MALCOM X: Combining maximum likelihood continuity mapping with Gaussian mixture models

摘要

GMMs are among the best speaker recognition algorithms currently available. However, the GMM's estimate of the probability of the speech signal does not change if the authors randomly shuffle the temporal order of the feature vectors, even though the actual probability of observing the shuffled signal would be dramatically different--probably near zero. A potential way to improve the performance of GMMs is to incorporate temporal information into the estimate of the probability of the data. Doing so could improve speech recognition, speaker recognition, and potentially aid in detecting lies (abnormalities) in speech data. As described in other documents (Hogden, 1996), MALCOM is an algorithm that can be used to estimate the probability of a sequence of categorical data. MALCOM can also be applied to speech (and other real valued sequences) if windows of the speech are first categorized using a technique such as vector quantization (Gray, 1984). However, by quantizing the windows of speech, MALCOM ignores information about the within-category differences of the speech windows. Thus, MALCOM and GMMs complement each other: MALCOM is good at using sequence information whereas GMMs capture within- category differences better than the vector quantization typically used by MALCOM. An extension of MALCOM (MALCOM X) that can be used for estimating the probability of a speech sequence is described here.