---
title: "REDDIT: Correcting Model-Generated Timestamp Drift in ASR without Forgetting via Replay-Based Distribution Editing"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.05364v1"
permalink: "https://feed.astralor.com/items/itm_6fa11e7e91fc8f45bb84bd24"
published: "2026-07-06T17:40:54.000Z"
score: 80
channel: "ai"
tags: ["ASR","OpenAI","Whisper-tiny","Post-training framework","Timestamp drift","Catastrophic Forgetting"]
---

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

## AI commentary

> This paper is highly recommended for ASR practitioners, especially those working on long-form audio or real-time captioning. It tackles a common yet often overlooked problem of timestamp drift in deployment, with a lightweight method that requires no expensive annotations. If you are using Whisper or other end-to-end ASR models, REDDIT's approach can be directly adapted to improve temporal alignment accuracy.

## At a glance

Modern autoregressive ASR systems can emit timestamps as decoded tokens, enabling timestamped transcription without frame-level aligners. However, these timestamps drift across long non-speech spans. This work proposes REDDIT (Replay-based Distribution Editing), a lightweight two-stage post-training framework that corrects timestamp drift while avoiding catastrophic forgetting. Using only 34.9 hours of targeted correction audio and updating 1.6% of model parameters, it improves long-gap mIoU from 38.7% to 95.0%, reduces mixed-gap out-of-domain AAS from 2752 ms to 223 ms, and preserves CV-en MER at 41.3%.

## Key points

- Autoregressive ASR systems exhibit timestamp drift across long non-speech spans, with plausible transcripts but misaligned time axes.
- Naive timestamp-corrected fine-tuning causes catastrophic forgetting, degrading non-target ASR behavior.
- REDDIT edits timestamp targets under replayed decoder context while matching frozen base distribution, requiring no human transcripts or timestamp annotations.
- On Whisper-tiny, only 1.6% of parameters updated; long-gap mIoU raised to 95.0% and drift nearly eliminated.

## Full text

Modern autoregressive ASR systems can emit timestamps as decoded tokens, enabling timestamped transcription without frame-level aligners or inference-time post-processing. We show that these generated timestamps can drift across long non-speech spans: the transcript may remain plausible, but the decoded time axis drifts away from the audio. We study this non-speech-induced timestamp drift with self-built gap and long-gap benchmarks across 15 evaluated timestamp-producing ASR and audio-language systems. Naive timestamp-corrected fine-tuning improves alignment but can severely degrade non-target ASR behavior, exposing a forgetting problem. We propose REDDIT(REplay-based Distribution eDITing), a lightweight two-stage post-training framework that corrects timestamps while avoiding this catastrophic forgetting: it first edits timestamp targets under the model's own replayed decoder context while matching the frozen base distribution on non-timestamp tokens, then applies a short edited-prefix refinement stage. In this framework, we construct correction supervision without human transcripts or human timestamp annotations by combining VAD-trimmed speech spans with inserted non-speech gaps and known concatenation offsets. On Whisper-tiny, 34.9 hours of targeted correction audio used and only 1.6% of model parameters updated, raising long-gap mIoU from 38.7% to 95.0% and reducing mixed-gap out-of-domain AAS from 2752 ms to 223 ms while preserving CV-en MER at 41.3% (versus 524.2% for ordinary SFT decoder tuning).

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_6fa11e7e91fc8f45bb84bd24