Improved robust speech recognition considering signal correlation approximated by taylor series

In this paper, an improved mismatch function by considering signal correlation between speech and noise is proposed to better estimate the noisy speech HMM’s. A linearized model based on Taylor series expansion approach is used to approximate the proposed mismatch function. The parameters of the noisy speech HMM’s can be estimated more precisely by combining the parameters of the clean speech and noise HMM’s in the log-spectral domain or cepstral domain. Experimental results show that improved robustness for speech recognition in the presence of white noise as well as colored noise can be obtained.