利用麥克風陣列與目標干擾比之強健型語音活動偵測方法

Abstract

本論文提出麥克風陣列與目標干擾比(Target-to-Jammer, TJR)的語音活動特徵搭配混和高斯模型(Gaussian mixture model, GMM)與目標干擾比的語音活動特徵搭配最小控制遞迴平均法 ( Minima Controlled Recursive Averaging , MCRA )的兩種穩健型語音活動偵測方法。並且將此方法與長時間訊號變動程度 (Long-Term Signal Variability)和訊號能量做比較，在大部分的狀況下使用干擾比當語音活動偵測的正確率都高於其它特徵。當訊雜比越來越低的時候，目標干擾比(Target-to-Jammer, TJR)的優勢會越來越顯著。

In this thesis, two methods detecting voice activity by microphone array are proposed. The first method combines target-to-jammer ratio with minima controlled recursive averaging. The second method combines target-to-jammer ratio with Gaussian mixture model. These two methods are compared with signal energy method and long-term signal variability method. In most situations, the correct rate by using target-to-jammer ratio is higher than other features. When signal to noise ratio (SNR) gets lower, the target-to-jammer ration method will be more robust than using other features.