Research on optimal deep learning modeling in HaiNan dialect recognition

Area of Science:

• Computational Linguistics
• Artificial Intelligence
• Speech Processing

Background:

• HaiNan dialect speech recognition faces challenges due to significant phonological, intonational, and grammatical variations.
• Current research is limited by insufficient corpus resources, particularly for multi-dialect recognition.
• Traditional models struggle with data scarcity and diverse dialectal characteristics.

Purpose of the Study:

• To explore deep learning models for HaiNan dialect-to-Mandarin conversion.
• To identify the optimal deep learning model for HaiNan dialect recognition.
• To propose an effective fusion model for multi-dialect recognition.

Main Methods:

• Application of multiple deep learning models.
• Extensive experimental comparative analysis.
• Development of a fusion model combining Convolutional Neural Networks (CNNs) and Multi-Head Self-Attention Mechanism (MHAN).

Main Results:

• The proposed fusion model, ConvMHANet, demonstrates excellent performance across different dialectal scenes.
• ConvMHANet effectively addresses complex dialectal phonetic and contextual dependencies.
• Achieved 97.58% accuracy and a character error rate of 0.0163 in a multi-classification mixing task.

Conclusions:

• ConvMHANet shows strong generalization capabilities for HaiNan dialect speech recognition.
• The fusion model offers a promising solution for dialect-to-Mandarin conversion tasks.
• This research contributes to advancing speech recognition for low-resource dialects.