Cepstrum third-order normalisation method for noisy speech recognition

Robust speech recognition using spatial–temporal feature distribution characteristics

“…For example, not content with using either CMN or CMVN merely to normalize the first or the first two moments of the probability distributions of speech features, some researchers have extended the principal idea of CMN and CMVN to the normalization of the third (Suk et al, 1999) or even more higher order moments of the probability distributions of speech features (Hsu and Lee, 2009). On the other hand, the histogram equalization (HEQ) techniques also have gained much attention, and have been widely investigated in recent years (Dharanipragada and Padmanabhan, 2000;Molau et al, 2003;Torre et al, 2005;Hilger and Ney, 2006;Lin et al, 2009).…”

Higher order cepstral moment normalization (HOCMN) for robust speech recognition

“…The row ( 4) is for CN with a segmental processing interval as mentioned previously with the segment length l being 86 frames, which gives better results. The next row (5) is for the third-order moment normalization (TMN) previously proposed [3] also with a full-utterance processing interval, in which the first, second, and the third order moments are all normalized. This approach is apparently even better.…”

“…When performing HOCMN with an even integer N, we simply scale the first-order moment normalized coefficients X [1] (n) by a constant, such that the N-th order moment of the coefficients can also be normalized as in equation (3). Such normalization therefore usually co-exists with the first-order normalization or CMS, and can be expressed as…”

“…The HOCMN for an odd order can be extended from the thirdorder cepstral moment normalization approach previously proposed [3]. It usually also co-exists with the first-order normalization or CMS, and can be expressed as a nonlinear transformation of the coefficients normalized with the first as well as the ( 1)…”

“…In addition, the variance normalization of CN may reduce the difference in probability density function (pdf) between the clean and noisy speech signals. It was also proposed that the normalization of the third-order cepstral moment may achieve better performance than CMS and CN [3], and with such normalization the pdf of noisy speech signals actually becomes even more closer to that of the clean ones. The pdf of cepstral coefficients of speech signals is usually regarded as a quasi-Gaussian distribution.…”