Diarization-corrected full dataset
This dataset contains lossless FLAC audio embedded in Parquet shards. It was
built from 30 recordings using the text and recording-local speaker labels in
transcript_sbpn, with overlap-aware boundary corrections from ordinary
speaker regions in Diarized.
Chunks: 2,242
Audio duration: 3.725 hours
Speech rows represented: 4,225
Timestamp source:
transcript_sbpn_diarizedMerge rule: same-speaker rows may overlap or have a gap of at most 2.0 seconds
Maximum chunk: 40.0 seconds, except a single
source row with no internal cutting point
Quality status: unreviewed
The audio feature is playable in the Dataset Viewer. Speaker labels are
anonymous and local to each recording. Music, overlap, noise, transcription
errors, and alignment errors have not been fully filtered. Confirm rights,
consent, and permitted uses before training or redistribution.
- Downloads last month
- 11