论文部分内容阅读
提出一种基于线性预测残差倒谱的基音周期检测算法.该算法对语音信号的线性预测残差信号做倒谱变换,将其作为基音检测特征.并综合残差倒谱峰、短时能量和短时过零率三种特征,构造一个清浊音判决函数,简化清浊音判决过程,提高判决精度.在基音周期检测过程中,根据基音连续原则,提出峰值重定位方法,有效降低基音倍频和半频的错误率.对比实验表明,本文算法的性能不仅较之传统的倒谱方法有明显改善,同时也优于目前效果较好的YIN算法和多尺度小波算法.
This paper proposes a pitch detection algorithm based on linear prediction residual cepstrum.The algorithm performs cepstrum transform on the linear prediction residual signal of speech signal as the pitch detection feature and combines the residual cepstrum, short-time energy And short-time zero-crossing rate, an unvoiced and voiced decision function is constructed to simplify the decision process of the voiced and unvoiced voices and improve the decision precision. In the process of the pitch cycle detection, the peak relocation method is proposed according to the principle of pitch continuous reduction, And half-frequency error rate.Comparison experiments show that the performance of the proposed algorithm is not only significantly improved compared with the traditional cepstrum method, but also better than the current YIN algorithm and multi-scale wavelet algorithm.