Volume 4, Issue 8, 2014

By | August 11, 2018

Probabilistic PCA Mixture under Variance Preservation

Mohamed Nour I. Ismail, College of Science, King Faisal University, Saudi Arabia.
Mohamed El-Hafiz Mustafa Muse, College of Computer Science and IT, Sudan University of Science and Technology, Sudan.

Abstract— modeling data heterogeneity by a mixture of local models and exploiting the correlation in the localized data subsets to reduce their subspace dimensionalities has been realized in many mixture models; like PCA mixture and FA mixture models. Determining the number of local models as well as the proper dimensionality for each subspace (local model space) are the most difficult questions of these models. Instead of using fixed ad-hoc dimensionality for all local models, this paper proposes using a global preserved variance percentage value to estimate the dimensionality that retains the given variability percentage in each subspace. We test the proposed method on classifying handwritten digit by a mixture of Probabilistic PCA model, the result shows that the proposed method outperforms fixed dimensionality probabilistic PCA mixture model.

Keywords-component; PCA; MPPCA; Gaussian mixtures; EM algorithm.

Adaptive Activation Function for Isolated Digit Recognition Based on Speaker Dependent System

Ummu Salmah, M.H, Department of Information Technology, Kolej Universiti Islam Sultan Azlan Shah, Malaysia.
Siti Mariyam, S., Department of Computer Graphics and Multimedia, Universiti Teknologi Malaysia, Malaysia.
Saira Banu, O.K., Computing Department, Universiti Pendidikan Sultan Idris, Malaysia.
Nor Azah, Center of Language for Foundational Studies (CELFOS), Kolej Universiti Islam Sultan Azlan Shah, Malaysia.

Abstract— An automatic speech recognition (ASR) system has been the goal in speech research for more than 6 decades. This study focuses on developing the robustness of the MLP neural network for the Malay isolated digit recognition system by proposing a simple novel approach. An adaptive sigmoid function is implemented to achieve this objective. A typical or fixed sigmoid function method is used in the learning phase. In the recognition phase, an adaptive sigmoid function is employed. In this sense, the slope of the activation function is adjusted to gain highest recognition rate. The outcome of the simulation reveals that adaptive sigmoidal function offers a number of advantages over traditional fixed sigmoid function, resulting in better generalization performance. The proposed approach implicates ASR is applicable for the task on Malay language continuous speech and the speaker independent task to fulfill the ultimate goal in speech technology, towards natural ASR.

Keywords- Automatic Speech Recognition; Multilayer Perceptron; Endpoint Detection; Artificial Neural Network.

Leave a Reply

Your email address will not be published. Required fields are marked *