METHOD OF SELECTING SPEECH PROCESSING SEGMENTS BASED ON ANALYSIS OF CORRELATION DEPENDENCIES IN SPEECH SIGNAL Russian patent published in 2012 - IPC G10L19/00 

FIELD: information technology.

SUBSTANCE: in the method of selecting speech processing segments based on analysis of correlation dependencies in a speech signal at the output of an analogue-to-digital converter with sampling frequency of 8 kHz and 256 quantisation levels in the speech signal, 20 ms analysis sections are selected, the autocorrelation function of signal segment up to the first negative reading is calculated, after which the value of the correlation interval is calculated via approximation with a second-order polynomial of discrete values of the autocorrelation function. The analysis segment is then shifted by 40 ms through 2.5 ms. The autocorrelation function and the correlation interval are calculated once more. This operation is repeated, while forming a sequence of values of segment correlation intervals, after which said sequence is analysed using an F-criterion (Fischer), where the boundary of the segment or analysing the speech signal is formed the moment there is principal change in value of the correlation interval.

EFFECT: reduced volume of data for low-rate encoding of a speech signal, low average speed for transmitting speech over digital communication channels.

4 dwg