A Study of Neighborhood Competing Models Based Verification Method

Utterance verification (UV) is an important portion in an intelligent speech recognition system, which role is determine if the input speech actual includes the word sound(s). In this study, we address the UV problem in the model neighborhood information viewpoint. We present a new robust verification method which can enhance the capability of UV in noise or other mismatch conditions by using the neighboring competing models information. Comparing with tradition likelihood ratio test (LRT) and online garbage model methods, experimental results show, the performance of proposed method is comparable to the LRT method in clean speech conditions, but explicitly outperforms other verification approaches in the noise speech conditions.