Maximum Likelihood Linear Regression (MLLR) for ASR Severity Based Adaptation to Help Dysarthric Speakers

Automatic speech recognition (ASR) for dysarthric speakers is one of the most challenging research areas. The lack of corpus for dysarthric speakers makes it even more difficult. The speaker adaptation (SA) is an alternative solution to overcome the lack of dysarthric speech and enhance the performance of ASR. This paper introduces the Severity-based adaptation, using small amount of speech data, in which data from all participants in a given severity type will use for adaptation of that type. The adaptation is performed for two types of acoustic models, which are the Controlled Acoustic Model (CAM) developed using rich phonetic corpus, and Dysarthric Acoustic Model (DAM) that includes speech collected from dysarthric speakers suffering from variety level of severity. This paper compares two adaptation techniques for building ASR systems for dysarthric speakers, which are Maximum Likelihood Linear Regression (MLLR) and Constrained Maximum Likelihood Linear Regression (CMLLR). The result shows that the Word Recognition Accuracy (WRA) for the CAM outperformed DAM for both the Speaker Independent (SI) and Speaker Adaptation (SA). On the other hand, it was found that MLLR is outperformed the CMLLR for both Controlled Speaker Adaptation (CSA) and Dysarthric Speaker Adaptation (DSA). Keywords-Component, Automatic Speech Recognition, Dysarthric Speakers, Severity Adaptation, Maximum Likelihood Linear Regression, Constrained Maximum Likelihood Linear Regression.