关键词: 基音频率, 反向传播（BP）神经网络, 自相关函数, 平均幅度差函数, 清浊音分离

Abstract: Traditional pitch detection methods present poor voicing detection accuracy and robustness when the signal-to-noise ratio is low. To overcome these shortcomings, a two-layer neural network based pitch detection algorithm is proposed in this paper. The method utilizes the BP artificial neural network to conduct endpoint detection and then the second BP neural network to separate the clear sonant. At last, the dynamic test is adopted to get pitch frequency. The experimental results show that the proposed algorithm can decrease double and half frequency errors and improve the pitch detection accuracy compared with traditional autocorrelation function method.

Key words: pitch detection, Back Propagation（BP） neural network, autocorrelation function, average magnitude difference function, separation of unvoiced sounds and voiced sounds