Optimization of a library of standards for speaker identification by cross-correlation portraits

The article discusses the application of a speech signal model in the form of a cross-correlation portrait to the problem of text-dependent speaker identification. These portraits are two-dimensional arrays consisting of sample values of the local correlation coefficients of two signals. Using the example of two speakers, it is shown that the features of a person’s voice appear in the portraits of his speech signals so that each speaker has his own unique portrait of his pronunciation of a speech command. The method of identifying a speaker based on portraits of his command utterances is based on this property. The method is based on comparing the team portraits of an “unknown” speaker (the speaker who needs to be identified and for whom his reference portraits are stored in the database) with predefined reference portraits for each class of speakers. The property of preserving the individuality of portraits against a background of fairly strong noise allows the method to be used in an environment of acoustic noise. The frequency of correct identification of speakers significantly depends on the choice of reference portraits. This raises the problem of choosing such utterances for each class of commands in which the portrait of the announcer’s command will be closest to all possible portraits of “your” announcer and most distant from the portraits of the commands of the “foreign” announcer. The paper proposes a method of directed enumeration, which allows one to select the most successful ones from the available set of utterances for use as reference ones. An experiment was conducted on real speech material, which proved the effectiveness of the method proposed in the work for optimizing the standard library.