作　　者： ; ; ; ;

机构地区： 深圳大学

出　　处： 《广西师范大学学报（自然科学版）》 2006年第4期26-29,共4页

摘　　要： 提出了一种新的基于高谱分辨率的谱减法,通常噪音帧和带噪语音帧具有相同的长度且都是短时帧,对短的噪音帧和带噪语音帧做傅立叶变换得到的谱分辨率低导致谱减法性能下降严重。基于平稳或接近于平稳的背景噪音下合并所有噪音帧再做傅立叶变换以提高噪音帧的谱分辨率;通过对带噪语音信号在时域进行前向和后向的延拓再做傅立叶变换以提高带噪语音信号的谱分辨率,在有效抑制噪音的同时减少了语音谱的失真并提高了系统识别率。 A new spectral subtraction algorithm with increased spectral resolution is proposed in this paper. Usually noise frames and noisy signal frames have the same length and are both short-time frames. The spectral resolution achieved by short-time Fourier transforms are very low and this seriously compromises the performance of spectral subtraction. With stationary or near-stationary background noise, the noise frame length can be increased until all the silence frames are covered. The FFT is then applied to the new noise frame so that the spectral resolution is more effective. By applying the FFT to the noisy signal frames after extrapolation of the noisy signal in time domain in both directions,a more accurate spectral resolution is obtained. The speech spectral distortion is much reduced and the noise effectively suppressed by this method. In this way, the recognition accuracy of system is enhanced.

关 键 词： 谱减法 线性预测 噪音估计 语音识别

领　　域： [电子电信] [电子电信]