Oído: Open-Vocabulary Speech Recognition on a $5 Microcontroller without a Neural Accelerator
Oído runs open-vocabulary English speech recognition on a $5 ESP32-S3 without a neural accelerator.
Oído is an inference engine that runs open-vocabulary English speech recognition on an ESP32-S3 with a 240 MHz dual-core CPU, 512 KB of SRAM, and no neural accelerator. Its 13-million-parameter int8 Conformer-CTC reaches 3.70%/8.23% word error rate on LibriSpeech test-clean/test-other, within 0.12 points of full precision, and 3.33%/7.15% with a 1.3M on-chip language model; a transducer head reaches 2.99%/6.67%. On a physical board, one core runs at 1.97x the audio duration, so a 4-second command finishes about 8.7 seconds later, while the two-core schedule is disabled because it fails on silicon. It makes 2.3-2.8x fewer errors than the vendor recognizer on the same chip and beats full-precision Whisper tiny.en and Moonshine tiny; code and models are released.
- 13M int8 Conformer-CTC runs on an ESP32-S3 without an NPU.
- LibriSpeech WER is 3.70% clean and 8.23% other.
- One core runs at 1.97x audio duration, not yet real time.
- Errors are 2.3-2.8x lower than the vendor recognizer.
- Spanish model makes 34-60% fewer errors than Whisper tiny.
- Vendors
- Espressif
- AI models
- Whisper tiny.enMoonshine tiny
From the original article
1 min read at arxiv.orgSpeech recognition on low-cost microcontrollers is, in practice, limited to keyword spotting or a fixed list of commands. We present Oído, an inference engine that runs open-vocabulary English speech recognition entirely on an ESP32-S3 (240 MHz dual-core Xtensa LX7, 512 KB of SRAM, 8 MB of PSRAM, 16 MB of flash, no neural accelerator).
By Daniel Varela Sánchez · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.