Use of HMMs in isolated recognition next up previous
Next: Training Up: Hidden Markov Models Previous: gradient wrt observation probabilities

Use of HMMs in isolated recognition

Isolated recognition in general means recognition of speech based on any kind of isolated speech unit, which can be a word or a sub word or even a concatenation of words. However only isolated word recognition has direct practical applications, while the isolated recognition based on other two types of units has basically a theoretical value. Specially the sub word recognition in isolated mode can give a good indication about the continuous recognition based on the same techniques.

In a simple isolated speech unit recognition task, where the vocabulary contains N speech units, we can use the system depicted in fig.1.1.

Figure 1.1: A simple isolated speech unit recognizer

There are however many possible solutions to the task, because we have many choices of optimization criteria, and several algorithmic implementations even for a particular criterion. Out of those, a gradient based method with MMI criterion will be described below.

Of course the task is comprised of two sub tasks; namely the training and recognition.

Narada Warakagoda
Fri May 10 20:35:10 MET DST 1996

Home Page