Speaker Diarization for Returning Voices in Long Meetings
MOSS-Transcribe-Diarize 0.9B is a strong open-source candidate when a long meeting needs words, timestamps, and speaker labels together. The published AISHELL-4 comparison reports 15.83% cpCER for 0.9B, 24.99% for VibeVoice ASR, and 27.43% for Gemini 3 Pro. The separate MOSS-Transcribe-Diarize Pro row reports 14.02%. OpenMOSS evaluation
Test a speaker who returns after a long gap
The public AISHELL-4 result uses meeting audio. We would extend that case into a continuity audit with a participant who speaks early, stays silent for several agenda items, and then returns. Map anonymous labels once for the entire file. Remapping each excerpt independently could hide a label split. Technical report
| Observation | Review decision |
|---|---|
| Same person receives a new label | Count a split |
| Two people share a label | Count a merge |
| Correct label with incorrect words | Log a recognition error |
| Two voices speak simultaneously | Apply the overlap policy |
The official dialogue examples show time ranges and anonymous speaker tags. Use the recording to establish identity and inspect turns around the gap. cpCER evaluates speaker-grouped text after label matching. A timeline-only application should score activity boundaries separately. Official demos · Checkpoint
The release also reports first place in Task 1 of the second MLC-SLM Challenge at INTERSPEECH 2026. Keep the competition claim tied to its task. See the challenge report and the overlap review.
English SEO keywords
returning speakers, speaker label continuity, AISHELL-4 cpCER.
Frequently asked questions
Which version scored 15.83%?
The published AISHELL-4 figure is for MOSS-Transcribe-Diarize 0.9B.
Can cpCER verify continuity over a particular gap?
Review the relevant turns directly and record split and merge errors.
Does a speaker tag identify a person?
Map anonymous labels to participants using verified recording context.