一种基于码书映射的高效语音转换方法

A Highly Efficient Voice Conversion Method Based on Codebook Mapping

为了使机器人在人-机语音交互过程中更为自然,利用语音转换技术改变源语音个性特征(机械音),进而变化为自然的目标人语音,是一种可行的方案。然而,当前的语音转换主流方法在实时性要求高且内核小的嵌入式机器人中并不适用。引入一种高效的改进型码书转换方法。该方法首先通过匹配线性谱频率参数的相对距离来求取加权系数,进而实现码字的预测重构;其次,对预测的码字进行带宽修正,克服频谱偏移问题。实验结果表明:该方法相比较传统方法,在转换性能相当的条件下,运行时间缩短75%左右。

In human -robot interaction, it is desired to have synthetic voices which sound natural and can be personalized for each user. One solution is to use voice conversion, in which the characteris- tics of a source mechanical voice are changed to produce a sound corresponding to a given target natural voice. However, the popular voice conversion method is computationally intensive, and not suitable for application in a robot with small kernel embedded. This paper introduces a high efficient improved segmental codebook conversion method. It firstly calculates the weighting coefficient by matching the relative distance of the Line Spectral Frequency （LSF） parameters to realize the prediction refactoring of code word. Secondly, the bandwidth correction for the predicted code word is used to solve the problem of spectrum shift. The test results show that the method is approximately 75% faster than the traditional Gaussian Mixture Model（GMM） under the comparative conversion performance.