CHG-0035: Only Words scribe Is Sure Of, and the reSpeaker Asked From Its Console
| Date | 2026-10-24 |
| Change type | Upgrade |
| Status | Complete 2026-10-24 |
| Systems | The services workload: the six services rebuilt, eds-scribe and eds-voiced changed, all restarted. respeaker-01: firmware flashed by cable |
| How | make -C deploy ship WHERE=builder; make -C firmware/ears build flash, from e9ed753 on voice-play-album |
| Risk | Low. Most likely to go wrong: scribe’s thresholds refusing speech it should take, which is a request said again, not a record played wrongly |
| Decisions | ADR-0042, ADR-0043 |
| Related | INC-0001, which this corrects |
Summary
The reSpeaker’s first request played Abbey Road, which nobody asked for (INC-0001): scribe’s model wrote its own prompt back, a request in words, and voiced acted on it. scribe’s prompt is now a list of names, and scribe says whether the words are worth acting on, by Whisper’s own thresholds; voiced acts on nothing else.
The reSpeaker is asked from its USB console until it has a wake word: the voice kit’s buttons are the XMOS chip’s, not the XIAO’s.
Goal
- Silence, noise and the request said outside the window are answered “not heard clearly”, and nothing plays.
- A request said in the window plays its album.
Procedure
make -C deploy ship WHERE=builder;make -C deploy status.make -C firmware/ears build flash.- Enter on the console with nothing said: “not heard clearly”.
- Enter, then “Play Aerosmith’s first album”, said by the owner.
Undo
Ship and flash 1badd8a’s scribe and voiced, and 1c6c645’s firmware;
which is to say, put INC-0001 back. Not done.
Outcome
| When (UTC) | What happened |
|---|---|
| 01:46:01–01:49:27 | ship WHERE=builder from e9ed753: six images, all six restarted; no service timed out connecting. |
| 01:49:50 | respeaker-01 flashed. |
| 01:50 | On the workload, scribe: macOS say’s “Play Aerosmith’s first album” confident (no speech 0.005); five seconds of silence not, and no words. |
| 01:51:41 | Two windows asked for on the console. In the room, someone else was speaking: the first written as talk and not confident (“not heard clearly”), the second as clear talk that was no request (“not a request”). Nothing played. |
| 01:54:23 | The owner, in the window: “Play Aerosmith’s first album.” The WAV reached the broker whole (160,812 bytes, tapped with voiced’s account); scribe wrote it confidently in 3.5 s; voiced asked for Aerosmith at 01:54:30.8; “Make It” played at 01:54:31.1. |
Departures from the plan
- Step 3, nothing said, was the room’s own talk instead of silence: a harder test, passed.
- The device took about 22 s to finish sending the request, though voiced had it within five; its console never showed what was heard. Left as a follow-up.
Follow-ups
- The wake word.
- The device’s send:
esp_mqtt_client_publishof 160 KB held the listening for ~22 s on a -69 dBm link, and the heard reply did not reach the console.