基于Gauss混合模型的清浊音恢复改进算法

Improved recovery algorithm for unvoiced/voiced parameters based on GMM

为提高子带清浊音(unvoiced/voiced,U/V)解码端恢复算法在不同能量电平下的鲁棒性,提出了一种改进型能量自适应U/V参数解码端恢复算法。通过跟踪长时能量的变化轨迹,在Gauss混合模型(Gaussian mixed model,GMM)下,用归一化的能量参数和线谱频率参数(line spec-tral frequency,LSF)对U/V参数的分布特性进行估计。测试结果表明:在较低的能量电平下,与用绝对能量对U/V参数进行恢复的算法相比,该能量自适应U/V参数恢复算法能够将清浊音误判率降低10%～25%,并将合成语音的平均意见得分(mean opinion score,MOS)提高0.03～0.09,改善了算法的性能。

The robustness of an unvoiced/voiced （U/V） speech classification recovery algorithm is improved by an energy self-adaption algorithm for the recovery of the U/V parameter. The algorithm traces the long-time changes of the energy level to estimate the statistical distribution of the U/V parameter from the normalized energy and the line spectral frequency （LSF） parameters based on the Gaussian mixed model （GMM）. Tests show that for relatively low energy levels, this energy self-adaption algorithm reduces the U/V classification error rate by 10% - 25% and improves the mean opinion score （MOS） of the synthesized speech signal by about 0.03 - 0.09 compared to the original method which uses the absolute energy value.