Limitations and Known Behaviors
Audio Stream Requirements
Individual input audio streams must contain only a single speaker’s audio. During silence, audio samples must be zero. If background noise is present, enable audio thresholding via useAudioThresholdToDetectActiveAudioStream and audioThresholdDb. For details, refer to Configuration Reference.
Latency
The Active Speaker Detection model requires an initial 3-second processing window to produce its first result. A look-ahead queue then buffers output frames—default 30 frames at 30 fps ≈ 1 second—to ensure smooth downstream transmission. This gives a fixed end-to-end latency of approximately 4 seconds at 30 fps with default settings.
The look-ahead queue depth is configurable via outputFrameBufferSize. For parameter details, refer to Configuration reference.
Throughput
The Active Speaker Detection NIM is optimized to adhere to SMPTE ST 2110 real-time processing standards. Performance scales with the number of tracked faces, diarized audio streams, and the GPU SKU used. Higher-end GPUs and lighter workloads achieve greater throughput, while more demanding combinations (more faces and more audio streams) require more compute.