CAs-Net: A Channel-Aware Speech Network for Uyghur Speech Recognition

Area of Science:

• Speech Recognition
• Artificial Intelligence
• Signal Processing

Background:

• Low-resource speech recognition faces challenges with limited data and noisy conditions.
• Existing models struggle to effectively process complex acoustic environments and linguistic nuances.

Purpose of the Study:

• To develop an advanced speech recognition model, CAs-Net, for improved performance in low-resource and noisy scenarios.
• To enhance the contextual modeling and temporal pattern recognition capabilities of speech recognition systems.

Main Methods:

• Proposed Channel-Aware Speech Network (CAs-Net) with a Channel Rotation Module (CIM) and Multi-Scale Depthwise Convolution Module (MSDCM).
• CIM reconstructs channel vectors for spatial structure modeling; MSDCM captures multi-scale temporal patterns within a Transformer framework.
• Utilized multi-branch depthwise separable convolutions and a lightweight self-attention mechanism.

Main Results:

• CAs-Net achieved the best performance on a Uyghur speech recognition dataset.
• Demonstrated a significant reduction in Word Error Rate (WER), reaching an average of 5.72%.
• Outperformed existing speech recognition approaches under challenging conditions.

Conclusions:

• CAs-Net proves effective and robust for low-resource speech recognition in noisy environments.
• The proposed modules enhance contextual understanding and temporal feature extraction.
• The model shows significant potential for improving speech recognition for underrepresented languages.