改进的基于增量估计的快速高斯计算

Maximum probability increase estimation method for fast Gaussian likelihood computations

该文分析讨论了连续语音识别系统中的快速高斯计算问题。语音信号的短时平稳特性,使得相邻语音帧可能共享相似的分布。最大概率增量估计算法利用该特性,估计当前帧与基准帧间似然值增量的最大值,以减少似然值的精确计算量。该文针对该算法中增量上界被高估的问题,在增量上界平滑、最优G auss候选、风险因子设定等方面进行了改进。实验结果表明,在几乎不损失识别率的情况下,改进后的M P IE算法可节约40%的维数计算,解码速度相对提高10%。

This paper presents a fast Gaussian likelihood computational method for a continuous speech recognition system.Adjacent speech frames likely share similar distributions due to the semi-stationary feature of speech signals.Therefore,the maximum probability increase estimation(MPIE) algorithm can estimate the maximum probability increase between the current frame and a reference frame to reduce the explicit Gaussian likelihood computations.Overestimates of the maximum probability increase are reduced by increased smoothness,a best candidate Gauss factor and risk factors and so on.The tests show that the improved MPIE reduces the Gaussian likelihood computations by 40% and speeds-up the decoding process by approximately 10%,with almost no loss of recognition rate.