Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Guo, Xiaoyong, Li, Nanjie, Zeng, Zijie, Wang, Kai, Huang, Hao, Xu, Haihua, Shi, Wei
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2603.24034
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866918407688945664
author Guo, Xiaoyong
Li, Nanjie
Zeng, Zijie
Wang, Kai
Huang, Hao
Xu, Haihua
Shi, Wei
author_facet Guo, Xiaoyong
Li, Nanjie
Zeng, Zijie
Wang, Kai
Huang, Hao
Xu, Haihua
Shi, Wei
contents Contextual automatic speech recognition (ASR) with Speech-LLMs is typically trained with oracle conversation history, but relies on error-prone history at inference, causing a train-test mismatch in the context channel that we term contextual exposure bias. We propose a unified training framework to improve robustness under realistic histories: (i) Teacher Error Knowledge by using Whisper large-v3 hypotheses as training-time history, (ii) Context Dropout to regularize over-reliance on history, and (iii) Direct Preference Optimization (DPO) on curated failure cases. Experiments on TED-LIUM 3 (in-domain) and zero-shot LibriSpeech (out-of-domain) show consistent gains under predicted-history decoding. With a two-utterance history as context, SFT with Whisper hypotheses reduce WER from 5.59% (oracle-history training) to 5.47%, and DPO further improves to 5.17%. Under irrelevant-context attacks, DPO yields the smallest degradation (5.17% -> 5.63%), indicating improved robustness to misleading context. Our code and models are published on https://github.com/XYGuo1996/Contextual_Speech_LLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2603_24034
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle From Oracle to Noisy Context: Mitigating Contextual Exposure Bias in Speech-LLMs
Guo, Xiaoyong
Li, Nanjie
Zeng, Zijie
Wang, Kai
Huang, Hao
Xu, Haihua
Shi, Wei
Computation and Language
Artificial Intelligence
Contextual automatic speech recognition (ASR) with Speech-LLMs is typically trained with oracle conversation history, but relies on error-prone history at inference, causing a train-test mismatch in the context channel that we term contextual exposure bias. We propose a unified training framework to improve robustness under realistic histories: (i) Teacher Error Knowledge by using Whisper large-v3 hypotheses as training-time history, (ii) Context Dropout to regularize over-reliance on history, and (iii) Direct Preference Optimization (DPO) on curated failure cases. Experiments on TED-LIUM 3 (in-domain) and zero-shot LibriSpeech (out-of-domain) show consistent gains under predicted-history decoding. With a two-utterance history as context, SFT with Whisper hypotheses reduce WER from 5.59% (oracle-history training) to 5.47%, and DPO further improves to 5.17%. Under irrelevant-context attacks, DPO yields the smallest degradation (5.17% -> 5.63%), indicating improved robustness to misleading context. Our code and models are published on https://github.com/XYGuo1996/Contextual_Speech_LLMs.
title From Oracle to Noisy Context: Mitigating Contextual Exposure Bias in Speech-LLMs
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2603.24034