Optimization of Korean Speech Synthesis Model Based on Neural Network

This article uses deep neural network (DNN) technology to perform parameter transformation on Korean speech in order to improve system performance. This article takes linear cepstral coefficients (LSF) as the core parameter, and provides a guarantee for the stability of the comprehensive filter based on its ordered arrangement characteristics. On this basis, a byte based mapping network is proposed, which effectively solves the problems of instability and training convergence in the parameter transformation process by introducing intra frame parameter normalization and step size adaptive learning strategies. The experimental results show that the recognition rate of this method reaches 98.3%; In terms of speech synthesis, the use of 48 dimensional LSF parameters and event vector transformation based speech quality evaluation indicators has significantly improved, with subjective evaluation results significantly higher than conventional spectral parameter direct transformation methods. The research results of this project indicate that the optimization algorithm based on parameter feature extraction and dynamic adjustment proposed in this project can significantly improve the naturalness and stability of Korean speech, providing a practical and feasible solution for further optimizing the comprehensive performance of Korean speech.