基于混合码书映射的高效语音转换方法

目前主流语音转换算法计算量大，复杂度高， 难以在内核小的嵌入式系统上运行。为了降低语音转换的计算复杂度，缩短训练时间，提出 一种基于混合码书映射的高效语音转换方法。在训练阶段,根据不同的参与训练的语音数据 量 建立不同的码书映射关系，节约训练时长，提高准确度。在转换阶段，系统依据训练阶段建 立的码书映射关系对浊音帧的声道参数进行转换。另外，为了提高转换语音的主观音质，系 统对清音帧的特征参数也作了相应转换，并且修正了转换语音的共振峰频率以克服帧间共振 峰抖动的问题。主客观测试结果表明：在保证转换音质的前提下，本文提出的语音转换方法 降低了计算复杂度、明显缩减了训练时间。

Efficient Voice Conversion Method Based on Mixture Mapping of Codebooks

The state of the art algorithms for voice conversion are computationally expensive and time consuming, thus they cannot be run in the embedded systems efficiently. An voice conversion method bas ed on mixture mapping of codebooks is proposed. In the training stage, different codebook mapping relationships are built according to the training speech amount, which saves training time and improves conver sion accuracy. In the transformation stage, the system converts the vocal tract parameters of voiced frames according to the corresponding codebook mapping buil t in the training stage. In addition, to improve the quality of the con verted speech, the system converts the feature parameters of unvoiced frames as well as correcting the formant frequency to overcome the formant jitters between frames. Both objective and subjective experiments show that the proposed method reduc es computational complexity and saves training time without degrading or deterio rating the quality of the converted speech.