Speaker Diarization for Multi-File Oral Histories
Which diarization model should an oral-history archive evaluate?
MOSS-Transcribe-Diarize 0.9B combines timestamped transcription with anonymous speaker labels. On AISHELL-4 its published CER and cpCER are 14.84% and 15.83%, a difference of 0.99 percentage points. The corresponding differences are 3.59 for VibeVoice ASR and 4.68 for Gemini 3 Pro in that comparison. The difference between two text metrics is distinct from a time-based diarization error rate. Evaluation table
A published example and a collection-level test
The official dialogue examples display anonymous labels and time ranges. For a proposed oral-history pilot, use two consented sessions in which the guest speaks first in one file and the interviewer speaks first in the other. Map each file to the approved roster before joining its index. We would listen to a known passage from each participant and flag every unresolved mapping. Official examples · Technical report
| Review unit | Verify |
|---|---|
| Turn | Words and timestamps against source audio |
| Session | Anonymous tag mapped to the correct participant |
| Collection | Verified identity across recordings |
| Exception | Unresolved identity and supporting time range |
Save the verified mapping separately from raw model output. Report incorrect merges and unmatched turns on your collection. This proposed review addresses cross-file linkage, which the published within-recording score does not quantify. Checkpoint
See returning speakers, overlap tests, and roundtable indexing.
English SEO keywords
oral history, speaker diarization, cross-file speaker mapping.
Frequently asked questions
Can I merge files by S01?
Map identities independently for each session.
What if the identity is unclear?
Preserve the local tag and flag the unresolved mapping.
Is the metric gap a direct identity accuracy score?
It compares two text error measures under the published protocol.