REDDIT: Correcting Model-Generated Timestamp Drift in ASR without Forgetting via Replay-Based Distribution Editing

arXiv cs.AI · 18 hours ago

Researchers developed REDDIT, a post-training method that corrects timestamp drift in autoregressive speech recognition systems caused by long non-speech spans, while preventing degradation of transcription quality. The method uses replay-based distribution editing and was tested on 15 ASR systems, improving long-gap alignment from 38.7% to 95.0% mIoU on Whisper-tiny while updating only 1.6% of model parameters. This approach enables better timestamped transcriptions without requiring human annotations or separate alignment tools.