[P2] ZARA-025 — Play a pre-generated immediate acknowledgement without blocking the turn #26

New issue

Closed

opened 2026-07-18 19:24:40 +00:00 by lost-rob0t · 1 comment

lost-rob0t commented

2026-07-18 19:24:40 +00:00

(Migrated from github.com)

Objective

After Zara detects a valid wake/turn start, respond immediately with a short acknowledgement such as “Okay” while speech capture and routing continue.

Required implementation

Add a small configurable acknowledgement set with a deterministic default.
Pre-generate/cache acknowledgement audio at startup for the active voice/provider, or ship an explicitly licensed local fixture for the default acknowledgement.
Trigger acknowledgement from the TurnCoordinator after wake acceptance, not after final transcription or command execution.
Do not send the acknowledgement through the LLM.
Ensure acknowledgement playback does not consume microphone frames as user speech, clip the user’s command, or claim successful execution.
If the user starts speaking immediately, prioritize capture and either suppress, duck, or stop the acknowledgement according to measured behavior.
Acknowledgement failure must not block command processing.

Required tests

Fake audio-output tests for cache hit/miss, startup generation failure, immediate user speech, duplicate wake events, cancelled turns, and disabled acknowledgement.
Verify acknowledgement contains no action-success wording.
Enforce the deterministic fixture budget from ZARA-022: wake-to-first-ack-audio p95 <= 350 ms, with a stricter target where hardware permits.
Add scripts/test-acknowledgement.sh and benchmark coverage.

Acceptance

Warm acknowledgement starts before final STT, Prolog, or LLM work.
Exactly one acknowledgement is associated with a turn.
The user can speak naturally without waiting for it to finish.

Branch

perf/zara-025-immediate-ack

Dependencies

ZARA-018, ZARA-022, ZARA-023, ZARA-024.

## Objective After Zara detects a valid wake/turn start, respond immediately with a short acknowledgement such as “Okay” while speech capture and routing continue. ## Required implementation - Add a small configurable acknowledgement set with a deterministic default. - Pre-generate/cache acknowledgement audio at startup for the active voice/provider, or ship an explicitly licensed local fixture for the default acknowledgement. - Trigger acknowledgement from the TurnCoordinator after wake acceptance, not after final transcription or command execution. - Do not send the acknowledgement through the LLM. - Ensure acknowledgement playback does not consume microphone frames as user speech, clip the user’s command, or claim successful execution. - If the user starts speaking immediately, prioritize capture and either suppress, duck, or stop the acknowledgement according to measured behavior. - Acknowledgement failure must not block command processing. ## Required tests - Fake audio-output tests for cache hit/miss, startup generation failure, immediate user speech, duplicate wake events, cancelled turns, and disabled acknowledgement. - Verify acknowledgement contains no action-success wording. - Enforce the deterministic fixture budget from ZARA-022: wake-to-first-ack-audio p95 <= 350 ms, with a stricter target where hardware permits. - Add `scripts/test-acknowledgement.sh` and benchmark coverage. ## Acceptance - Warm acknowledgement starts before final STT, Prolog, or LLM work. - Exactly one acknowledgement is associated with a turn. - The user can speak naturally without waiting for it to finish. ## Branch `perf/zara-025-immediate-ack` ## Dependencies ZARA-018, ZARA-022, ZARA-023, ZARA-024.

lost-rob0t commented

2026-07-20 04:12:09 +00:00

(Migrated from github.com)

Cache the TTS providers for the responses

Use random phrases like
"ok"
"yep"
"Alright"
"one second..."
so on and so forth do like 15+ ones

Cache the TTS providers for the responses Use random phrases like "ok" "yep" "Alright" "one second..." so on and so forth do like 15+ ones