Quantifying and Improving the Performance of Speech Recognition Systems on Dysphonic Speech.

Area of Science:

• Speech-language pathology
• Artificial intelligence
• Machine learning

Background:

• Current speech recognition systems exhibit performance disparities when processing dysphonic speech compared to typical speech.
• The underlying cause is hypothesized to be the underrepresentation of dysphonic voice data in the training datasets of these systems.

Purpose of the Study:

• To quantify the performance of existing speech recognition systems on dysphonic speech input.
• To investigate methods for improving speech recognition accuracy for individuals with voice disorders.

Main Methods:

• A retrospective database of dysphonic speech recordings was utilized.
• Three commercial speech recognition platforms were evaluated against both normal and dysphonic voice inputs.
• A custom speech recognition model was developed using transfer learning on data from patients with spasmodic dysphonia and vocal cord paralysis.

Main Results:

• Commercial platforms demonstrated significantly lower accuracy on dysphonic speech, with accuracy rates ranging from 84.55% to 93.56%.
• Performance deficits were particularly noted for spasmodic dysphonia and vocal fold paralysis.
• The custom-trained model achieved a high accuracy of 96.43% for dysphonic voices and 97.62% for normal voices.

Conclusions:

• Existing speech recognition technology underperforms with dysphonic speech due to limited training data.
• Transfer learning and custom model development can substantially enhance speech recognition performance for dysphonic voices, addressing specific pathologies.