How Robust Are Intent Classifiers to ASR Noise? A Systematic Comparison of Traditional Models, BERT, and Large Language Models

Voice-based systems depend on ASR transcription before intent classification. ASR errors degrade downstream accuracy, but no prior work has systematically compared TF-IDF classifiers, BERT, and instruction-tuned large language models across controlled noise levels on the same benchmark. This paper presents such a comparison on the SLURP spoken language dataset (66 intent classes, 13,006 test utterances) at word error rates of 0%, 10%, 20%, 30%, 40%, and 50%. We fine-tune Qwen2.5-3B-Instruct using QLoRA and compare it against TF-IDF + SVM, TF-IDF + Logistic Regression, and BERT-base. Results show that the fine-tuned LLM maintains the highest Macro F1 at every noise level above zero, degrading by 20.8 points from clean to WER=50% compared to 28.2 points for BERT. TF-IDF models show a two-phase degradation pattern not previously reported in the literature. We also establish the first published LLM classification baselines on SLURP. Code, noisy test sets, and results are included in the supplementary ZIP file.