Systematic Study of Dysarthric Speech Recognition: Spectral Features and Acoustic Models

Researchers have achieved significant improvements in dysarthric speech recognition by systematically combining acoustic features with the Factorized Time Delay Neural Network (F-TDNN) model, demonstrating 4.65% relative improvement in word recognition and 4.63% in sentence recognition. The study identifies pitch features as particularly effective for handling the acoustic variability characteristic of impaired speech, advancing accessibility technology for individuals with speech disorders.

This research addresses a critical accessibility challenge in speech recognition technology. Dysarthric speech—speech affected by neurological conditions like cerebral palsy, Parkinson's disease, or stroke—presents unique acoustic variability that confounds standard recognition systems. The researchers' systematic approach to feature engineering represents meaningful progress in making speech technology inclusive for individuals with disabilities. By methodically testing different acoustic feature combinations against various acoustic models, they identified that pitch features significantly enhance performance, particularly for sentence-level tasks where context matters. The improvements achieved through the F-TDNN architecture suggest that architectural choices combined with thoughtful feature selection can substantially overcome recognition barriers. This work builds on established hybrid DNN/HMM training approaches but demonstrates concrete performance gains through deliberate frame overlap optimization. Beyond academic interest, dysarthric speech recognition has substantial real-world applications in assistive technology, enabling individuals with speech impairments to control devices and communicate more effectively. The accessibility technology sector continues expanding as aging populations and disability awareness drive demand. These incremental improvements in recognition accuracy translate directly into usability gains for affected populations. The research demonstrates that specialized models need not require extensive labeled data if feature engineering is sufficiently sophisticated. Future development may incorporate these findings into commercial accessibility platforms, improving quality of life for millions globally. The methodology itself—systematic feature evaluation across model architectures—offers a reusable framework for other specialized speech recognition challenges.

• →Pitch features significantly improve dysarthric speech recognition performance across both word and sentence-level tasks.
• →F-TDNN models with optimized frame overlap achieve 4.65% relative improvement over previous state-of-the-art methods.
• →Systematic acoustic feature engineering can compensate for impaired articulatory precision in speech recognition systems.
• →The TORGO database analysis provides validated benchmarks for evaluating dysarthric speech recognition advances.
• →These improvements have direct accessibility applications for individuals with neurological speech impairments.