CAJ | 학술논문

用线性预测的方法求出语音信号的LPC(Linear Predictive Coding)谱,然后根据候选的声门激励与LPC谱卷积重构语音信号的短时频谱,当重构频谱与原始语音频谱之间的畸变最小时,声门激励之间的间隔为基音周期.为了提高计算效率,采用频域动态搜索的方法搜索基音周期的候选值.数值实验表明,采用线性预测和极大似燃估计 (Maximum Likelihood, ML)的基音检测算法可保留更多的基音信息,并能有效地减少基音检测的错误,并且该算法比传统的ML法有更强的鲁棒性.
용선성예측적방법구출어음신호적LPC(Linear Predictive Coding)보,연후근거후선적성문격려여LPC보권적중구어음신호적단시빈보,당중구빈보여원시어음빈보지간적기변최소시,성문격려지간적간격위기음주기.위료제고계산효솔,채용빈역동태수색적방법수색기음주기적후선치.수치실험표명,채용선성예측화겁대사연고계 (Maximum Likelihood, ML)적기음검측산법가보류경다적기음신식,병능유효지감소기음검측적착오,병차해산법비전통적ML법유경강적로봉성.
A robust pitch detection algorithm based on LPC spectrum and ML method is proposed in the paper. Firstly, the LPC spectrum is obtained, and then the short time frequency spectrum is reconstructed based on the convolution of glottis and the LPC spectrum. When the aberration between the reconstructed spectrum and the original frequency spectrum reaches minimum, the interval between excitations of the glottis is the pitch. In order to improve the computational efficiency, the dynamic seeking algorithm is adopted to ensure the probable pitch values. Numerical experiment indicates that the proposed method of pitch detection is effective, more information of pitch is remained and the pitch detection error is reduced effectively. In addition, the method is more robust than the traditional method of ML.