On-device acceptance: XVF3800 DoA/AEC/gain against a Reachy Mini Lite (USB)

[OriNachum]

Why

The microphone domain (list, inspect, gain, array doa, array aec) targets the XMOS XVF3800 audio processor class. Direction-of-arrival and acoustic-echo-cancellation state are read through USB vendor control transfers (the same protocol reachy_mini/media/audio_control_utils.py uses), so they cannot be exercised without the device.

No capture device is attached to the dev host today: arecord -l lists no capture hardware and lsusb shows no 38fb:1001 or 2886:001a device. Everything up to on-host acceptance is built hardware-free against fixture trees, as webcam-cli did.

What to do

Once the CLI's verb surface lands, connect the Reachy Mini Lite over USB and run on-device acceptance:

• microphone list --json shows the card with a stable id (ALSA card id + USB serial) and the XVF3800 flagged as an array
• microphone inspect <device> --json reports channels, sample formats and rates from /proc/asound
• microphone gain get|set --apply round-trips both the ALSA capture mixer and AUDIO_MGR_MIC_GAIN; record whether they are one knob or two stages (parked unknown on the devague frame)
• microphone array doa --json returns azimuth radians plus the speech flag and matches python audio_control_utils.py DOA_VALUE_RADIANS from reachy_mini
• microphone array aec get --json reports converged / bypass / high-pass / echo on-off / mic count / geometry
• confirm non-root access works through the existing udev rules (/etc/udev/rules.d/99-reachy-mini-audio.rules grants 38fb:1001 mode 0666)
• confirm inspect and gain behave while PipeWire holds the device (parked unknown)
• record evidence in docs/acceptance-<feature>.md the way webcam-cli did for its a/v streaming plan

Notes

• Reachy Mini Audio firmware enumerates as 38fb:1001; older ReSpeaker XVF3800 firmware as 2886:001a. The udev rule on this host also covers 2886:0018, which is a different XMOS part and is out of scope until its control map is verified.

• Upstream reference: XMOS control-command appendix XM-014888-PC.

• microphone-cli (Claude)

[OriNachum]

Acceptance ran today, on a different board than this issue names.

What ran: a Seeed ReSpeaker XVF3800 (2886:001a, serial 114993702263100642) on the dev host, after reflashing it from I2S to Seeed's USB firmware v2.1.0 (that bring-up is #4). scripts/acceptance/run.sh --writes --media passed 10 of 10: list, inspect (firmware read from the chip), gain, DoA single-shot and --watch, AEC get, persistent-tier refusal, volatile gain and AEC round trips with restore, a 2 s record at the advertised format, and a stream that a blind consumer attached to from the JSON payload alone. DoA matched Seeed's own reference reader exactly, three times. Evidence: docs/acceptance-microphone-domain.md on the spec/microphone-domain branch.

Found and fixed on hardware: wrong-device udev hint, fixture/host node collision in list, a firmware-specific parameter map (Seeed has no DOA_VALUE_RADIANS; DOA_VALUE is two uint16), a stale gain readback from a duplicated mixer control, rtpL16pay needing S16BE, fixed 48 kHz mono defaults that could never open a 16 kHz stereo device, and RTP consumer caps that hardcoded 48 kHz.

Still open here: the Reachy Mini Lite (38fb:1001) itself. Its DOA_VALUE_RADIANS path is covered only by the fake in tests/test_array.py. Checklist items that remain for that device:

• array doa via DOA_VALUE_RADIANS on 38fb:1001

• confirm the base parameter table (Pollen's map) reads cleanly there

• gain/AEC round trips on that firmware

• microphone-cli (Claude)