Training-Free Intelligibility-Guided Observation Addition for Noisy ASR

Researchers propose a training-free method for improving automatic speech recognition in noisy environments by intelligently fusing noisy and speech-enhanced audio based on intelligibility estimates. The approach eliminates the need for trained neural predictors, reducing complexity while maintaining robustness across diverse speech enhancement and ASR model combinations.

This research addresses a fundamental challenge in automatic speech recognition: maintaining accuracy when background noise is present. While speech enhancement front-ends have become standard for noise suppression, they frequently introduce artificial artifacts that paradoxically degrade recognition performance. The proposed intelligibility-guided observation addition method tackles this trade-off by dynamically blending the original noisy signal with enhanced speech, weighted by real-time intelligibility assessments derived from the ASR backend itself.

The innovation lies in eliminating the training requirement that plagued previous observation addition approaches. Traditional methods relied on neural network predictors trained on substantial datasets, introducing computational overhead and limiting generalization to unseen acoustic conditions. By extracting intelligibility signals directly from the ASR model's internal representations, this training-free approach achieves superior generalization while reducing system complexity and deployment friction.

For ASR practitioners and developers, this method offers practical benefits across multiple dimensions. The approach works agnostically with different speech enhancement and ASR model combinations, providing flexibility in system architecture choices. Extensive experimental validation across diverse datasets demonstrates consistent robustness improvements over existing baselines without requiring model retraining or architectural modifications.

The research validates both frame-level and utterance-level fusion strategies, with analyses of intelligibility-guided switching alternatives providing insights into optimal implementation decisions. These findings establish a foundation for more resilient speech recognition systems in real-world deployment scenarios where environmental noise remains unavoidable. The training-free design particularly benefits resource-constrained applications and edge deployment scenarios where computational efficiency and rapid adaptation matter.

• →Training-free intelligibility-guided observation addition improves ASR performance in noisy environments without retraining speech enhancement or ASR models
• →The method dynamically fuses noisy and enhanced speech based on intelligibility estimates derived directly from the ASR backend
• →Approach demonstrates superior generalization compared to neural predictor-based observation addition methods
• →Works agnostically across diverse speech enhancement and ASR model combinations
• →Reduces system complexity while maintaining robustness through frame and utterance-level fusion strategies