Improving Acoustic Models in TORGO Dysarthric Speech Database.

Area of Science:

• Speech Technology
• Biomedical Engineering
• Computational Linguistics

Background:

• Dysarthria, a motor speech disorder, significantly impacts communication and quality of life.
• Existing automatic speech recognition (ASR) systems struggle with dysarthric speech due to its variability.
• The TORGO database provides a valuable resource for developing and evaluating ASR systems for dysarthria.

Purpose of the Study:

• To significantly improve ASR performance for dysarthric speech using the TORGO database.
• To explore and implement advanced techniques for enhancing Gaussian mixture model (GMM) and deep neural network (DNN) based hidden Markov model (HMM) ASR systems.
• To establish new benchmarks for ASR accuracy on dysarthric speech datasets.

Main Methods:

• Trained speaker-specific acoustic models by optimizing parameters and using speaker-normalized cepstral features.
• Developed complex DNN-HMM models incorporating dropout and sequence-discrimination strategies.
• Applied a generalized distillation framework to integrate dysarthric speech characteristics into models trained on both dysarthric and normal speech.

Main Results:

• Achieved significant performance improvements compared to previous ASR attempts on the TORGO database.
• Demonstrated the effectiveness of speaker-specific models and advanced DNN-HMM architectures.
• Showcased substantial gains for severe and severe-moderate dysarthric speakers through knowledge distillation.

Conclusions:

• The proposed methods yield state-of-the-art recognition accuracies for the TORGO dysarthric speech database.
• Advanced ASR techniques, including DNN-HMMs and distillation, are highly effective for improving speech recognition in dysarthria.
• This research contributes to developing more effective assistive speech technologies for individuals with motor speech disorders.