Hierarchical deep belief networks based point process model for keywords spotting in continuous speech

Abstract

Summary Point process model keyword spotting (KWS) system has attracted considerable attentions in the areas of keyword spotting by its capacity that can generalize from a relatively small numbers of training examples. But unfortunately, the accuracy level of the point process model is not comparable with the state‐of‐the‐art KWS systems because of the poor modeling capacity of the phoneme detector, which are based on Gaussian Mixture Models. In this paper, focus on improving the performance of detector in point process model, we propose an enhanced version of point process model, which is based on hierarchical deep belief networks (DBNs). Hierarchical DBNs are used as the phoneme detector in this system, and they combine the advantages of both the DBN and the hierarchical architecture for capturing complex statistical patterns in speech while overcoming the inherent flaws of conventional hidden Markov models and multilayer layer perceptron. Experiments results on TIMIT database show that the proposed method can yield 2% improvement. Furthermore, in the case when training examples are extremely limited, it can achieve better results over state‐of‐the‐art KWS systems. Copyright © 2013 John Wiley & Sons, Ltd.

On Speech Recognition Algorithms

“…Deep Belief Networks (DBNs) are neural networks consisting of a stack of restricted Boltzmann machine (RBM) layers that are trained one at a time, in an unsupervised fashion to induce increasingly abstract representations of the inputs in subsequent layers [22]. The design of a DLN is basically a stack of RBM where each machine has two layers to it [24]. Each node contains "one layer of (typically Bernoulli) stochastic hidden units and one layer of (typically Bernoulli or Gaussian) stochastic visible units," such that each hidden layer is connected to every visible layer and each visible layer is connected to every hidden layer [24].…”

“…The design of a DLN is basically a stack of RBM where each machine has two layers to it [24]. Each node contains "one layer of (typically Bernoulli) stochastic hidden units and one layer of (typically Bernoulli or Gaussian) stochastic visible units," such that each hidden layer is connected to every visible layer and each visible layer is connected to every hidden layer [24]. However, no layer has connections within itself.…”

“…However, no layer has connections within itself. This layered structure uses a greedy learning strategy called unsupervised pre-training [24]. This strategy can theoretically "improve a variation lower bound on the likelihood of the training data under the composite model" [24].…”

“…This layered structure uses a greedy learning strategy called unsupervised pre-training [24]. This strategy can theoretically "improve a variation lower bound on the likelihood of the training data under the composite model" [24]. Once the unsupervised training has been completed, the supervised training begins which adjusts "resulting weights using gradient descent learning to improve the performances of DBN" [24].…”

“…This strategy can theoretically "improve a variation lower bound on the likelihood of the training data under the composite model" [24]. Once the unsupervised training has been completed, the supervised training begins which adjusts "resulting weights using gradient descent learning to improve the performances of DBN" [24]. The overall purpose of pre-training, which is unsupervised in the DBN, is to initialize the weights in the network to have a slightly better approximation than to have a random initialization [24].…”