Who Said What, and Will It Be Remembered? Evaluating Persistent Speaker Attribution Across Meetings
SI-cpWER measures whether the same speaker keeps one identity across meetings, and ThyVoice leads commercial cascades.
The paper argues that meeting transcripts used as long-term memory must preserve both words and stable speaker identities. Speaker Identified cpWER scores a corpus under one global speaker-ID assignment, unlike metrics that remap anonymous speakers independently in each recording. On the full 129-meeting CHiME-8 NOTSOFAR set, in clean and noise-augmented form, plus CHiME-6, ThyVoice records a mean SI-cpWER of 47.13 versus 54.75 for the next system and beats every evaluated commercial diarize-then-identify cascade.
- Existing metrics cannot track one speaker identity across meetings
- SI-cpWER scores a corpus under one global speaker-ID assignment
- Benchmark covers five commercial cascades and two academic baselines
- ThyVoice mean SI-cpWER is 47.13 versus 54.75 for the next system
From the original article
1 min read at huggingface.co
Speech transcripts used as long-term memory must preserve both words and stable speaker identities. Existing meeting-transcription metrics either ignore speakers or remap anonymous speakers independently in each recording, so they cannot measure whether the same person retains one identity across meetings.
By Shantanu Vispute, Aditya Mishra, Siddhartha Saxena · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.