Adaptation of HMM dynamic parameters in reverberant environment

This paper presents a new adaptation method for HMM-based automatic speech recognition system in a reverberant environment. Unlike the conventional approach that estimates dynamic mean vectors by adopting a spline interpolation technique, the proposed algorithm uses the transform derived by the mathematical property. Additionally, we introduce the adaptation for covariance matrices with the domain conversion process induced by log-normal distribution, because the statistical parameters are affected by not only mean vectors but also covariance matrices. Consequently, all statistical parameters in HMM can be adapted by the exact same transform structure. Experimental results show that the proposed method improves the recognition rate, in spite of having much simple adaptation process. Also it is robust to the estimation error that is unavoidable while extracting the reverberation time related parameters.