Each row contains one shared Yoruba transcript and two time-aligned audio columns. audio comes from the first archive's audio folder; audio_2 comes from the second quality-version archive. SBPN and diarization were run once on the first variant and their final timestamps were reused exactly for both columns. HTDemucs was run independently once on every complete recording in each variant. Each column independently applies the documented finite Librosa delta-SNR >= 0 dB vote between original and saved full-vocals audio. Non-music tags remain in published text according to the existing same-speaker association logic; music remains a hard boundary and is excluded from derived speech text.

Rows: 1,783
Recordings: 25
Duration per audio column: 3.049 hours
audio Demucs selections: 1,338
audio_2 Demucs selections: 1,323

Downloads last month: 32

Total file size:

255 MB