论文部分内容阅读
针对在连续丢包情况下声码器合成语音质量较差的问题,提出了一种特征参数的分模式线性预测技术。该方法利用参数的短时相关性,以子带清浊音参数为模式信息,计算各特征参数在不同模式下的预测系数,并根据获得的分模式预测系数用上一个正确接收帧的特征参数预测当前丢失帧的参数,最后用恢复的参数重建丢失语音帧。测试结果表明:当丢包长度的范围在75~200 ms时,与传统的抗丢包处理算法相比,该方法能够将合成语音的平均意见得分(mean opinion score,MOS)提高0.03左右。
Aiming at the poor voice quality of vocoder in case of continuous packet loss, this paper proposes a sub-mode linear prediction technique of feature parameters. The method uses the short-term correlation of the parameters and uses the sub-band unvoiced and unvoiced parameters as the model information to calculate the prediction coefficients of the various characteristic parameters in different modes and predicts the characteristic parameters of a correctly received frame according to the obtained sub-mode prediction coefficients The parameters of the current frame are lost, and finally the lost speech frame is reconstructed with the recovered parameters. The test results show that the proposed algorithm can improve the mean opinion score (MOS) of synthesized speech by around 0.03 when compared with the traditional anti-packet-loss processing algorithm when the packet loss length ranges from 75 ms to 200 ms.