基于HMM状态聚类均值替代的宽带语音ISF参数补偿算法

ISF Parameters Concealment Algorithm Based on the HMM Clustering Means for Wideband Speech

宽带语音在Internet传输中不可避免会出现丢帧现象,由于错误传播的影响,使接收语音质量急剧下降。该文采取大型连续分布隐马尔可夫模型(LCDHMM)对宽带语音ISF参数建模,采用Viterbi算法确定丢失帧之前若干语音帧ISF参数观察值的最佳状态序列。由于状态的冗余度较大,用丢帧前最近接收的正确帧ISF参数的HMM状态对应的聚类均值和真实值的加权,代替丢失帧的ISF参数值。将采取该算法的补偿语音和采取G.722.2标准附件I所提算法的补偿语音进行比较,仿真结果表明该算法具有较好的补偿效果,其波形与谱失真更小。

There exist inevitably frame losses in the transmission of wideband speech. Due to the influence of error propagation, the quality of the received speech decays rapidly. In this paper, large hidden Markov model is adopted to model the ISF parameter in wideband speech, and the best state sequences of ISF parameters of the several speech frames before the lost frame are determined using Viterbi Algorithm. Because of the large redundancy of the HMM states, the lost ISF parameters are substituted by the weighted values between the clustering means and the real values of the ISF parameters of the nearest received frame. The speech compensated by this algorithm is compared to which by Annex I of G.722.2 specification. Simulation shows that the algorithm in this paper can result in best speech, and smaller waveform and spectrum distortion.