Oído is an open‑source speech recognition system that outperforms Whisper‑tiny while running on a $5 ESP32‑S3 microcontroller with 8 MB PSRAM and no GPU or NPU. Using the NVIDIA Conformer‑CTC Small model (13 M parameters, int8), it achieves LibriSpeech WER of 3.7/8.2 versus 6.3/15.9 for Whisper tiny.en on a laptop, and under real‑world noise its mean WER is 8.4 compared to 12.1 for Whisper. The code and a live demo are available on GitHub.
Read original
reddit/r/LocalLLaMA