Automatic phone segmentation and labeling of continuous speech

Abstract

To obtain an accurate phone sequence from a continuous speech signal, we suggest a novel approach consisting of tightly coupled bottom-up and top-down processing. The bottom-up path consists of segmentation, recognition and labeling. Also the top-down path consists of labeling, speech generation and segmentation. In this manner, the four processes form a closed feedback loop achieving an optimal interpretation efficiently for a given noisy observation of speech signal and a priori knowledge. The major goal of this paper is to identify the system model using both the stochastic estimation theory and the mean field theory. Experimental results are obtained in terms of the TIMIT database. It is shown that introducing the top-down path to the traditional bottom-up path can improve the recognition rate by 19.7%, and reduce the error (substitution, deletion and insertion) rate by 16.1%. As a result, the overall system can transform the incoming continuous signal into one of the 61 phone classes at the rate of 73.7%.