Skip to main content
Quail Voice Focus targets one primary speaker while suppressing competing speech, background noise and residual echo. Use it when a voice agent should follow one participant, rather than transcribe every voice in a room. Voice Focus uses recent audio context to select the primary speaker. At the beginning of a stream, suppression can be weaker while that context builds. Test startup with short utterances and quiet speakers; there is no fixed warm-up duration guaranteed by this guide. Keep processing continuous across adjacent frames so the model can retain context. Primary-speaker selection can change when another speaker becomes more prominent, such as when someone moves closer to the microphone. It is not a persistent identity lock. If your application requires identity verification or a fixed enrolled speaker, model selection alone does not establish that behavior.

​
Choose a variant

For core SDK 0.26.0, start with quail-vf-2.3-l-16khz or compare the smaller quail-vf-2.3-s-16khz on your target hardware. Both have a native rate of 16 kHz and an optimal native block size of 240 samples. 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 acquisition details.

​
Check speech retention

Compare original and enhanced recordings using your speech-to-text (STT) provider. Include quiet target speech, overlapping speech, speaker changes and the first seconds after a reset. Measure target-word deletions as well as background-word insertions; stronger suppression can remove speech you intended to retain. For dedicated ai-coustics voice activity detection (VAD), use the original input audio. VAD Voice Focus targets primary-speaker activity when only the target speaker should drive speech activity. A third-party VAD may behave differently on enhanced audio; evaluate its speech boundaries and added delay separately. Follow Voice Focus tuning and the SDK quickstart for the next step.