RMS (int16 scale) at or above which a chunk counts as speech.
Measured on a real Twilio a-leg call against an ElevenLabs callee (100 ms
windows, int16 scale after the SDK's own µ-law→PCM16 24 kHz path, which
applies no gain): between-utterance noise p95 ≈ 740 (-33 dBFS), speech p10 ≈
940 (-31 dBFS), p50 ≈ 1380 (-27.5 dBFS). The margin is thin, which is why the
gate ALSO needs DEFAULT_SPEECH_GATE_HANGOVER_MS: speech dips under the
threshold for up to ~200 ms on plosives and breaths, and a bare threshold
would split one utterance into hundreds of fragments. Tune via
speechGate.rmsThreshold on the adapter options.
RMS (int16 scale) at or above which a chunk counts as speech.
Measured on a real Twilio a-leg call against an ElevenLabs callee (100 ms windows, int16 scale after the SDK's own µ-law→PCM16 24 kHz path, which applies no gain): between-utterance noise p95 ≈ 740 (-33 dBFS), speech p10 ≈ 940 (-31 dBFS), p50 ≈ 1380 (-27.5 dBFS). The margin is thin, which is why the gate ALSO needs DEFAULT_SPEECH_GATE_HANGOVER_MS: speech dips under the threshold for up to ~200 ms on plosives and breaths, and a bare threshold would split one utterance into hundreds of fragments. Tune via
speechGate.rmsThresholdon the adapter options.