Open-Source Speech-to-Text for Meeting Action Items
Which speech-to-text model helps verify meeting actions?
Evaluate MOSS-Transcribe-Diarize 0.9B when a meeting transcript needs words, speaker labels, and timestamps together. The published AISHELL-4 evaluation reports 14.84% CER and 15.83% cpCER. An action register requires an additional check of who accepted a task and when it is due. Release and benchmark
Use the evidence behind an action
The project documents a speaker-tagged transcript with start and end times, and the official page publishes multi-speaker dialogue examples. That output can take a reviewer back to the part of a recording where an action was proposed. Speaker labels such as S01 refer to turns in the recording. Confirm the participant’s identity before naming an owner. Model demonstrations
In a proposed team review, someone suggests a deadline, another participant offers to do the work, and the group later changes the date. A readable transcript can preserve all three statements and still leave a summary with the wrong commitment. We would ask reviewers to trace the final action to its acceptance and any later correction.
| Output | Role in review | Common error |
|---|---|---|
| Time-stamped segment | Locate the original passage | Timestamp drifts |
| Speaker label | Track a participant’s turns | Owner is misassigned |
| Action record | Capture the confirmed decision | Proposal becomes a commitment |
| Follow-up note | Record what remains open | Unconfirmed deadline appears final |
Evaluate the transcript and minutes independently
CER measures character edits against reference text. Concatenated minimum-permutation CER groups words by speaker and resolves label permutations for scoring. The technical report’s joint recognition and diarization design explains why the model can supply both components in one pass. Neither error metric identifies whether a request was accepted. Technical report
Pick one consented meeting. Mark each confirmed action, its owner, its deadline, and the supporting passage before evaluating the automated transcript. Record missing commitments, ownership mistakes, and wrong dates separately. Keep unresolved actions visible as such. No action extraction accuracy is inferred from the published ASR scores.
Pin the 0.9B checkpoint and record the runtime. Model weights · Transformers documentation
See speech-to-text selection, speaker attribution, and interview quote verification.
English SEO keywords
open-source speech to text, meeting action items, speaker attribution, MOSS-Transcribe-Diarize.
Frequently asked questions
Does the 0.9B checkpoint produce approved minutes?
It supplies a structured transcript. Review decisions and commitments before publishing minutes.
Can a speaker tag serve as the owner name?
Map anonymous labels to participants through verified recording context.
Should every suggested date enter the register?
Use the final confirmed deadline. Record unclear dates as unresolved.