To reduce VAD commit latency: lower vad_silence_threshold_secs (default 1.5s, allowed 0.3–3.0s). You can also tune min_silence_duration_ms and min_speech_duration_ms (50–2000 ms) to make shorter pauses count, then validate against your audio conditions. If VAD still doesn’t meet your needs, switch to manual commit; best practice is to commit during silences and roughly every 20–30 seconds (there is an automatic commit at 90s). Also note that transcript processing starts after the first ~2 seconds of audio are sent. (elevenlabs.io)
Could chunk size 3200 affect VAD or commit timing? Larger chunks increase streaming latency; recommended chunk duration is 0.1–1.0 s, where smaller chunks lower latency but add overhead. If your 3200-sized chunks correspond to ~0.1 s at your format, you’re already at the low end and chunking is unlikely to be the main cause. (elevenlabs.io)
Best practices for tuning vad_threshold vs vad_silence_threshold_secs:
vad_silence_threshold_secs controls how long silence must persist before a commit; reducing it commits sooner. Range: 0.3–3.0 s (default 1.5). (elevenlabs.io)
vad_threshold controls VAD sensitivity; lower values are more sensitive. Range: 0.1–0.9 (default 0.4). There’s no one-size-fits-all recommendation in the docs—evaluate on your audio. You can also adjust min_speech_duration_ms and min_silence_duration_ms (50–2000 ms). (elevenlabs.io)
Is 0.1 s for vad_silence_threshold_secs supported/safe? No—the documented range is 0.3–3.0 s; values below 0.3 s are not accepted. (elevenlabs.io)
Related references
VAD strategy and example config; chunk-size guidance. (elevenlabs.io)
Realtime STT query parameters (ranges and defaults), and manual-commit guidance. (elevenlabs.io)
API reference confirming the same parameter ranges. (elevenlabs.io)
humble furnace
serene dawn