Bayes Rule and Speech Recognition
Bayes Rule and Speech Recognition
https://pdfs.semanticscholar.org/5ba5/7ff3c3e6e319586b86a990b6e082f4ecf972.pdf
In this paper, Wantanabe dives into the application of Bayesian theories to speech recognition – in particular, the Bayesian Information Criterion and the Variational Baysian Estimation and Clustering for Speech Recognition. Using the Bayesian approach allows researchers to increase the ability for generalization in speech classification. The procedure is as follows : First, The raw speech waveform is parameterized to observation vectors, which is then pattern matched to words. Using Bayes rule, this word prediction is modeled by
which calculates the probability given the prior probability of the word sequence (denoted as evidence). For example, “This morning, I had _____” would yield a higher prediction for the word breakfast, for example, compared to dinner based on the frequency of occurrence in the English language.
This relates to the Bayes Theorem covered in the textbook, introduced as a model of decision making under uncertainty. The textbook’s coverage of Bayes Rule talks about conditional probability or the chances of one event happening given that another event has already occurred. Graphically, when studying the conditional probability of A given B, this is represented by assuming we are in in B’s sample space, the fraction of the area B occupied by the intersection of A and B. We get a similar formula,Pr [A | B] = Pr [A] · Pr [B | A]/ Pr [B] . This links to the speech recognition where A= the word, and B is the evidence. Bayes rule is said to be the fundamental theory of Speech Recognition.