A speech recognition model with time retrenching and dual recognition

Abstract

In this paper, We Propose a new speech recognition model with time retrenching and dual recognition. This model needs not use the endpoint detection and dynamic time warping (DTW) algorithms. Based on the properties of chinese speech, the contacts between consonant and vowel of each word are found out first, and then the sequence of word are retrenched in the time domain. By this approach, not only the endpoint detection and DTW algorithms can be omitted, but the redundance of speech signals can be reduced significantly, so that the computation time and memory capacity can be reduced considerably. In addition, for the purpose of large vocabulary recognition, a dual recognition method is proposed in this paper. The short-time average zerocrossing rate is defined as coarse feature parameter, and the LPC coefficients are referred to the fine feature parameters and they are used to constitute coarse and fine templates, respectively. It is shown by experiments that this model is successful and efficient. The recognition rate is above 94 percent for isolated word and is above 86 percent for connected speech. The percentage of retrenching ratio is about 40 percent.