Oído is an open-source speech-to-text engine from Lokutor, written in C, that transcribes English sentences entirely on an ESP32-S3 (240 MHz dual-core, 8 MB PSRAM, 16 MB flash) with no cloud service, no fixed command list and no neural accelerator. The model is NVIDIA's stt_en_conformer_ctc_small, ported to PyTorch and quantized to int8, with a front end of log-mel features and convolution subsampling, 16 Conformer layers and CTC decoding over 1024 BPE tokens. The C engine uses the chip's PIE vector unit for its int8 and int4 kernels, with dual-core scheduling, a VAD/AGC utterance segmenter and an optional SSD1306 OLED. The README reports word error rate on the full LibriSpeech test sets, computed with the on-chip engine's exact arithmetic (a host build bit-identical to the firmware, matched word for word under Espressif's QEMU emulator; physical-board measurements are still to come), with the same text normalization for every system: the int8 model with greedy decoding scores 3.7 on test-clean and 8.2 on test-other at 14.0 MB, and the int4 quantization-aware fine-tune scores 4.6 and 10.0 at 8.3 MB, leaving a 6 MB app partition. An optional 1.3 M-parameter GRU language model used with beam search (beam 4) lowers the int8 result to 3.3 and 7.2. For comparison, the README lists Espressif MultiNet7, whose API takes fixed command lists, at 8.5 and 21.3 on the same chip per Espressif's own ESP-SR benchmark, and Whisper tiny.en at 6.3 and 15.9 and Vosk small (Kaldi) at 9.9 and 21.6 on a laptop, the laptop baselines using 500 evenly spaced utterances per set. Real-time speed is estimated rather than measured: the README gives a factor of 0.7 to 0.95 derived from emulator instruction counts, says it has not yet been measured on silicon, and describes utterance-mode latency of about 3 s for a 2 to 4 s command. It runs on an ESP32-S3-DevKitC-1 N16R8 with an INMP441 I2S microphone (ESP-IDF v5.5), on a laptop with the chip's exact arithmetic, or under QEMU; it is English-only and needs the 16 MB flash / 8 MB PSRAM variant. The code is GPL-3.0; for products that cannot meet its terms, such as consumer devices that do not let users install modified firmware, Lokutor offers commercial licenses. Model weights carry separate terms: nemo8.tnm and the language model are CC-BY-4.0, and nemo4.tnm is CC-BY-SA-4.0. Both models are also on Hugging Face.