Improving Dysarthric Speech Segmentation With Emulated and Synthetic Augmentation

Area of Science:

• Speech processing
• Neurological disease diagnosis
• Machine learning

Background:

• Acoustic features from speech aid in diagnosing neurological diseases and monitoring symptoms.
• Temporal segmentation of audio signals into words is crucial for feature extraction.
• Automatic Speech Recognition (ASR) and sequence alignment can automate speech segmentation.

Purpose of the Study:

• To explore augmentation methods for improving ASR performance on dysarthric speech.
• To assess the feasibility of using emulated and synthetic data for ASR model fine-tuning.
• To reduce reliance on scarce and privacy-sensitive clinical data.

Main Methods:

• Fine-tuning pre-trained ASR models using two augmentation strategies: 1) healthy speakers altering rate/loudness, 2) synthetic speech with varied rate/accent.
• Evaluating model performance on dysarthric speech data.
• Comparing augmentation methods against fine-tuning with real clinical data.

Main Results:

• Augmentation with emulated and synthetic data outperformed models fine-tuned solely on real clinical data.
• Performance matched models fine-tuned on real data combined with synthetic speech.
• The best model achieved 5.7% word error rate and 94.4% accuracy on neurological disease data.
• Mean intersection-over-union of 89.2% for word segmentation against ground truth.

Conclusions:

• Emulated and synthetic data augmentations effectively reduce the need for real clinical data in ASR model fine-tuning for dysarthric speech.
• These methods enhance speech segmentation accuracy for neurological disease assessment.
• The approach offers a scalable and privacy-preserving solution for clinical speech analysis.