Method of voice source coding with data compression based on the linear prediction model

Within the framework of a dynamically developing direction of research in the field of acoustic measurements – analysis and evaluation of parameters of the excitation signal of acoustic oscillations in the vocal tract of a speaker – the problem of coding a voice source of speech with data compression based on a linear prediction model is considered. Using the criterion of minimum average the voice source power in the speech production process, the problem is reduced to real-time coding of the linear prediction error signal. A new method of voice coding has been developed: with clipping of the linear prediction error, which is not associated with computationally expensive procedures for measuring the initial phase and frequency of the fundamental tone of the speech signal. An example of its technical implementation in soft real-time mode is considered. A full-scale experiment was set up and carried out, during which a comparative analysis of the effectiveness of the proposed method and the widely used discrete cosine transform method was performed. It is shown that due to the weakening of data compression artifacts in the reconstructed speech signal, the accuracy of coding the voice source using the developed method is one and a half to two times higher, and there is no need to detect vowel sounds of speech and pauses in the speech signal. The obtained results will be useful in the development of new and modernization of existing systems and algorithms in the fields of automatic speech processing and synthesis, mobile speech communication, artificial intelligence and other applications of speech technologies with data compression based on the linear prediction model.