论文部分内容阅读
语音存在概率的估计是语音增强的核心技术之一,针对传统的存在概率估计方法是启发式的,没有把存在概率的估计统一到一个理论框架之中,不能保证估计最优,提出了一种基于序贯隐马尔可夫模型(SHMM)的存在概率估计方法,在每一子带上构建一个SHMM模型描述对数功率谱包络的时间序列,把谱包络序列看作一个在语音和噪声状态之间转移的动态一阶马尔可夫链,采用单高斯函数构建每一状态的概率模型,语音状态的后验概率即为语音信号的存在概率。为了满足算法实时性要求,SHMM参数估计简化为一阶回归过程,根据极大似然准则逐帧更新模型参数。实验表明:SHMM所描述的时序相关性对存在概率的估计起到关键作用,它优于一般的启发式估计方法;SHMM算法的语音增强分段信噪比(SegSNR)和对数谱失真(LSD)性能优于经典的改进型最小统计量控制递归平均(IMCRA)算法。
The speech existence probability estimation is one of the core technologies of speech enhancement. In view of the traditional method of existence probability estimation being heuristic, the estimation of the existence probability is not unified into a theoretical framework and the estimation is not guaranteed, and a Based on the existence probability estimation method of sequential hidden Markov model (SHMM), a SHMM model is constructed on each subband to describe the time series of the logarithmic power spectrum envelope, and the spectral envelope sequence is regarded as a speech signal and noise The dynamic first-order Markov chain which is transferred between states, uses the single Gaussian function to construct the probability model of each state, and the posterior probability of the speech state is the existence probability of the speech signal. In order to meet the requirement of real-time algorithm, the SHMM parameter estimation is reduced to a first-order regression process, and the model parameters are updated frame by frame according to the maximum likelihood criterion. Experiments show that the sequence correlation described by SHMM plays a key role in estimating the existence probability, which is superior to the general heuristic estimation method. The Segmental Signal-to-Noise Ratio (SegSNR) and Log Spectral Distortion (LSD) ) Outperforms the classical Modified Minimum Statistic Control Recursive Mean (IMCRA) algorithm.