雲间烟火Efficient Acoustic Front-End Processing for Tamil Speech Recognition using Modified GFCC Features
I.J. Image, Graphics and Signal Processing, 2016, 7, 22-31
Published Online July 2016 in MECS (/) DOI: 10.5815/ijigsp.2016.07.03
Efficient Acoustic Front-End Processing for Tamil Speech Recognition using Modified GFCC
Features
Vimala. C, V. Radha
Department of Computer Science, Avinashilingam Institute of Home Science and Higher Education for Women,
Coimbatore – 641043, Tamil Nadu, India.
E-mail: vimalac.au@, radhasrimail@
Abstract—Giving suitable input and features are always essential to obtain better accuracy in Automatic Speech Recognition (ASR). The type of signal and feature vectors given as an input is highly essential as the pattern matching algorithms strongly depends on these two components. The primary goal of this paper is to propose a suitable Pre-processing and feature extraction techniques for speaker independent speech recognition for Tamil language. The five pass Pre-processing and three types of modified feature extraction techniques are introduced using Gammatone Filtering and Cochleagram Coefficients (GFCC) to achieve better recognition performance. The modified GFCC features using multi
taper Yule walker AR power spectrum, combinational
Fig.1. Basic Components of an ASR System
features using Formant Frequencies (FF), combined frequency warping and feature normalization techniques Pre-processing is performed in signal processing using Linear Predictive Coding (LPC) and Cepstral Mean applications, in order to make useful spectral analyzes. Normalization (CMN) are investigated. The experimental The general Pre-processing techniques applied for signal results prove that the proposed techniques have produced processing are, Pre-emphasis, framing and windowing. high recognition accuracy when compared with the Feature extraction technique is used to transform the conventional GFCC feature extraction technique. speech signal into useful parametric representations. These parameters are used to group the similar patterns Index Terms—Gammatone Filter banks, Multi Taper and to recognize them [4]. The most popular feature window, Yule Walker AR, Formant Feature extraction, extraction techniques used for speech related applications Cepstral Mean Normalization and Tamil Speech are Mel Frequency Cepstral Coefficients (MFCC), Linear Recognition. Predictive Coding (LPC), Perceptual Linear Predictive
I. INTRODUCTION
Speech recognition became an active topic of research in recent years. It has been applied in many research areas like dictation, dialog system and voice based information search etc. Developing an ASR system comprises of two significant components, namely, speech front-end and back-end processing. The front-end processing is used to build unique models for each speech patterns and back–end processing is used to perform pattern matching [1] [2]. Figure 1 shows the basic components of an ASR system.
Speech front-end processing includes various tasks, namely, speech acquisition, Pre-processing and feature extraction. Back-end processing is used to create an unique model for each speech patterns by extracting the most essential properties of a speech signal based on Pre-
processing and feature extraction techniques [3].
(PLP) coefficients, wavelet features and auditory features. Even though many feature extraction techniques are available for speech recognition, selecting suitable features is always important, as the entire process depends on the feature vectors given as an input. Speech recognition systems never give good performance, if the selected features are not suitable [4].
To avoid this, psychological studies of the human auditory and articulatory systems are always necessary [1] [4]. During the recent years, GFCC features that are work based on an inner Ear model is widely used because of their significant improvements in the speech related applications [5].
However, for Tamil Speech Recognition most of the experiments are based on conventional MFCC and LPC features with HMM and Neural Networks (NN). Hence, the particular contribution of our work is involving GFCC features with other machine learning techniques, like MLP and SVM. This research work is mainly focused
