Frame-level speaker encoder for streaming diarization. Confidence-gated online clustering at a fixed 960 ms delay.
Most accurate streaming diarizer on low-overlap benchmarks. Degrades less than cache-based systems beyond four speakers. Needs no diarization corpus in training.
Diarization only. No transcript.
Audio is 16 kHz. Every chunk goes through three steps:
VAD (Silero) marks speech versus silence. About 250 ms of confirmation delay.
Encoder turns each 80 ms frame into a 192-dimensional speaker embedding.
Clustering assigns a speaker id only when the embedding is confident enough. Labels come out about 960 ms behind the audio, then consecutive frames of the same speaker become one RTTM line.
