Severity-based adaptation with limited data for ASR to aid dysarthric speakers

Area of Science:

• Speech and Language Processing
• Assistive Technologies
• Biomedical Engineering

Background:

• Automatic Speech Recognition (ASR) systems face challenges with speech-impaired individuals due to limited speech databases.
• Adaptation techniques are crucial for developing effective ASR acoustic models for impaired speech.
• Existing research has not fully addressed optimal adaptation techniques or suitable source models for impaired speech.

Purpose of the Study:

• To identify the most effective adaptation technique for impaired speech.
• To determine suitable source models for building effective impaired-speech acoustic models.
• To investigate these issues specifically for dysarthria, a common speech impairment.

Main Methods:

• Applied Maximum Likelihood Linear Regression (MLLR) and Constrained-MLLR (C-MLLR) adaptation techniques.
• Utilized both unimpaired and impaired speech as source models.
• Measured recognition accuracy using Word Error Rate (WER), including phoneme insertion, substitution, and deletion rates.

Main Results:

• Combining unimpaired speech with limited impaired data improved ASR for severely dysarthric speech.
• C-MLLR outperformed MLLR for mildly and moderately impaired speech based on WER analysis.
• Phoneme substitution was the primary contributor to WER across all dysarthria severity levels.

Conclusions:

• Speech acoustic models developed with appropriate adaptation techniques enhance ASR performance for impaired speech, even with limited data.
• The choice of adaptation technique and source model significantly impacts ASR accuracy for dysarthric speakers.
• Further research into phoneme-level errors can refine ASR systems for speech-impaired populations.