Some Applications in Speech Recognition

The speech signal is the result of a voluntary and well-controlled movement of the speech apparatus. The design of an adequate modeling of speech patterns has been a constant concern since the beginning of automatic speech recognition research. In template methods [Sakoe and Chiba, 1978], the acoustic variability modeling of vocabulary consisted in storing several references for the same lexical unit, or in deriving typical sequences of acoustic frames by resorting to some kind of averaging method. These solutions were rather inefficient and expensive, even though they can provide a viable solution for a variety of applications. The idea of the statistical modeling of the spectral properties of speech gave a new dimension to the problem. The underlying assumption for all statistical methods is that speech can be adequately characterized as a random process whose parameters can be estimated effectively. We consider that the intelligent movement of the vocal track controlled by a speaker uttering a word in a specific acoustic and prosodic context is a temporal sequence of random events emitted by a Markov source. We therefore adopt a Bayesian attitude and quantify the uncertainty about the occurrence of an event by a probability.