Improving Hybrid CTC/Attention Architecture for Agglutinative Language Speech Recognition

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• End-to-end (E2E) automatic speech recognition (ASR) models offer an alternative to traditional systems but require substantial training data.
• Hybrid CTC/attention ASR systems show promise for low-resource conditions but are underutilized for Central Asian languages like Turkish and Uzbek.

Purpose of the Study:

• To enhance the performance of E2E ASR systems for agglutinative languages.
• To address the data scarcity challenge in low-resource ASR scenarios.

Main Methods:

• Dataset augmentation through noise addition and speed perturbation.
• Proposed a novel multi-scale feature extraction method (MSPC) using varied convolution kernel sizes.
• Improved attention mechanisms and utilized CTC objective function with BERT-initialized language models for training and decoding.

Main Results:

• The MSPC feature extractor outperformed VGGnet.
• The proposed model demonstrated accelerated convergence and improved accuracy.
• On LibriSpeech, character error rate (CER) and word error rate (WER) increased by 2.42% and 2.96% respectively.
• Significant WER reductions of 7.07% and 7.08% were achieved on Common Voice-Turkish and Uzbek datasets, respectively.

Conclusions:

• The developed E2E ASR system shows competitive performance, approaching advanced systems.
• The proposed methods are effective in improving ASR for low-resource agglutinative languages.