Investigation on the Robustness of Acoustic Foundation Models on Post Exercise Speech

2026-03-29

AI Summary

This study addresses the degradation in robustness of automatic speech recognition (ASR) systems caused by post-exercise physiological changes—such as disordered breathing, vocal instability, and non-semantic pauses—and presents the first systematic evaluation of mainstream acoustic foundation models under this challenging condition. The assessment encompasses both out-of-the-box and domain-finetuned settings across fluent and non-fluent speakers, covering sequence-to-sequence models (e.g., Whisper, FunASR/Paraformer) and self-supervised encoders (e.g., Wav2Vec2, HuBERT, WavLM) paired with CTC decoding. Experimental results show that FunASR achieves the best performance on the Post-All test set (WER 14.57%, CER 8.21%); CTC-based models benefit substantially from finetuning, whereas Whisper exhibits inconsistent gains. Recognition of non-fluent speakers proves significantly more difficult, highlighting the critical influence of model architecture, finetuning strategy, and speaker fluency on ASR performance in post-exercise scenarios.

Abstract

Automatic speech recognition (ASR) has been extensively studied on neutral and stationary speech, yet its robustness under post-exercise physiological shift remains underexplored. Compared with resting speech, post-exercise speech often contains micro-breaths, non-semantic pauses, unstable phonation, and repetitions caused by reduced breath support, making transcription more difficult. In this work, we benchmark acoustic foundation models on post-exercise speech under a unified evaluation protocol. We compare sequence-to-sequence models (Whisper and FunASR/Paraformer) and self-supervised encoders with CTC decoding (Wav2Vec2, HuBERT, and WavLM), under both off-the-shelf inference and post-exercise in-domain fine-tuning. Across the Static/Post-All benchmark, most models degrade on post-exercise speech, while FunASR shows the strongest baseline robustness at 14.57% WER and 8.21% CER on Post-All. Fine-tuning substantially improves several CTC-based models, whereas Whisper shows unstable adaptation. As an exploratory case study, we further stratify results by fluent and non-fluent speakers; although the non-fluent subset is small, it is consistently more challenging than the fluent subset. Overall, our findings show that post-exercise ASR robustness is strongly model-dependent, that in-domain adaptation can be highly effective but not uniformly stable, and that future post-exercise ASR studies should explicitly separate fluency-related effects from exercise-induced speech variation.

Problem

• post-exercise speech
• automatic speech recognition
• robustness
• physiological shift
• speech variation

Innovation

• post-exercise speech
• acoustic foundation models
• robustness
• in-domain fine-tuning
• fluency stratification