COMPOSITIONAL RECOGNITION OF AZERBAIJANI SPEECH SIGNALS USING PHONETIC ACOUSTIC COMPONENTS

The paper proposes an approach to recognizing Azerbaijani speech signals based on structural comparison between phonetic acoustic components and segmented speech signal families. Unlike traditional methods, which rely on direct comparison of the signal to be recognized with words in the lexicon, the proposed approach performs recognition through the structural comparison of phonetic acoustic components corresponding to the letters of the Azerbaijani alphabet. The experiments use speech signals containing acoustic realizations of the selected Azerbaijani phonemes “Q”, “Ə”, “L”, “M”, and “K”. Segmentation is performed directly on the analyzed speech signals at various scales. The digitized Azerbaijani words “kitab”, “dəftər”, and “qələm”, voiced by a native speaker, are used as test speech signals. For each phoneme-speech signal pair, sequences of difference functions are calculated, reflecting the dynamics of the structural similarity between phonetic acoustic components and segmented speech signal representations. Experiments have shown that more detailed segmentation allows for the identification of stable local structural features of speech signals and improves the robustness of recognition. The obtained results confirm the feasibility of recognizing speech signals based on a limited set of segmented phonemic components without using a full dictionary of words. The proposed approach potentially reduces the search space and computational costs of speech recognition systems by shifting from comparing entire words to comparing phonetic structural components (pp.3-12).