Automatic Speech Recognition Method Based on Deep Learning Approaches for Uzbek Language.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Speech Technology

Background:

• Speech recognition systems predominantly focus on high-resource languages like English, neglecting low-resource languages.
• Uzbek language and its dialects lack comprehensive speech recognition models, limiting technological accessibility.
• Effective communication and globalization necessitate advancements in speech recognition for diverse linguistic backgrounds.

Purpose of the Study:

• To develop and evaluate advanced speech recognition models for the Uzbek language.
• To address the underrepresentation of low-resource languages in current speech recognition research.
• To improve the accuracy and efficiency of speech recognition systems for Uzbek.

Main Methods:

• Implementation of an End-To-End Deep Neural Network-Hidden Markov Model (DNN-HMM) for Uzbek speech recognition.
• Development of a hybrid Connectionist Temporal Classification (CTC)-attention network to enhance model training.
• Utilizing the CTC objective function within the attention model to optimize training time and accuracy.

Main Results:

• The proposed hybrid CTC-attention model achieved a word error rate (WER) of 14.3% on the Uzbek language dataset.
• The model demonstrated improved speech recognition accuracy and reduced training time compared to existing methods.
• Performance was evaluated using a newly collected Uzbek language dataset comprising 207 hours of recordings.

Conclusions:

• The developed speech recognition models show significant promise for the Uzbek language and its dialects.
• The hybrid CTC-attention approach offers an effective strategy for improving low-resource speech recognition systems.
• This research contributes to bridging the gap in speech technology accessibility for underrepresented languages.