基于噪声倒谱阈值频谱估计的语音活动检测

Voice Activity Detection Based on Noise Cepstrum Thresholding Spectral Estimation

针对低方差频谱估计的语音活动检测(VAD)中Welch频谱估计方法计算量大的问题,提出利用倒谱阈值方法估计VAD中的噪声功率谱。该方法在静音时期为噪声的倒谱设置阈值,利用快速傅里叶变换计算频谱,再更新VAD中的判决阈值。算法复杂度分析与仿真结果表明,该方法的检测性能与Welch方法相当,计算量降低约18%,同时降低整个VAD的时间复杂度。

Aiming at high computational complexity in Voice Activity Detection（VAD） which uses low variance spectral estimation, this paper proposes a method using noise cepstrum thresholding spectral estimation to compute noise Power Spectral Density（PSD）. In speech absent period, a threshold is set for smoothing the noise cepstrum, then noise PSD is estimated by using Fast Fourier Transform（FFT）, and detecting threshold in VAD system is updated. Computational complexity analysis and simulation results indicate that compared with Welch method, the computation of the method is reduced by 18%, and total time complexity of VAD is decreased.