Colloquium Computer Science, Professor M. Sodanil (King Mongkut's University of Technology North Bangkok)

Title: Neural Networks in Speech Recognition for Acoustic Modeling of Tonal Language

                           

Abstract:

The baseline system of an automatic speech recognition (ASR) normally uses Mel- Frequency Cepstral Coefficients (MFCC) as feature vectors. However, for tonal language like Thai, tone information is one of the important features which can be used to improve the accuracy of recognition. This topic related to a method of building an acoustic model for Thai-ASR using a combination of MFCC and tone information as an input feature vector. In addition,  Artificial Neural Network (ANN) multilayer perceptrons is appled to estimate the posterior probabilities of a class model given a sequence of observation input. The performance of the ANN approach is compared with the Gaussian Mixture Model (GMM) with the Hidden Markov Model Toolkit (HTK). The experiments were carried out with 2-grams and 3-grams of a language model. The training and test data sets were recorded from male and female speakers. The results showed that the combination method for ANN input can be used to improve the performance of Thai-ASR in terms of reducing word error rate.

Colloquium coordinators are Prof.dr. M. Aiello (e-mail : M.Aiello rug.nl ) and
Prof.dr. M. Biehl (e-mail: M.Biehl rug.nl )

http://www.rug.nl/research/jbi/news/colloquia/computerscience