Title: Acoustic model combinations for continuous speech recognition system

Authors: R.K. Aggarwal; Mayank Dave

Addresses: Department of Computer Engineering, National Institute of Technology, Kurukshetra – 136119, Haryana, India. ' Department of Computer Engineering, National Institute of Technology, Kurukshetra – 136119, Haryana, India

Abstract: System combination is a promising way to obtain a significant improvement in performance as compared to the conventional form of single system model. In the field of automatic speech recognition (ASR), various approaches have been studied focusing on different aspects of feature extraction and acoustic modelling. These approaches can be combined to utilise their complementary information and to cope with the limitations of individual technique. In this paper we have proposed a novel approach in which three acoustic models based on maximum likelihood, discriminative and margin-based estimation are combined using a technique called as confusion network combination. Further, each acoustic model is associated with a different type of feature extractor to derive observation vectors for training and testing. Experimental results show 2%-5% reduction in error rate for Hindi ASR.

Keywords: automatic speech recognition; ASR; acoustic models; confusion network combination; Hindi; system combinations; Mel-frequency cepstrum coefficient; MFCC; perceptual linear prediction; PLP; discriminative training; large margin estimation; modelling; error rate.

DOI: 10.1504/IJCSYSE.2012.050231

International Journal of Computational Systems Engineering, 2012 Vol.1 No.2, pp.79 - 90

Published online: 28 Aug 2014 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article