSpeech, превосходя существующие решения и поддерживая полностью открытый исходный код.
Задача
On microcontrollers, speech recognition is limited to wake words or a short command list, preventing open-ended English speech.
Результат
Oído achieves 3.7 % WER on LibriSpeech test‑clean and 8.4 % WER in noisy conditions on a $5 ESP32‑S3, outperforming existing on‑chip recognizers and comparable laptop models, while remaining fully open source.
Рабочий процесс
- 1. Clone repository and build host demo:
git clone https://github.com/lokutor-ai/oido && cd oido/esp32/host && make python live_demo.py - 2. Flash firmware to ESP32‑S3‑DevKitC‑1 N16R8 with INMP441 microphone using one‑line script in README
- 3. Run on‑device inference; model weights stored in flash, processed in int8 or int4 arithmetic
- 4. Optionally enable on‑chip language model for lower WER
Доказательства
Oído, int8 – ESP32‑S3 – test‑clean 3.7%
Oído makes 2.3 to 2.6 times fewer errors than Espressif's own recognizer
The int8 version is within about a tenth of a point of the original full‑precision model (3.70% vs 3.68% on test‑clean).
Everything is on GitHub at github.com/lokutor-ai/oido
- Дата записи: 2026-10-01T11:41:03.795Z