BADCH: Backtracking and Adaptive chunking for Longform speech recognition

Developing efficient Automatic Speech Recognition (ASR) systems for long-form audio transcription faces challenges such as overlapping speech, background noise, and boundary errors in segments, impacting accuracy and runtime. Current chunking techniques-such as hard chunking, serial stitching, and overlapping inference-often face issues at segment boundaries. This paper addresses these issues using BADCH, a back-tracking and adaptive chunking approach that reduces boundary errors while decreasing computational demands across diverse audio conditions. Evaluations on TED-LIUM and Hindi datasets indicate that BADCH is better at handling boundary errors with less computing time.