A new approach to discriminative HMM training for pathological voice classification

Area of Science:

• Medical acoustics
• Machine learning
• Speech processing

Background:

• Pathological voice identification is crucial for diagnosing voice disorders.
• Traditional Hidden Markov Models (HMMs) have limitations in accurately classifying pathological voices.
• Improving discriminative training criteria for HMMs can enhance classification performance.

Purpose of the Study:

• To develop and evaluate a novel discriminative training criterion for Hidden Markov Models (HMMs) tailored for pathological voice identification.
• To enhance the accuracy of voice disorder classification systems.

Main Methods:

• The proposed technique adjusts HMM parameters using Mahalanobis distance and the distance between probability density function means as objective functions.
• The Area Under the Curve (AUC) of the receiver operating characteristic (ROC) curve is maximized.
• The method was evaluated using the MEEIVL voice disorders database.

Main Results:

• The novel training criterion significantly improved classification accuracy compared to existing methods.
• The optimized HMMs demonstrated superior performance in distinguishing pathological voices.
• Results validated on the MEEIVL database.

Conclusions:

• The proposed discriminative training approach offers a significant advancement for pathological voice identification.
• This technique enhances the effectiveness of HMM-based classification systems for voice disorder detection.
• The method shows promise for clinical applications in audiology and speech pathology.