FEATURES OF USING EXTENDED FORMS OF THE TRANSFORMER MODEL IN KAZAKH SPEECH RECOGNITION

In recent years, speech recognition technologies have significantly advanced due to artificial intelligence and machine learning methods. These technologies enable the automatic understanding of human speech and its conversion into text. The field of speech recognition, driven by machine learning and deep learning techniques, has garnered substantial interest from researchers worldwide. Both domestic and international scientists have conducted studies on Kazakh speech recognition, testing various models and proposing solutions for automatic recognition. Despite these advancements, Kazakh speech recognition still faces challenges due to the limited availability of large-scale datasets and the morphological and syntactic complexity of the Kazakh language. Consequently, improving recognition accuracy remains a relevant research area. This study provides a comprehensive review of the automatic speech recognition (ASR) models applied to Kazakh speech and their performance outcomes. Key challenges in the current Kazakh speech recognition process were identified, and flexible models for addressing these challenges were analyzed. The study examines the CHiformer model in speech recognition and evaluates the performance of Transformer, Hiformer, and Conformer models trained on the 400-hour KazASR Kazakh acoustic-text corpus. The results indicate that the Hiformer model achieved the highest accuracy in recognizing Kazakh speech, demonstrating its suitability for the linguistic characteristics of the Kazakh language. Specifically, for the 400-hour corpus, the word error rate (WER) of the Conformer model was reduced to 12.9%, while the WER of the Hiformer model further decreased to 11.6%