Improving End-to-End Speech Recognition for Dysarthric Speech through In-Domain Data Augmentation

Automatic Speech Recognition (ASR) systems often struggle with dysarthric speech due to high acoustic variability and limited training data. This study investigates whether applying traditional data augmentation techniques during the fine-tuning of pre-trained Wav2Vec2 models can improve recognition accuracy across different levels of dysarthria severity.

The researchers evaluated four data augmentation methods—Speaking-Rate Modification (SRM), Pitch Modification (PM), Formant Modification (FM), and Vocal Tract Length Perturbation (VTLP)—on the TORGO dysarthric speech corpus. They fine-tuned the Wav2Vec2-large model separately for each severity level (low, medium, and high). By systematically varying the modification factors for each technique, the authors identified the optimal augmentation parameters for each severity class to maximize performance.

The study demonstrates that the efficacy of augmentation techniques is highly dependent on the severity of the dysarthric speech. SRM (with a scaling factor of 0.8) yielded the best results for low and medium severity, achieving relative Word Error Rate (WER) improvements of 30.02% and 16.64%, respectively. For high-severity speech, PM (with a factor of 0.8) proved most effective, providing a 15.47% relative improvement. The results confirm that severity-specific augmentation helps mitigate data scarcity and enhances the robustness of end-to-end ASR systems.

This research provides a practical framework for improving accessibility tools for individuals with motor speech disorders. By demonstrating that targeted, severity-aware data augmentation can significantly reduce transcription errors, the study offers a clear path for developers to optimize ASR performance for users with varying degrees of dysarthria, even when labeled training data is limited.