用于语音降噪的级联滤波器的设计与实现

Design and Implementation of Cascade Filter for Speech Denoising

将ETSI ES 202 050 V.1.1.3版本规定的Mel域上的两级维纳滤波结构推广到分数Mel域上,基于3周期的离散分数余弦变换,提出了分数Mel域上的两级维纳滤波结构,获得了性能的提高;针对以上两种方法的计算量较大和引入了时域截断噪声等缺点,提出了分数余弦变换域上的两级最佳线性滤波。理论分析和实验结果表明,该算法具有明显的降噪作用,能够应用于语音识别系统前端处理中。

Based on the three cycles discrete Fractional Cosine Transform, the two stages structures in the fractional Mel domain was proposed, which generalizes from two stages cascade Wiener structures in Mel domain defined in ETSI ES 202 050 V. 1.1.3 version. To reduce the large amount of calculation and truncation noises of the two kinds of methods, the two-stage optimal filter based on Fractional Cosine Transform was proposed. The analysis of theory and experiment results showed that the algorithms can obviously decrease the noises, which can be applied in the front-end processing of speech reorganization system.