ADR-0043: The reSpeaker Runs Firmware of EdS’s Own, and Sends a Request When Its Button Is Held, Then When It Hears Its Name

Context

The reSpeaker Lite is two chips. An XMOS XU316 takes two microphones and gives one cleaned-up voice: echo cancelled, noise suppressed, beam steered. Its firmware (1.0.8, read over DFU on 2026-10-24) sends that over I2S to a XIAO ESP32-S3 (8 MB PSRAM, 8 MB flash), which came with an Arduino test sketch from Seeed. The XIAO is what can be programmed.

Options considered

ESPHome, which Seeed and Home Assistant publish for the board. Its voice pipeline is Home Assistant’s, and its microWakeWord is the best on this chip; the rest would be a second way of doing what EdS’s three devices already do.

Firmware of EdS’s own, in ESP-IDF and C, like the stand, the screen and the beacon: credentials in NVS (ADR-0004), updates through the broker (ADR-0035), presence repeated (ADR-0034), events to the tenant’s log store. The wake word is the one new thing, and microWakeWord is a TensorFlow Lite model that runs without ESPHome.

Decision

The reSpeaker runs EdS firmware, firmware/ears. It sends what it heard as one WAV on voice/respeaker-01/utterance: first while a button is held, then after a wake word.

• It listens to the XMOS chip over I2S, the XMOS driving the clock. The firmware measures the rate it is given and says it, and makes 16 kHz mono 16-bit from whichever channel carries the voice, as ADR-0042 asks.
• Held button first. The XIAO’s own button sends what was said while it was held, up to fifteen seconds. Every part of the path can be shown from the device before a wake word exists, and the button stays as the way to ask when the room is too loud for one.
• Then a wake word, microWakeWord with a stock model, then “Hey Ed”: what follows it, until a pause, is sent the same way.
• The rest is the screen’s, copied: the creds partition and make provision from the tenant’s outputs, WiFi on the IoT segment, presence on voice/respeaker-01/status, events on log/respeaker-01, two OTA slots with rollback in the bootloader from the first flash.
• What was heard is answered with a light: the XIAO’s LED while it listens and sends, and a pattern for “not understood”.

Consequences

• Four firmwares, sharing copied files, as the three before do.
• A request is audio on the IoT segment for as long as it takes to send, as ADR-0042 already accepts.
• The XMOS chip is left as Seeed made it. Its firmware is updated, if ever, over its own USB port with dfu-util.

What would settle it

• The rate and channel the XMOS chip sends, measured: 16 kHz on BCLK 8, WS 7, DIN 44, the voice the same on both channels (2026-10-24).
• A request said into it plays its album: asked from the console, not a button, since the voice kit’s buttons are the XMOS chip’s; Aerosmith, 2026-10-24 (CHG-0035).
• The same after a wake word: “Hey Jarvis, play Aerosmith’s first album” played it, 2026-10-24, 21 s from the wake word to the music (CHG-0036).
• An update through the broker, and a rollback.