Researchers Just Gave Dysarthric Speech Recognition a 30% Boost — Here's Why That Could Change Lives

Dysarthric speech recognition remains a critical challenge for assistive communication, plagued by high variability in severity and scarce training data. Researchers Paban Sapkota, Hemant Kumar Kathania, Sudarsana Reddy Kadiri, and Shrikanth Narayanan tackle this by fine-tuning the End-to-End pre-trained Wav2Vec2 model with in-domain data augmentation. They systematically evaluate four techniques: Speaking-Rate Modification (SRM), Pitch Modification (PM), Formant Modification (FM), and Vocal Tract Length Perturbation (VTLP), each designed to mimic the acoustic distortions present in dysarthric speech. The study builds separate baseline models for low, medium, and high severity, then applies severity-specific fine-tuning with augmented data.

The results reveal clear performance patterns. SRM with a scaling factor of 0.8 works best for low and medium severity, achieving WERs of 9.02% and 38.11% — improvements of 30.02% and 16.64% over baselines. For high severity, PM with a pitch shift factor of 0.8 yields the best WER of 55.15%, a 15.47% relative gain. These findings demonstrate that targeted augmentation can significantly close the accuracy gap for severely impaired speakers, making Wav2Vec2-based ASR more practical for real-world assistive technology.