Confronting representative spectral structures of vocal tract pulse responses

Abstract

Before the resources of a number of ethnic language systems can be confronted, multiple linguistic procedures must be applied to identify the possible common acoustic features in the utterances of the language users. There is a clear difference in the articulatory habits characteristic of the particular languages, which is why significant similarities must be sought in the oblique fragments of speech that convey the information code (sense) rather than in the acoustic representation of free speech. The paper discusses methods and results of automatic classification of selected lexemes of three language systems by confronting their digital representations. Digital representation includes sets of acoustic parameters as discussed in the previous OSA papers, which the authors termed as: representative spectral structures of vocal tract pulse responses. Representative structures... were produced by cepstral smoothing of averaged acoustic parameters taken from multiple utterances by speakers of different genders and ages. The paper includes spectrograms of the material used for confrontation and graphic illustrations of the results. Automatic classification of speech sounds of a pair of lexemes of two different languages is in fact the process of identifying the percentage proportions of convergent and divergent energies of lines in a set of parallel elements of both spectral matrixes being confronted, with time and frequency coordinates.