Samaali — ASR Post-Processing (Whisper + Alignment + Confidence + Semantics)
Audio
Original Text (Ground Truth)
Whisper model size (forced)
compute_type (forced)
VAD filter
Use MARBERT (forced)
Transcribe & Evaluate
Corrected Transcript (ASR errors restored)
Raw ASR Transcript
Report (scores & thresholds)
Token-level Decisions
Token-level Decisions
ASR_word
⋮
GT_word
⋮
status
⋮
reason
⋮
used
⋮
ASR_word
⋮
GT_word
⋮
status
⋮
reason
⋮
used
⋮