A method for the asynchronous analysis of a voice source based on a two-Level autoregressive model of speech signal

We consider the problem of analysis of the voice source of speech within the range of short-term observations. The problem of insufficient speed of the available methods for the analysis of voice source is described, regardless of the method of data preparation: either synchronous with the main tone of speech sounds or asynchronous. We propose a method for the analysis of voice sources based on the two-level autoregressive model of the speech signal. We describe a software realization of the developed method based on the Berg-Levinson high-speed procedure of numerical calculations. It is shown that this procedure is characterized by a relatively low level of computation costs and its application does not require synchronization of the sequence of observations with the main tone of speech signal. With the help of software implementation of the proposed method, we designed and performed full-scale experiment aimed at analyzing the vowel sounds in the speech of a reference speaker. The results of this experiment confirmed the elevated speed of the proposed method and enabled us to formulate the requirements to the duration of speech signal for the real-time voice analysis. Thus, the optimal duration of the speech signal should vary within the range 32–128  msec. The obtained results can be used for the development and investigation of digital speech communication systems, systems of voice control, biometrics, biomedicine and other speech systems in which specific voice features of speaker’s speech are of primary importance.