Lecture 5: Clustering and Adaptation

The state index should encode all context information that might influence the acoustics: the third state of the /t/ in “a tree” should be different from the third state of the /t/ in “a tip,” because they are acoustically different. Likewise, lexical stress, phrase position, glottalization, and dialect might matter. Unfortunately, we never have training examples sufficient to learn the likelihood function associated with every possible combination of context variables. Therefore, we use a three-step strategy: (1) learn the mean and variance of MFCCs associated with every combination of context variables available in the training database, (2) cluster the context vectors, using acoustic similarity as a metric, ensuring that we have an adequate number of training examples from each phone, (3) learn a complete mixture Gaussian PDF for each clustered context-dependent phone. There are two ways in which similarity-based clustering can be performed in HTK: