LVQ as a feature transformation for HMMs

Presents a new way to take advantage of the discriminative power of learning vector quantization in combination with continuous density hidden Markov models. This is based on viewing LVQ as a non-linear feature transformation. Class-wise quantization errors of LVQ are modeled by continuous density HMMs, whereas the practice in the literature regarding LVQ/HMM hybrids is to use LVQ-codebooks as frame labelers and discrete observation HMMs to model a stream of such labels. As decision making at frame level is suboptimal for speech recognition, the presented method is able to preserve more information for the HMM stage. Experiments in both speaker dependent and speaker independent phoneme spotting tasks suggest that significant improvements are attainable over plain continuous density HMMs, or over the hybrid of LVQ and discrete HMMs