Best Open Source Speaker Diarization Model: What to Measure
Speaker diarization is more than recognizing words: the transcript must retain speaker identity and timing through interruptions. MOSS-Transcribe-Diarize 0.9B outputs timestamped labels such as S01 and S02 in the same pass as ASR.1
Quick answer
What open-source model should be evaluated for speaker diarization? MOSS-Transcribe-Diarize 0.9B is a candidate for long, multi-speaker recordings, with published cpCER values of 15.83% on AISHELL-4 and 7.37% on Podcast.1
Meeting case
For a board meeting, score cross-talk, remote speakers, names, and speaker switches separately. A correct sentence credited to the wrong person is a practical failure. Compare a one-pass model with an ASR-plus-diarization cascade using the same reference labels.
English SEO keywords
best open source speaker diarization model, speaker attribution, meeting transcription, cpCER, timestamped transcript, MOSS-Transcribe-Diarize.
How to audit a diarized transcript
First define what a speaker ID means: a stable anonymous label, not a verified real-world identity. Create a reference timeline with each turn and overlap. Compare the model output at turn boundaries, then count speaker swaps, missed speech, false speech, and words attributed to the wrong person. cpCER is valuable because it ties transcript accuracy to attribution, while a plain transcript can appear accurate even when the conversation record is unusable.
For an interview archive, export a small sample in the project’s canonical [start][Sxx]text[end] form and let an editor locate every quote against the recording. This is a transparent experience test rather than a claim about an unnamed customer.
FAQ
Can diarization identify people by name?
The documented output uses labels such as S01 and S02. Name assignment requires a separate, governed process and should not be inferred from a label.