论文部分内容阅读
为了解决低速声码器合成语音的偶发性嘶哑或变调问题,对参数提取进行改善,采用有监督学习的Fisher判决法,利用多个特征值组成的特征向量为判据;基音周期平滑的准确度在利用了更准确的清浊音信息后大有提高。测试结果表明:该算法能够大大降低清浊音误判率,减少严重基音周期错误数;应用该算法的SELP(sinuous excitationlinear prediction)2.4 kb/s的PESQ-MOS分优于2.4 kb/s的MELPe(mixed excitation linear prediction)和AMBE+(advanced multi-band excitation)算法,DRT(diagnosticrhythm test)分数达95%,具有良好的可懂度和自然度。
In order to solve the problem of occasional hoarseness or pitch change of low-speed vocoder speech synthesis, the parameter extraction is improved. Fisher discriminant method with supervised learning is used as the criterion based on eigenvectors composed of multiple eigenvalues. The precision of pitch smoothing In the use of more accurate voiced sound information greatly improved. The experimental results show that this algorithm can greatly reduce the misjudgment rate of voiced and unvoiced voices and reduce the number of serious pitch errors. The 2.4 kb / s PESQ-MOS with SELP (sinuous excitation linear prediction) is better than 2.4 kb / s MELPe mixed excitation linear prediction and advanced multi-band excitation (AMBE +) algorithm, DRT (diagnosticrhythm test) score of 95%, with good intelligibility and naturalness.