This dataset combines six independently aligned source archives. Each row embeds its selected MP3 in the audio Parquet column. SBPN-derived word timestamps are observational and do not control chunk edges or the Demucs vote. Accepted hard-word verbalizations are projected back to the original written forms; pronunciation_alignment_dictionary_json records the winning spoken form. Non-music tags are preserved using the existing same-speaker association rules, while music remains a hard boundary. has_any_overlap detects every positive-duration overlap between finalized speech rows assigned to different speakers; same-speaker row overlaps do not count. overlap_percentage reports the union of that cross-speaker duration as a percentage of the chunk. HTDemucs runs once on each complete recording and each chunk selects the original or saved vocals stem from the finite delta-SNR vote.

Rows: 50
Recordings: 1
Audio duration: 0.157 hours
Parquet shards: 1

Downloads last month: 18

Total file size:

39.6 MB