fix(voice): rolling-window VAD to prevent false barge-in during TTS

[beardedeagle]

Problem

The barge-in VAD used one-shot calibration that froze the noise floor from the first 800ms of TTS playback. Later louder passages exceeded the stale floor and false-triggered barge-in, cutting TTS mid-sentence. The barge-in monitor also opened the mic immediately when TTS started, causing speaker bleed to falsely trigger VAD. Additionally, when token streaming was enabled, TTS sentences were rendered twice — once by the token stream and once by the TTS display callback.

Solution

Rolling-window VAD — 90th percentile of a rolling deque window (~3s), 8x multiplier, 4000 RMS trigger ceiling, min_floor clamped to SILENCE_RMS_THRESHOLD * 2, sustained_ms=1000, calibration_ms=800.

Barge-in grace period — barge_in_grace_seconds (default 2.0s) delays VAD activation so TTS playback establishes first.

Duplicate render suppression — pass display_callback=None to stream_tts_to_speaker when streaming_enabled, making the token stream the sole display path.

Normal-exit flag — _tts_normal_exit prevents the exception finally block from setting stop_event on clean exit.

TTS CUT diagnostic logging at all barge-in cut paths.

Regression tests

• test_quiet_then_loud_playback_does_not_trip
• test_8x_multiplier_absorbs_tts_volume_spikes
• test_trigger_ceiling_lets_genuine_speech_trip
• test_silence_calibration_does_not_false_trip_on_tts

[Copilot]

Pull request overview

Improves voice-mode TTS barge-in behavior by making the VAD floor adaptive during playback, adding a short grace period before the mic starts listening, and preventing duplicate text rendering when token streaming is enabled.

Changes:

• Replaces one-shot VAD calibration with a rolling-window (90th percentile) floor and adds diagnostic logging for VAD trips.
• Adds a configurable barge-in grace period and tunes barge-in VAD parameters from the CLI monitor.
• Suppresses duplicate sentence rendering during streaming TTS and adds a “normal exit” flag to avoid cutting audio on clean shutdown.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.

File	Description
tools/voice_mode.py	Implements rolling-window VAD floor + trigger ceiling and adds diagnostics.
tests/tools/test_voice_mode.py	Adds regression coverage for new VAD behavior (needs one test adjustment to truly cover the multiplier).
cli.py	Adds barge-in grace period + TTS cut diagnostics; avoids duplicate rendering; refines streaming-TTS shutdown behavior (normal-exit flag logic needs a small fix).

Comments suppressed due to low confidence (1)

tools/voice_mode.py:1226

• The comment says the multiplier "was 5.0", but the removed code used a different calibration approach (median * 3.5). Since it's not a direct like-for-like multiplier change, the "was 5.0" parenthetical is misleading and should be removed.

                # 8.0x multiplier (was 5.0): TTS speaker bleed has wide
                # volume variation between sentences and within sentences.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[beardedeagle]

Note on suppressed Copilot comment about 'was 5.0' parenthetical in voice_mode.py: True positive — the old code used median * 3.5, not 5.0. Removed the misleading '(was 5.0)' from the comment. Fixed in 070370f.

[teknium1]

Salvaged into PR #74000 (merged, b6729ba) with your commit authorship preserved in git history. Your rolling-window VAD floor, multiplier tuning, RMS ceiling, and barge grace window were exactly the missing piece that made spoken barge-in reliable — they shipped alongside the fix that armed the barge monitor on the whole-file playback paths, and voice.barge_in_grace_seconds is now a documented config key. Thanks for the careful signal-processing work; live-tested as part of the voice UX polish wave.