Skip to main content
VAD Multi Speaker detects voice activity from any speaker. Use it when speech from multiple participants should drive transcription or turn-taking, such as in a meeting or speakerphone application. It produces speech probabilities and speech activity decisions while leaving the input audio unchanged. Feed it the original audio, alongside any enhancement processor. It runs independently of Quail Multi Speaker or other enhancement models. If only one primary speaker should trigger activity, use VAD Voice Focus.

​
Model and compatibility

For core SDK 0.24.0, use vad-ms-2.1-xxs-16khz. Its native sample rate is 16 kHz and its optimal native block size is 240 samples (15 ms). See the model reference for artifact details and the compatibility matrix for your language package.

​
Evaluate speech detection

Test quiet speech, pauses, overlapping speakers, noise and music. Track missed speech and false detections separately. Speech from another participant is expected to count as activity with this model. Read the loaded model’s parameter defaults before tuning sensitivity, minimum speech duration and speech hold duration. VAD supplies speech activity; your application combines that signal with transcription and turn detection to decide when a turn is complete. Follow Use and tune VAD for a runnable example, signal routing, timing controls and language API references.