Skip to main content
Rook Multi Speaker targets speech naturalness and intelligibility for human listeners. Use it for calls and other applications where people hear the enhanced output. Its multi-speaker models aim to retain speech from both foreground and background participants while reducing noise and reverberation. For machine listening and speech-to-text (STT), evaluate Quail Multi Speaker. For isolating one primary speaker, evaluate Quail Voice Focus.

​
Choose a variant

Rook Multi Speaker is available in small and large variants at 8 kHz, 16 kHz and 48 kHz. Examples are rook-ms-s-16khz and rook-ms-l-48khz. The model reference lists all current IDs and their native block sizes. Match the model’s native rate to the frequency range your application needs, then measure quality and CPU use on target hardware. Processing 48 kHz audio through a 16 kHz model does not make that model enhance frequencies above its native range.

​
Evaluate for listeners

Use paired recordings and listening tests with representative speech, noise, reverberation and overlapping speakers. Check speech retention and artifacts as well as noise reduction. Measure latency through the complete capture, processing and playback path. Rook Multi Speaker can also precede a third-party voice activity detector, but test that detector’s false starts, missed speech and timing with the enhanced signal. Do not assume one enhancement level gives the best result for every detector. Dedicated ai-coustics VAD should receive the original input audio. Start with the SDK quickstart, then use the latency guide to evaluate your call path.