Eliminating inter-speaker variability prior to discriminant transforms

Abstract

This paper shows the impact of speaker normalization techniques, such as vocal tract length normalization (VTLN) and speaker-adaptive training (SAT), prior to discriminant feature space transforms, such as LDA (linear discriminant analysis). We demonstrate that removing the inter-speaker variability by using speaker compensation methods results in improved discrimination as measured by the LDA eigenvalues and also in improved classification accuracy (as measured by the word error rate). Experimental results on the SPINE (speech in noisy environments) database indicate an improvement of up to 5% relative over the standard case where speaker adaptation (during testing and training) is applied after the LDA transform which is trained in a speaker independent manner. We conjecture that performing linear discriminant analysis in a canonical feature space (or speaker normalized space) is more effective than LDA in a speaker independent space because the eigenvectors carve a subspace of maximum intra-speaker phonetic separability whereas in the latter case this subspace is also defined by the inter-speaker variability. Indeed, we show that the more normalization is performed (first VTLN, then SAT), the higher the LDA eigenvalues become.