Convolutional Neural Network Based Automatic Speech Recognition for Tamil Language

Abstract

AbstractAutomatic Speech Recognition (ASR) systems recognize the text transcript for the given speech signal. Most of the ASR systems developed for Indian regional Language recognition follows the traditional model such as Hidden Markov Model (HMM) and Gaussian Mixture Model (GMM) till now. In recent years English, Chinese, German Language based ASR system used deep learning models for recognition with improved recognition rate. In the proposed ASR system for Indian regional language we designed Acoustic model by using Convolutional Neural Network (CNN). Acoustic Model identifies the phones for the given speech signal. To implement the CNN model we used first and second order derivations of Mel frequency spectral coefficient (MFSC) and Gammatone frequency cepstral coefficient (GFCC) as a feature extraction technique. CNN have significant properties such as feature sharing, learning important features automatically and polling that will improve the accuracy of ASR system. The proposed work is carried out by using Open sourced high-quality multi-speaker speech data set and achieved 90.63% training accuracy and 81.25% testing accuracy. For the training and testing purpose we used isolated tamil words and also applied modification like changing the pitch, backward and forward shift in time and adding noise to the data. KeywordsASRCNNMFSCGFCCHMMGMMNatural Language Processing (NLP)Speech processingAcoustic model