New ways to use LVQ-codebooks together with hidden Markov models

We introduce a novel way to employ codebooks trained by learning vector quantization together with hidden Markov models. In previous work, LVQ-codebooks have been used as frame labelers. The resulting label stream has been modeled and decoded by discrete observation HMMs. We present a way to extract more information out of the LVQ stage. This is accomplished by modeling the class-wise quantization errors of LVQ by continuous density HMMs. Experiments in a speaker dependent phoneme spotting task verify that significant improvements are attainable over plain continuous density HMMs, or over the hybrid of LVQ and discrete HMMs