Live-captioning tuning

Finalization is VAD-silence-gated, not time-gated.

The Silero v5 VAD declares speech-end only after a silence gap of vad_min_silence_ms (default 500 ms). Continuous / broadcast audio has few gaps ≥500 ms, so the default yields roughly one final per 20–30 s (finalizing at CloseStream flush). That is the conservative default working — not a broken pipeline.

Recommendation

For continuous / broadcast captioning, lower the gate:

DDX_STT_VAD_MIN_SILENCE_MS=200   # 200–300 ms recommended

Verified 2026-06-13 on live BBC World Service: 200 ms → 3× the finals (1 → 3 over 30 s), first final 29.4 s → ~14 s, with RTF (~1.05) and TTFT (~0.12 s) unchanged. Trades a few false-splits for usable cadence on gapless audio.

Baseline + methodology: bench/reports/BROADCAST_KPI_BASELINE.md.