sahilfarib/aimscribe-bengali-eval

AIMScribe Bengali evaluation data

Evaluation data from my AIMScribe RA (Speech AI) assessment. Code: https://github.com/sahilaf/Aimscribe

Folder	What
consultation/	2 real role-play doctor–patient recordings (my scripts, 2 speakers), RTTM + transcripts. Timestamps drafted from a system output, then corrected by ear.
tts/	Clinical test: 100 sentences × 6 voices (Indic Parler-TTS), with drug / dose / test names listed per sentence
tts_consultation/	6 synthetic consultations (patient first, silent patient, 2 attendants, overlap, noise), RTTM + transcripts
samples/diarization/	9 clips with audio, reference RTTM, overlap RTTM and predicted RTTMs before / after
samples/asr/	10 OpenSLR 53 clips with transcripts

Sources: OpenSLR 53 (CC BY-SA 4.0) and Indic DiarBench (sarvamai/indic-diarbench, CC BY 4.0) for the sample clips; the rest was made for this assessment. The synthetic sets are cleaner than real speech.