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

Researchers developed data augmentation techniques to improve automatic speech recognition (ASR) for people with dysarthria by fine-tuning the Wav2Vec2 model. Using methods like speaking-rate modification, pitch modification, and formant modification tailored to different severity levels, the study achieved significant word error rate reductions across low, medium, and high severity dysarthric speech.

This research addresses a critical accessibility challenge in speech technology by tackling dysarthric speech recognition, a notoriously difficult domain due to acoustic variability and limited training data. The study's systematic exploration of four distinct data augmentation techniques demonstrates that different modification approaches yield optimal results depending on speech severity, suggesting that one-size-fits-all solutions are inadequate for assistive speech technology.

Dysarthria affects millions of individuals with conditions like cerebral palsy, Parkinson's disease, and stroke recovery, yet commercial ASR systems perform poorly on these populations. The research builds on Wav2Vec2, a self-supervised pre-trained model that has shown promise in low-resource speech scenarios. By investigating severity-specific augmentation strategies, the authors move beyond generic data multiplication toward intelligent augmentation that mirrors the acoustic characteristics of dysarthric speech patterns.

The results carry significant implications for accessibility and inclusive AI development. Achieving 30% relative improvement for low-severity cases demonstrates that targeted augmentation can meaningfully enhance usability for people with mild dysarthria who might benefit from communication assistance. The diminishing returns at higher severity levels (15.47% improvement) indicate where fundamental model architecture changes may be necessary, guiding future research priorities.

Developers building accessibility features can leverage these augmentation techniques to improve dysarthric speech recognition in real-world applications. The severity-stratified approach provides a framework for other acoustic variation challenges beyond dysarthria. Future work should explore combining multiple augmentation methods and investigating whether severity-agnostic models could match severity-specific performance through better augmentation strategies.

• →Speaking-rate modification proved most effective for low and medium severity dysarthric speech, achieving 9.02% and 38.11% WERs respectively
• →Pitch modification yielded best results for high-severity dysarthria at 55.15% WER, demonstrating technique-to-severity matching requirements
• →Data augmentation produced 15-30% relative improvements across severity levels, confirming viability for addressing speech recognition data scarcity
• →Severity-specific fine-tuning of Wav2Vec2 models outperformed generic approaches, establishing baseline methodology for dysarthric ASR development
• →Research provides practical framework for improving accessibility features in speech technology applications targeting people with speech disabilities