Comparison on Neural Network based acoustic model in Mongolian speech recognition

Deep Neural Networks (DNNs) beat the Gaussian Mixture Models (GMMs), and become the state-of-the-art techniques for acoustic model. Then various neural networks based acoustic models are proposed to make the speech recognition systems better and better. However these successes are not adopted in the...

Full description

Saved in:
Bibliographic Details
Published in2016 International Conference on Asian Language Processing (IALP) pp. 1 - 5
Main Authors Hongwei Zhang, Feilong Bao, Guanglai Gao, Hui Zhang
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.11.2016
Subjects
Acoustics
Hidden Markov models
Logic gates
Long Short-Term Memory (LSTM)
Mongolian speech recognition
Neural Network (NN)
Neural networks
Speech
Speech recognition
Training
Vocal Tract Length Normalization (VTLN)
Online AccessGet full text

Cover

More Information
Summary:Deep Neural Networks (DNNs) beat the Gaussian Mixture Models (GMMs), and become the state-of-the-art techniques for acoustic model. Then various neural networks based acoustic models are proposed to make the speech recognition systems better and better. However these successes are not adopted in the researches of Mongolian speech recognition. This study fills in this gap. We study a series of neural networks based acoustic models, apply them in the Mongolian speech recognition systems, and compare their performance. We find out the Long Short-Term Memory (LSTM) is the best model among them. Finally, by using the LSTM acoustic model together with data augmentation technique, which uses various combinations of Vocal Tract Length Normalization (VTLN) warping factor and time-warping factor to artificially expand the amount of data, we refresh the recode of the Mongolian speech recognition. Compared with the best DNN-based speech recognition system, we cut the Word Error Rate (WER) nearly by half.
DOI:10.1109/IALP.2016.7875921