The preservation of regional dialects in China has become increasingly important due to the dominance of Mandarin, driven by urbanization and globalization, which threatens local languages and cultures. This paper addresses the challenge of dialect restoration by leveraging advanced AI models, including Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, to restore and maintain regional dialects in the digital era. The objective of this research is to develop a scalable, real-time framework for dialect restoration that ensures both linguistic accuracy and cultural preservation. The proposed method utilizes FastSpeech2 for text-to-speech synthesis and HiFi-GAN for high-fidelity speech generation, overcoming the limitations of traditional models. The framework also integrates Mel-Frequency Cepstral Coefficients (MFCCs) for feature extraction and the Arithmetic Optimization Algorithm (AOA) for efficient optimization, which improves both accuracy and processing speed. The results demonstrate the effectiveness of the proposed method, achieving an accuracy of 0.9812, precision of 0.9811, recall of 0.9808, and F1-score of 0.9805, indicating high performance in dialect classification and restoration. The Mean Opinion Score (MOS) for speech quality ranges from 4.52 to 4.72, with Mandarin achieving the highest score of 4.72. The Word Error Rate (WER) is between 1.58% and 2.45%, with Mandarin showing the lowest error rate of 1.58%. These results confirm the potential of the proposed framework in preserving regional dialects and ensuring their continued cultural significance.
Long Liu (Sun,) studied this question.