Mitigating Speaker Leakage in Cascaded Multi-talker ASR with Diarization-based Transcript Correction
The method introduces a pruning-based correction for speaker leakage in cascaded multi-talker Automatic Speech Recognition (ASR) systems using an independent, pre-trained speaker diarization model. It identifies and removes leaked segments by requiring a tripartite consensus of acoustic containment, lexical cross-validation, and temporal alignment. This approach significantly improves transcript accuracy by ensuring only high-confidence leakage is pruned.