Skip to main content
VAD Voice Focus detects voice activity from the primary speaker. Use it when a voice agent should respond to one participant while avoiding activity triggered by background conversations or competing speakers. It produces speech probabilities and speech activity decisions while leaving the input audio unchanged. Feed it the original audio. It runs independently of Quail Voice Focus, which produces enhanced audio. If speech from every participant should count, use VAD Multi Speaker.

​
Model and compatibility

For core SDK 0.26.0, use vad-vf-2.1-s-16khz. Its native sample rate is 16 kHz and its optimal native block size is 240 samples (15 ms). Core SDK 0.26.0 requires model file format 8. Match the model to your package in the compatibility matrix; see the model reference for artifact details.

​
Evaluate primary-speaker detection

Test quiet target speech, overlapping speech, background conversations and changes in speaker prominence. The primary speaker can change as speakers become more or less prominent; this model does not identify or enroll a particular person. Track missed target speech and activity triggered by interfering speakers separately. Tune sensitivity and speech timing against labeled recordings, starting from the loaded model’s defaults. Follow Use and tune VAD for the shared processing example and parameters. Use this model’s ID in the example to evaluate primary-speaker activity.