Optimising Speaker-Dependent Feature Extraction Parameters to Improve Automatic Speech Recognition Performance for...

Area of Science:

• Speech processing
• Computational linguistics
• Biomedical engineering

Background:

• Automatic Speech Recognition (ASR) systems struggle with dysarthric speech.
• Standard ASR approaches are often ineffective for impaired speech.
• Dysarthria presents significant challenges in human-computer interaction.

Purpose of the Study:

• To validate a new speech analysis technique for improving ASR performance in speakers with dysarthria.
• To investigate correlations between speaker voice characteristics and optimal ASR parameters.
• To enhance the effectiveness of speaker-dependent ASR systems for impaired speech.

Main Methods:

• Fine-tuning spectral analysis window size and shift parameters for Short-Time Fourier Transform (STFT).
• Utilizing a speaker-dependent ASR system.
• Experimenting with Italian speech data from 30 dysarthric speakers (IDEA database) and 10 unimpaired speakers (CLIPS database).

Main Results:

• The proposed speech analysis technique significantly improves ASR performance for speakers with moderate to severe dysarthria.
• The new approach is ineffective for unimpaired or mildly impaired speech.
• A correlation was identified between specific speaker voice features and the optimal window/shift parameters for minimizing ASR errors.

Conclusions:

• The novel speech analysis method offers a viable solution for enhancing ASR in the context of dysarthria.
• Personalized parameter optimization based on voice features can further improve ASR accuracy for dysarthric speakers.
• This research contributes to more inclusive and effective speech technology for individuals with speech impairments.