基于子带卡尔曼滤波的语音增强方法

Speech enhancement based on subband kalman filtering

与基于短时谱的语音增强方法相比,卡尔曼滤波的语音增强方法是基于语音生成模型的增强方法,这种基于模型的递推计算,导致卡尔曼滤波时的计算量很大。为了减少卡尔曼滤波的计算量,本文给出一种基于子带卡尔曼滤波的语音增强方法。先将带噪语音分解成子带信号,并通过子带频域谱减后估计低阶AR模型参数,然后利用卡尔曼滤波对子带信号进行滤波,最后由滤波后的子带信号重构全带语音信号,实现语音增强。实验表明该方法在提高语音质量的同时,通过子带分解降低了卡尔曼滤波的模型阶数,明显减少了语音增强系统的计算量,更容易实时实现。

Compared with the speech enhancement algorithm based on short speech spectrum, the kalman filtering is based on the speech produce model. The recursion operations based on the model introduce a big complex computation. To reduce the computation, this paper proposes a subband kalman filtering method. First, the noisy speech is decomposed into subbands, and then the subband noisy signal is subtracted in frequency domain and modeled by low order AR model. Second ,these subband signals are filtered by kalman filter. Last, the enhanced speech is reconstructed from these enhanced subband signal. This method improves the quality of the enhanced speech, and also largely reduces the computation complexity for the low orders of models in subbands, and then it can be easily realized on real time.