一种基于正弦激励的线性预测模型的语音转换方法

Voice Conversion Based on Linear Prediction Model with Sinusoidal Excitation

在正弦激励模型的线性预测(LP)残差转换的基础上,提出了一种改进语音特征转换性能的语音转换方法。基于线性预测分析和综合的构架,该方法一方面通过谱包络估计声码器提取源说话人的线性预测编码(LPC)倒谱包络,并使用双线性变换函数实现倒谱包络的转换;另一方面由谐波正弦模型对线性预测残差信号建模和分解,采用基音频率变换将源说话人的残差信号转换为近似目标说话人的残差信号。最后由修正后的残差信号激励时变滤波器得到转换语音,滤波器参数通过转换得到的LPC倒谱包络实时更新。实验结果表明,该方法在主观和客观测试中都具有良好的结果,能有效地转换说话人声音特征,获得高相似度的转换语音。

By using a sinusoidal excitation method for voice spectral linear prediction （LP） residual transformation, an algorithm for the voice conversion technology is proposed to improve the target characteristics in the converted voice. The algorithm is based on the LP coding （LPC） analysis/synthesis framework and achieves LPC cepstral spectral envelope of the source speaker by the spectral envelope estimation vocoder （SEEVOC）. The spectral envelope is converted by the bilinear transform function. LP residual signals are modeled and decomposed by the harmonic sinusoidal model. Pitch modification is applied to the source speaker residual to approximate the target speaker pitch range. Then, the modified LP residual is used to excite the time varying filter. Filter parameters are updated according to the desired LPC cepstral spectral envelope. Experimental results indicate that the proposed method has a good performance in both objective and subjective tests and can convert the speaker personality with high similarity.