Development of Language Models for Continuous Uzbek Speech Recognition System.

Area of Science:

• Natural Language Processing
• Speech Recognition
• Computational Linguistics

Background:

• Large vocabulary automatic speech recognition and natural language processing applications require language models.
• Research on pre-trained language models predominantly focuses on high-resource languages, neglecting low-resource languages like Uzbek.
• There is a scarcity of publicly available Uzbek speech datasets, hindering the development of language models for this language.

Purpose of the Study:

• To develop a low-resource language model for the Uzbek language.
• To address the limitations in Uzbek natural language processing by creating a robust Uzbek language model.
• To investigate and understand linguistic occurrences specific to the Uzbek language.

Main Methods:

• Proposed UzLM, an Uzbek language model, by evaluating statistical and neural-network-based approaches.
• Developed an Uzbek-specific linguistic representation to enhance the robustness of the language model.
• Utilized a corpus of 80 million words, comprising approximately 68,000 unique words and 15 million sentences, for training.

Main Results:

• The developed UzLM leverages Uzbek-specific linguistic features for improved performance.
• The model was trained using a substantial corpus, demonstrating efficient use of training data compared to previous studies.
• Experimental results on continuous Uzbek speech recognition showed a reduction in character error rate to 5.26% when using neural-network-based models compared to manual encoding.

Conclusions:

• The development of UzLM represents a significant advancement for Uzbek natural language processing and speech recognition.
• The study highlights the effectiveness of neural-network-based language models for low-resource languages.
• UzLM provides a foundation for future research and applications in Uzbek speech technology.