Speech recognition picks the model maximizing the probability of observed features

Evidence

If an HMM is used for recognition rather than generation of observations, the problem is that only observations, typically the feature vectors analysed from a speech signal, are known. The joint probability that the known observation O = o1, o2, … oN is emitted by model M and state sequence X = 1 … 6 is the product P(O, X|M) = a12b2(o1) a22b2(o2) … A simple way to use HMMs in pattern recognition is to select the recognition result for the model M that gives the highest probability P(O, X|M) for the feature vector sequence O = oi. The name ‘hidden’ Markov model comes from the fact that the state sequence Xi is hidden, not known. — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 447

Categorization

• DEFINE: Clarifies why the model is called ‘hidden’: the underlying state sequence is never observed, only the emitted feature vectors are.