Robust F0 estimation based on time-varying complex speech analysis and its application for IP telephony and musical signal

语音F0估计在语音处理中起着重要的作用。一位作者提出了分析语音信号的时变复语音分析，并在时变复语音分析的基础上提出了新的F0估计方法，其中通过语音分析估计复残差，通过残差的相应AMDF的倒数加权自相关取峰来估计F0。它被称为基于框架的方法。在本研究中，我们提出了更准确和快速的F0估计算法。该方法采用基于时变复分析的F0和F1轮廓估计方法，实现F0和F1的预选择。它被称为基于样本的方法。F0的最终选择采用基于帧的方法，在估计F0和F1的基础上缩短F0的范围。在两级F0估计中，更短的范围使估计更准确，计算量更小。此外，为了研究时变分析的有效性，基于框架的方法对每一帧进行了评估，并根据声音强度分为4种模式。实验结果表明，时变分析对强浊音帧具有较好的效果。

F0 estimation of speech plays an important role on speech processing. One of the authors has proposed time-varying complex speech analysis for analytic speech signal and has proposed novel F0 estimation based on the time-varying complex speech analysis in which complex residual is estimated by the speech analysis and F0 is estimated by peak-picking the weighted auto-correlation with a reciprocal of the corresponding AMDF for the residual. It is called frame-based method. In this study, we propose more accurate and fast F0 estimation algorithm. In the method, pre-selections of F0 and F1 are realized by using F0 and F1 contour estimation method based on the time-varying complex analysis. It is called sample-based method. Final-selection of F0 is realized by the frame-based method with shorten range of F0 based on the estimated F0 and F1. The shorten range results in more accurate estimation with smaller computational amount in the two-stage F0 estimation. Furthermore, in order to investigate the efficacy of the time-varying analysis, the frame-based method is evaluated for each frame that is categorized into 4 modes with respect to the voiced strength. The experimental results confirm that the time-varying analysis can perform better for strong voiced frames.