METHOD FOR EXTRACTING SPEECH PROCESSING SEGMENTS BASED ON SEQUENTIAL STATISTICAL ANALYSIS Russian patent published in 2019 - IPC G10L19/22 

FIELD: calculating; counting.

SUBSTANCE: invention relates to digital communication and can be used in telecommunication information systems with low-speed coding of a speech signal. At the output of the analogue-to-digital converter with a sampling frequency of 8 kHz and 256 quantisation levels, the instant read sequence is subjected to buffering and segmentation, with selection of the analysis portion of 20 ms with subsequent increase thereof and checking for "homogeneity". Initial boundary of the analysis portion is stored in the memory buffer. Parzen kernel smoothening method is used to estimate a curve of probability density of the initial section of analysis. Number of smoothing windows is determined using the Akaike information criterion. Data on distribution curve of instantaneous values of initial section of analysis are stored in memory buffer. First section of analysis is obtained by increasing the length of the analysis segment. Initial likelihood logarithm for the first analysis segment is calculated from density curve data for the initial segment. Curve of probability density of the first analysis site is evaluated. Obtained data are stored in a memory buffer and the first value of the log likelihood logarithm is calculated.

EFFECT: technical result consists in reduction of data volume at low-speed coding of speech signal.

1 cl, 5 dwg, 1 tbl