Environment normalization for robust speech recognition using direct cepstral comparison

In this paper we describe and evaluate a series of new algorithms that compensate for the effects of unknown acoustical environments (or changes in environment) through the use of compensation vectors that are added to the cepstral representations of speech that is input to a speech recognition system. These compensation vectors are obtained from direct frame-by-frame comparisons of the cepstral representations of speech that is simultaneously recorded in the training environment and various testing environments, but the algorithms do not make use of such “stereo” speech data in analyzing speech from an unknown environment. In the proposed paper we will compare the improvement in recognition accuracy provided by the algorithms using common standard ARPA speech recognition corpora. For example, the normalization algorithm known as MFCDCN provided a 22% reduction in word error rate when compared to results obtained using cepstral mean normalization on the 1992 ARPA WSJ/CSR corpus, and a 56.6% reduction in error rate compared to baseline processing. A family of new algorithms, PDCN, which accomplish the environment normalization inside the decoder are described and evaluated in the same corpus. A substantial word error rate reduction, 66.8%, can be achieved by combining MFCDCN and PDCN in the system with cepstral mean normalization compared to baseline system.