提出了一种新的非平稳噪声环境下的噪声功率谱估计方法。该方法通过采用非固定长度的时间窗跟踪含噪语音功率谱的最小值,解决了传统跟踪时延较大的问题。不同频带采用不同的阈值计算语音存在概率,从而利用语音存在概率值的大小调节噪声和语音的混合程度。实验证明,本文提出的方法较基于语音活性判决(voice activity detectors,VAD)的一系列方法和传统的最小统计(Minimal statistic,MS)算法有更好的效果,从而有效地改善了增强后语音的质量。
A new noise power spectrum estimation method is introduced for noise estimation in non-stationary noise environment. The noise estimation is realized by averaging past spectral power values and using a smoothing factor adjusted by the signal presence probability in every bins. Presence of speech is determined by the ratio between the energy of the smoothened noisy speech and its minimum. Experiments prove that the proposed method is superior to a series of exsiting methods based on voice activity detectors and the traditional minimal statistic algorithm.