Who Said What, and Will It Be Remembered? Evaluating Persistent Speaker Attribution Across Meetings
The paper argues that meeting transcripts need a single speaker identity that survives across recordings, and tests systems on that harder requirement.
The authors introduce SI-cpWER, a corpus-level score that uses one global speaker-ID assignment instead of remapping speakers per meeting. They evaluate commercial diarize-then-identify systems, academic baselines, and ThyVoice on CHiME-8 NOTSOFAR and CHiME-6 conditions. Under persistent identity scoring, ThyVoice beats every evaluated commercial cascade in all three conditions, with a 47.13 mean SI-cpWER versus 54.75 for the next system. The paper says this shows persistent attribution has to be measured directly when transcripts are reused as long-term memory. HF Daily Papers' note
The authors introduce SI-cpWER, a corpus-level score that uses one global speaker-ID assignment instead of remapping speakers per meeting. They evaluate commercial diarize-then-identify systems, academic baselines, and ThyVoice on CHiME-8 NOTSOFAR and CHiME-6 conditions. Under persistent identity scoring, ThyVoice beats every evaluated commercial cascade in all three conditions, with a 47.13 mean SI-cpWER versus 54.75 for the next system. The paper says this shows persistent attribution has to be measured directly when transcripts are reused as long-term memory. HF Daily Papers' note
score 4