Datasets:
Akhand Paath sp01-03: pangti keys and Shabad windows (two-model aligned, audited rule)
#7
by JashanSingh126 - opened
The three Akhand Paath recordings whose truth was only the sequence of Shabads read (onsets unknown) get pangti keys from whole-clip forced alignment with two speech models (the app's karansea int8 and surindersinghssj pa-v3-kirtan), kept only where both models put the pangti within 1 s and the heard words match (check >= 0.5): sp01 125 pangtis / 11 Shabads, sp02 124 pangtis / 11 Shabads, sp03 120 pangtis / 12 Shabads. shabads.csv keeps the Shabads read inside the 10-minute cut, windowed first-to-last pangti. manifest.csv changes lines_n, lines_source and shabads_n on those three rows only. Same rule as GAB's other aligned-2model-audited keys (97% right pangti against GAB-Hard's human keys; owner spot check 20 of 20).
Arash2348 changed pull request status to merged