Approach of features with confident weight for robust speech recognition

The enhancement of speech has become one of the focuses of automatic speech recognition (ASR) development. In recent studies, the missing feature approach (MFA) has been proved to be a suitable method. However the hard mask decision in the MFA is mostly a rough binary classifier on the basis of a certain threshold value that could cause a failed decision of reliability and result in a signal screening risk. As improvements of the hard mask the effectiveness of soft masks, including soft mask works with a Bayesian classifier, attempt to compensate the loss of real speech in the hard mask decision by discovering the probability density function (p.d.f.) of the unreliable feature component. Unfortunately, this is a very difficult task because of the overlap of at least two complex random processes. The sigmoid function suggested by some soft masks is not a reasonable p.d.f. In this paper, we provide an analysis of the confident degree of a feature component in a subband based on four criteria and then propose four types of confident weight (CWs). Based on CWs, we introduce four classes of approaches of feature with confident weight (AFCWs), which estimate the confidence degree of each feature vector simply and efficiently, describe the effect of noise in a rigorous manner, and eliminate the risk of selecting thresholds and the difficulty of finding a joint p.d.f. of reliable and unreliable components. Experimental results have shown that the proposed approaches improve the performances of ASR systems even in an adverse environment.