Authors
Navdeep Jaitly, Patrick Nguyen, Andrew Senior, Vincent Vanhoucke
Publication date
2012/3/12
Publisher
Tech. Rep. 001, Department of Computer Science, University of Toronto
Description
The ANN/HMM hybrid model was first used for automatic speech recognition (ASR) over two
decades ago [Morgan and Bourlard, 1990]. This model computes generative emission
probabilities for acoustic data from the states of an HMM using Bayes rule to invert
discriminative probabilities from a neural network trained to predict posterior states of the
HMM from acoustic data. In spite of their early promise, ANN/HMM hybrids were eventually
overtaken by GMM/HMM systems because of several factors which led to the superior ...
Total citations
2012201320142015201635472