Lokutor's Oído transcribes any English sentence on an emulated $5 ESP32-S3

A 14 MB model on a chip with 512 KB of fast internal memory is 28 times larger than the memory where compute is cheap ^[4][1]. The weights stay in flash. The ESP32-S3 reads flash through a small cache at a few tens of megabytes per second ^[5]. "The obstacle isn't arithmetic, it's memory bandwidth," Lokutor wrote ^[6].

Lokutor's new C inference engine for the chip's vector instructions is designed around that limit ^[7]. Its integer matrix kernels do 16 multiply-adds per instruction. Attention runs in integers too, softmax included, so the whole model stays in int8 ^[7]. The schedule is the clever part. Audio goes through in blocks of 64 frames, and each weight is read from flash once per block instead of once per frame ^[8]. Weight traffic fell from 18 to 7.5 MB per second of audio, a 58% cut, with both cores working in parallel ^[8][2][7].

Quantisation cost less than the post's own "about a tenth of a point". The int8 model scores 3.70% on LibriSpeech test-clean against 3.68% at full precision, a 0.02-point gap ^[11][5]. The post's text does not give an error rate for the smaller int4 build. An optional on-chip language model brings the reported error rates down to 3.3% and 7.2% ^[13].

The benchmarks describe Lokutor's test audio, so whether they hold for a product depends on how close that product's audio is. The noise result, about 31% fewer errors than Whisper tiny ^[4], applies to rooms that sound like the car, kitchen and cafeteria recordings, background chatter and echo Lokutor built its conditions from ^[10]. The Espressif comparison is harder to read. A command matcher is built for its own list, and LibriSpeech is the standard English benchmark ^[9][20]. A product that only ever needs its command list should test both recognisers on its own phrases.

Lokutor argues that a device streaming audio to a GPU "pays for inference for as long as it exists" ^[21]. Its verification is careful. Every transcript and accuracy figure comes from the exact arithmetic of the on-chip engine, and the firmware reproduces the same transcripts word for word in Espressif's QEMU emulator ^[14]. A host build runs the chip's integer math on a laptop, with a live microphone demo ^[18]. At the slow end of the speed estimate, about 5% margin remains before processing a second of audio takes a full second ^[15][6]. Oído is what cooks in a Spanish kitchen call out to confirm an order ^[19]. So far its firmware has done its listening inside QEMU ^[14].

I think the host build is enough to start a prototype. Hardware needs an ESP32-S3-DevKitC-1 N16R8 board and an INMP441 microphone ^[18], and the model handles only English for now ^[16].