Export of Word-Level Timestamps

[lterfloth]

The code seems to use word-level timestamps internally. For our research, word-level timestamps are more appropriate than the rather ambigous segment level timestamps. We annotate data with ELAN and having word-level timestamps makes annotation single words precise.

Would it be possible to get an export that preserves these word level timestamps? In the past, we used a manual approach of whisperx and diarization, then merged the data together. That worked, but was rather tedious and error prone. Having it included in noScribe would be lovely.

[kaixxx]

In situations of overlapping speech, word-level timestamps are often ambiguous regarding speaker attribution. I use the segmentation from whisper as additional information to group words together and make speaker attribution more robust.
It would be difficult to reintroduce word-level-timestamps without loosing this advantage. Which output format are you using to export to ELAN? WebVTT?

[lterfloth]

I understand. Using our previous workflow, we manually fixed the wrong speaker attribution after importing the transcripts in ELAN. If noScribe would serve the feature, we would most likely use WebVTT, yes. Maybe an additional export that includes the word level timestamps but maybe has no speaker attribution at all (or just a guesstimate)? It would be a very neat feature

[lterfloth]

I created some experimental changes and provided them through #299 . I hope this is fine for you, otherwise please reject. All discussions should, I believe, happen there.