Using Articulatory Knowledge in Automatic Speech Recognition

Over the years different types of speech recognizers have been proposed and tested. During the last decade (or maybe even longer) hidden Markov models (HMMs) seem to have a better performance than other types of speech recognizers, like e.g. rule-based speech recognizers. This state of affairs has led to a gap between speech technology on the one hand, and phonetics and phonology on the other. Clearly, this is not an ideal situation, because both fields could and should benefit from each other. The important question then is: why is the performance of rule-based recognizers not as good as that of HMM-based recognizers? Probably this is due to a combination of several factors like: (1) at the moment, there isn’t enough knowledge available; (2) most of the knowledge available is derived from lab speech and cannot always be generalized to ther types of speech; (3) the kn owledge available is generally transformed into deterministic rules in rule-based systems, while HMM-based systems mainly use a kind of stochastic rules; (4) and finally, and probably most important, in many rule-based recognizers local decisions often have to be made (i.e. whether a segment is voiced or not), while in HMM-based systems one overall probabilistic decision is made.