Inverse problem of speech recognition

Gutenkov Roman Leonidovich

This work is devoted to the inverse problem of finding a segment of speech in a signal. The original signal is guaranteed to contain speech, but it is not known which codec it was encoded with. The main goal is to form a method for determining the codec and pre-processing parameters, such as inversion, byte and frame reversal, with which the signal was originally encoded. Several commonly used speech activity detectors are considered, namely the method for calculating the energy in the signal section and the spectral method. Based on them, an assumption was made that one of the speech activity detectors should be used to determine the codec, since the signal immediately begins with speech. Using one of the detectors and entering a numerical estimate, you can determine which codec and which pre-processing parameters need to be used in order to correctly decode the signal. After that, it will be necessary to test the method on different signals. As a result, the paper formulated a general approach for determining the required codec, provided that the signal is guaranteed to contain speech.

Keywords:speech recognition, speech recognition methods, Fourier transform, spectrograms, speech activity detector.