Comparison of two LP parametric representations in a neural network-based speech recognizer
Kuldip K. Paliwal · The Journal of the Acoustical Society of America · 1988
Although the different linear prediction (LP) parametric representations provide equivalent information about the short-time spectral envelope of speech, these representations are known to show differences in their speech recognition performance when used with conventional linear pattern classifiers. Recently, an error backpropagation algorithm has been reported in the literature for training the artificial neural networks, and it has been shown that the multilayer perception (MLP) classifiers that are nonlinear in nature can provide arbitrarily shaped decision surfaces in the multidimensional pattern space. The aim of the present paper is to see whether the different LP parametric representations show differences in their speech recognition performance for these nonlinear MLP classifiers, also. For this, the two-, three-, and four-layer perception classifiers are studied for the following two LP parametric representations: (1) the LP coefficient representation and (2) the cepstral coefficient representation. The results for the conventional linear pattern classifiers are also provided for the sake of completeness. It is shown that, like the conventional pattern classifiers, the MLP classifiers also result in better recognition performance for the cepstral coefficient representation than for the LP coefficient representation.