Three ESP32 Builds Show How Far On-Device AI Has Come in a Week

A $5 ESP32-S3 Learns Open-Vocabulary Speech, No Cloud

Developer lokutor-ai has released Oído, an open-vocabulary speech recognition system built to run entirely on a $5 ESP32-S3, with no cloud connection required. The project's own benchmark claims a 3.7% word error rate on the LibriSpeech test set, the standard yardstick used in speech recognition research.

That number is self-reported, and LibriSpeech is a relatively forgiving benchmark: read, scripted speech recorded in quiet conditions, not the noisy, accented audio a device actually hears in a kitchen or workshop. The repository doesn't yet publish results on noisier, real-world test sets.

Still, open-vocabulary recognition — understanding words outside a fixed command list — has historically needed far more memory than a microcontroller offers. Squeezing it onto a chip this cheap, caveats and all, is a meaningful data point for anyone building a voice interface that never phones home.

A Freenove ESP32-S3 Board Gets a Full Offline Voice Chain

Developer iamankushpandit published esp_ai, a fully offline voice assistant for the Freenove FNK0104B ESP32-S3 board, chaining free-form speech-to-text, a small on-device language model, and text-to-speech into one loop with no network calls. The project was created on September 19, 2026, with commits as recent as October 26, 2026.

The repo doesn't publish word-error-rate or latency numbers for the combined pipeline, so how it holds up on open-ended questions versus simple commands is unverified. Running all three stages back to back on one chip is also likely to be slower than the single-task demos that dominate this space.

Even so, stitching a full STT-LLM-TTS loop onto a single ESP32-S3 board, instead of offloading any stage to a server, is exactly the kind of incremental hobbyist proof that keeps nudging what counts as possible on a $10 chip.

An ESP32-P4 Compute Module Plugs Into Raspberry Pi Carrier Boards

Espressif's developer blog highlighted SCINTIX P4, a compute module from RELOC built around the ESP32-P4, on October 28, 2026. The module is designed to drop onto Raspberry Pi CM4/CM5 carrier boards and bundles memory, power regulation, multiprotocol wireless and Ethernet connectivity onto one package, letting existing Pi-compatible carrier designs swap in Espressif's vision-oriented chip.

Espressif's post doesn't list a price or ship date, so this reads as a development-kit announcement rather than a retail launch. The P4 has no radio of its own, so like other recent P4 boards, SCINTIX P4 pairs it with a companion wireless module to get online.

The CM4/CM5 socket format matters because it gives an established ecosystem of carrier boards and enclosures built for Raspberry Pi's Compute Module line a path to host an ESP32-P4 running local vision or voice models instead.

None of these three ship with a benchmark anyone outside the project has verified, but together they sketch the same shape: speech, language and vision workloads keep getting pushed further onto chips that cost less than a cup of coffee.