Subspace-based speaker-independent vowel recognition

In this paper, we present a subspace-based approach for speaker-independent vowel recognition. Five vowels (/aa/,/eh/,/iy/,/ow/ and /uw/) from the TIMIT database were considered for the task. The subspaces representing two different vowel classes may have a large common subspace due to speaker variability, noise and coarticulation. We use common principal component (CPC) [1] and its extension i.e., partial-Common principal component (pCPC) to obtain a specific subspace for each vowel which is insensitive to variations. We perform CPC analysis on the covariance matrices of the vowels. pCPC gives q eigenvectors which are common to all vowels and (p - q) vowel specific eigenvectors. For each value of q, vowel specific subspaces are obtained. An input vector from an unknown vowel is classified based on the maximum length of its projection on the specific subspaces. We have choosen 18-dimensional Mel-Frequency Cepstral coefficients as a feature in our recognition task. The specific subspace is treated as a transformation matrix which enhances the vowel-specific information in the feature vector and, inturn. increases signal-to-noise ratio. Recognition experiments were performed on vowels extracted from a multiple speaker set taken from different dialect regions in the TIMIT database. Results for each vowel-specific subspace are presented for different values of q ranging from 1 to 5. The results are encouraging in the context of a speaker-independent framework. Visual Analysis of the vowel basis spectra provides useful and interesting information by highlighting the importance of different frequency regions.