Automatic Speech Recognition for Intelligibility Assessment in Children With Dysarthria.

Area of Science:

• Speech-language pathology
• Computational linguistics
• Assistive technology

Background:

• Accurate speech intelligibility assessment is crucial for children with cerebral palsy-related dysarthria.
• Traditional methods like human transcription and perceptual ratings are time-consuming and subjective.
• Automatic Speech Recognition (ASR) offers a potential objective and efficient alternative.

Purpose of the Study:

• To evaluate the efficacy of ASR systems for assessing speech intelligibility in children with dysarthria.
• To identify optimal ASR systems that correlate with human listener judgments.
• To explore ASR's potential as a clinical tool for this population.

Main Methods:

• Five ASR systems transcribed speech samples from 20 children with dysarthria.
• 168 adult listeners provided orthographic transcriptions and ease of understanding (EoU) ratings.
• Word Recognition Rate (WRR) was calculated for ASR and human transcriptions; Spearman correlations assessed relationships.

Main Results:

• Four ASR systems (WhisperX-small, -medium, -large, and Google Cloud) showed strong correlations with human WRR, with WhisperX-medium being highest.
• These four systems also demonstrated moderate-to-strong correlations with human EoU ratings, led by Google Cloud ASR.
• Wav2Vec2 showed weak correlations with both human WRR and EoU ratings.

Conclusions:

• ASR holds significant promise for speech intelligibility assessment in children with dysarthria.
• WhisperX-medium is recommended for approximating human transcription accuracy.
• Google Cloud ASR is suggested for aligning with perceptual ease of understanding ratings.
• Careful selection of ASR systems is vital for effective clinical application.