Phone Labeling Based on the Probabilistic Representation for Dysarthric Speech Recognition
Year: 2016
Type: article
Abstract: In this paper, we discuss speech recognition for persons with articulation disorders resulting from athetoid cerebral palsy. Because the speech style for a person with this type of articulation disorder is quite different from a physically unimpaired person, a conventional speaker-independent acoustic model for unimpaired persons is hardly useful to recognize it. Therefore, a speaker-dependent model for a person with an articulation disorder is necessary. In our previous work, a feature extraction method using a convolutional neural network was proposed for dealing with small local fluctuation of dysarthric speech, and its effectiveness was shown in a word recognition task. The neural network needs a training label (teaching signal) to train the network using back-propagation, and the previous method used results from forced alignment using HMMs as the training label. However, as the phoneme boundary of an utterance by a dysarthric speaker is ambiguous, it is difficult to obtain the correct alignment. If a wrong alignment is used, the network may be inadequately trained. Therefore, we propose a probabilistic phoneme labeling method using the Gaussian distribution. In contrast to the general approach, we deal with the phoneme label as the soft label, that is, our proposed label takes the continuous value. This approach is effective for the dysarthric speech which is the ambiguity of the phoneme boundary. The effectiveness of this method has been confirmed by comparing its effectiveness with that of forced alignment.
Institution: Kobe University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W2342540126