Paper Overview
Research Area: NLP
Authors: Cheng-Kang Chou, Ming-To Chuang, Ke-Han Lu, Chan-Jan Hsu, Hung-yi Lee
Published: 2026-07-06
arXiv: 2607.05364
Summary
Modern autoregressive ASR systems can emit timestamps directly as decoded tokens, enabling timestamped transcription without frame-level aligners or inference-time post-processing. However, this paper finds that such generated timestamps drift over long non-speech spans: the transcribed text may still look reasonable, but the decoded timeline diverges from the actual audio.
The authors propose REDDIT (REplay-based Distributional EDITing), a lightweight two-stage post-training framework that corrects timestamps while avoiding catastrophic forgetting:
1. Stage 1: Edit the timestamp targets conditioned on the model's own replayed decoder context, while matching the frozen base distribution for non-timestamp tokens. 2. Stage 2: Apply a short edit-prefix refinement phase.
Results
Experiments on Whisper-tiny, using only 34.9 hours of target correction audio and updating just 1.6% of model parameters, show:
- Long-interval mIoU improved from 38.7% to 95.0%
- Mixed-interval out-of-domain AAS reduced from 2752ms to 223ms
- CV-en MER preserved at 41.3%, compared to 524.2% for naive SFT decoder tuning
*Auto-collected on 2026-07-06*