Open-Source Speaker Diarization for Overlapping Dialogue
Which diarization model should you test on overlapping dialogue?
Include MOSS-Transcribe-Diarize 0.9B if the deliverable is a transcript with timestamps and speaker labels. The Apache-2.0 release combines speech recognition and diarization. If a project requires an explicit timeline of simultaneous speakers, verify that its output format and scoring policy cover that use case. Release repository
Read the public dialogue demo at turn level
The official page presents a scene where two speakers discuss buying a watermelon. The transcript alternates anonymous labels across short turns and includes start and end times. The demo illustrates how quickly attribution can change during a conversation. It provides a place to inspect the output structure. Official demonstrations
For a proposed panel-recording audit, choose passages containing interruptions, brief acknowledgments, and actual simultaneous speech. An interruption followed by silence and two speakers talking at once need different annotations. We would keep both in the review set because a single total can hide which kind caused the error.
| Review question | Suitable evidence |
|---|---|
| Is speech assigned to the right speaker? | Reference speaker timeline and diarization error rate |
| Are spoken words retained under the right label? | cpCER and error examples |
| Does the output represent overlapping intervals? | Inspection against marked overlap regions |
| Does a voice get split into several identities? | Speaker mapping across distant segments |
Compare diarization with speaker-attributed recognition
Diarization error rate measures time-based errors such as missed speech, false alarms, and speaker confusion under a stated scoring setup. The scoring collar and treatment of overlap affect the result. The published AISHELL-4 numbers for MOSS-Transcribe-Diarize 0.9B are 14.84% CER and 15.83% cpCER. Those figures describe text and speaker-grouped text errors. An overlap-specific DER claim requires another evaluation. Technical report · Benchmark table
Have two reviewers mark the difficult segments and resolve disagreements before scoring. Test a dedicated diarization route alongside joint recognition if a multi-speaker activity timeline is mandatory. Keep that comparison tied to the same recording and scoring settings.
Use the official checkpoint and consult the serving cookbook for output handling. For adjacent decisions see diarization selection, SOTA attribution evidence, and meeting action review.
English SEO keywords
open-source speaker diarization, overlapping dialogue, DER, cpCER, MOSS-Transcribe-Diarize.
Frequently asked questions
Does cpCER measure overlapping speech directly?
It measures speaker-grouped text error after matching labels. Use overlap-marked reference intervals to assess simultaneous speech.
Should I include short acknowledgments?
Include them if they matter to the deliverable. Report the scoring policy for short turns.
Is a joint model a good starting point?
Yes when the output needs readable words, timestamps, and attribution together. Assess dedicated diarization for a timeline-only workflow.