Oído: open-source speech recognition for ESP32-S3 without a command list

Lokutor released Oído, an open-source open-vocabulary English speech-to-text engine for the ESP32-S3 N16R8 that runs entirely on-device with no cloud, GPU or NPU. The Conformer-CTC Small model ships in int8 (14 MB) and int4 (8.3 MB) formats, but only host and QEMU arithmetic are verified so far; physical-board speed measurements are still pending.

• Runs on ESP32-S3 N16R8: 240 MHz, 16 MB flash, 8 MB PSRAM
• Conformer-CTC Small model in int8 (14 MB) and int4 (8.3 MB)
• Reported lower WER than Whisper tiny.en on LibriSpeech, but not independently benchmarked
• Code is GPLv3; int8 weights CC-BY-4.0, int4 weights CC-BY-SA-4.0