Noise-Augmented Training Cuts LLM Degradation Under ASR Errors by 47%: A Study on Spoken Intent Classification

Large language models (LLMs) fine-tuned for intent classification degrade under automatic speech recognition (ASR) noise, but prior work has not examined whether noise-augmented training can systematically reduce this degradation. We investigate this question using the SLURP spoken language benchmark with 66 intent classes. We train two versions of a Qwen2.5-3B QLoRA model: one on clean transcripts only, and one on a mixture of clean and synthetically noisy transcripts at word error rates (WER) of 10%, 20%, and 30%. The noise-augmented model (40% clean training data) reduces total Macro F1 degradation from 20.8 percentage points to 11.1 points, a 47% relative improvement in robustness. A mixing ratio ablation reveals that 60% clean data achieves the best overall trade-off: it attains the highest clean Macro F1 (0.806) while degrading only 10.6 points to WER=50%. At WER=50%, the best noise-augmented model outperforms BERT-base by 17.9 Macro F1 points. All augmented training data, noisy test sets, and model adapter weights are released for full reproducibility.