Image Approach to Uzbek Speech Recognition

This paper proposes the method to extract the features of speech signal from its spectrograms by applying a two-dimensional discrete-cosine transform over them. The calculated informative features are classified by three machine learning algorithms with a teacher, i.e. KNN, RF and SVM. As part of the scientific research, a database of Uzbek syllables, words and sentences is formed. Experiments have shown that when the proposed method is applied in combination with KNN classification method, the recognition accuracy of syllables, words and sentences together averaged 98,6%, thus outperforming RF (91,2%) and SVM (94,5%) algorithms.