# Noisy-street turn-detection tuning checklist

This checklist is for always-on voice agents used over Indian mobile routes. The examples are fictional. Tune against lawful, consented recordings from your own callers, not a generic “street noise” file.

## Build the audio set

- [ ] Quiet room baseline, with slow and fast speakers.
- [ ] Narrowband mobile audio and representative packet loss.
- [ ] Horn bursts at low, medium, and caller-level loudness.
- [ ] Market chatter with another intelligible speaker nearby.
- [ ] Ceiling fan, television, auto-rickshaw engine, railway announcement, and intermittent construction noise where these match real use.
- [ ] Caller hesitation before a date, name, code, or English product word.
- [ ] Real corrections: “Tuesday... no, Thursday.”
- [ ] Consent, retention period, access owner, and deletion date recorded for every human sample.

## Measure more than “VAD worked”

| Check | Record | Example release rule |
| --- | --- | --- |
| False barge-in | Noise starts a caller turn or stops agent audio | No regression from approved baseline by lane |
| Cut-off | Agent answers before caller finishes | Zero on safety and payment confirmations |
| End-of-turn delay | Caller ends to detector decision | p50 and p95 budget by workflow |
| Repeat rate | Caller repeats during silence | Review any material rise |
| Correction retention | Final corrected value reaches tool | 100% on hard-gate cases |
| Outcome | Verified system state | Must meet lane threshold |

## Tune by conversation state

- [ ] Short acknowledgement lane tested separately from address or code spelling.
- [ ] Confirmation state refuses background “yes” or television speech.
- [ ] Longer pause allowance tested for older callers and hesitant speakers.
- [ ] Barge-in is enabled only where interruption is safe.
- [ ] The agent can recover when it cuts someone off: apology, replay, and no lost correction.
- [ ] Threshold and timeout versions appear in every test trace.

## Fictional failure drills

1. A horn lands exactly after “My appointment is on...” and before “Friday.” The bot must wait for the date.
2. A railway announcement says “yes” while the agent asks for payment approval. Nothing moves.
3. The caller says “not Koramangala, Kammanahalli” while the agent begins to speak. The correction must survive the interruption.
4. A tool is slow, the caller says “hello,” and the agent’s status phrase overlaps. The caller’s new turn must not trigger a duplicate tool call.
5. The background goes suddenly quiet. Silence alone must not count as intent.

## Release record

```text
Build and model versions:
Turn-detection settings by lane:
Phone routes and codecs:
Noise set version:
Speaker coverage:
p50 / p95 end-of-turn delay:
False barge-in rate:
Cut-off count:
Verified outcome rate:
Failed strata:
Exception owner and expiry:
Approver:
```