CHG-0038: A Misheard “Play” Is Taken for One Where the Rest Is an Album by Its Place, and the Pre-roll Is 600 ms

Date2026-10-24
Change typeUpgrade
StatusIn progress
SystemsThe services workload: nowplayd and voiced sent and all restarted. respeaker-01: flashed by cable
Howmake -C deploy ship WHERE=builder, the first with load.sh (INC-0002); make -C firmware/ears build flash
RiskLow. Most likely to go wrong: a sentence that is not a request read as one; only “X’s Nth album” and “the Nth album by X” are read behind a misheard “play”
DecisionsADR-0042, ADR-0043
RelatedCHG-0037

Summary

After CHG-0037 and ADR-0044, “Hey Jarvis, play Aerosmith’s first album” played in 8 s from the wake word: 2 s heard, sent in 89 ms, 4 s to write down, 1.7 s to count. The music dipped while it listened. “Play Van Halen’s first album”, straight after, was written “Tae Van Halen’s first album.” and was not a request.

  • voiced takes “tae”, “clay”, “pray” and their like for “play” at the start of a request, but only where the rest is an album by its place with the possessive said, or “the Nth album by X”; never after a negation, and never for “ALBUM by ARTIST”.
  • The device keeps 600 ms from before the wake was decided on, not 350.

Goal

  • “Hey Jarvis, play Van Halen’s first album” plays it.
  • Only nowplayd’s and voiced’s images cross to the workload.

Procedure

  1. make -C deploy ship WHERE=builder.
  2. make -C firmware/ears build flash.
  3. The owner asks.

Undo

git revert this change’s commit; ship and flash.

Outcome

When (UTC)What happened

Departures from the plan

Follow-ups