The relationship between perceptual disturbances in dysarthric speech and automatic speech recognition performance

Area of Science:

• Speech processing
• Biomedical engineering
• Computational linguistics

Background:

• Automatic speech recognition (ASR) systems excel with typical speech but falter with dysarthric speech.
• Dysarthric speech exhibits significant acoustic variability linked to specific subtypes, impacting ASR accuracy.
• Understanding the relationship between perceptual characteristics and ASR performance is crucial for improving accessibility.

Purpose of the Study:

• To investigate the correlation between perceptual features of dysarthric speech and the performance of ASR systems.
• To identify key perceptual dimensions that most influence ASR accuracy in dysarthric speakers.
• To develop a predictive model for ASR performance based on speech perception.

Main Methods:

• Collected perceptual ratings for 32 dysarthric speakers across various dimensions.
• Analyzed the performance of a standard ASR algorithm on the speech data from these speakers.
• Utilized statistical analysis to explore the relationship between perceptual ratings and ASR outcomes.

Main Results:

• ASR performance on dysarthric speech was found to be predictable from perceptual disturbances.
• Articulatory precision emerged as the strongest predictor of ASR performance.
• Prosody was identified as the second most significant factor influencing ASR accuracy.

Conclusions:

• Perceptual assessment of dysarthric speech can offer insights into ASR system performance.
• Targeting articulatory precision and prosody in ASR development may enhance recognition of dysarthric speech.
• This research provides a foundation for more robust and inclusive speech recognition technologies.