There are a plenty of research experiments and achieved results in various languages throughout the world regarding speech recognition. But, in Bangla language, early researchers in this field had qualified success, though the scenario is being changed in recent years. This research work aims at developing a neural network based connected digit recognition system in Bangla language. Firstly, a Bangla digit corpus has been developed comprising of male and female speakers. Speech is recorded in connected fashion and words are extracted through automatic segmentation. Then MFCC features of the segmented words are calculated and these feature values are sent as the input to the back-propagation neural network (BPNN). BPNN learning algorithm is used to train the network. The required time to train the network, number of hidden layers, error threshold and number of epochs are considered while training the network to reach the best possible recognition accuracy. This proposed system has been implemented using object oriented programming and the achieved recognition accuracy is very much satisfactory and consistent. The network has been tested for three different setups and the best recognition accuracy achieved for digit dataset is 98.46%.
Varieties of environmental sources of noise and distortion can degrade the quality of the speech signal in a communication system. This research work explores the effects of these interfering sounds on speech applications and introduces a technique for reducing their influence and enhancing the acceptability and intelligibility of the speech signal. In this work, a noise reduction system using single microphone method in time domain to improve SNR of noise contaminated speech is proposed. Traditional Spectral Subtraction method has been reviewed very well and the relationship with wiener filter is also illustrated. The Spectral Subtraction method has been generalized and the focus is put on reducing noise from speech in single channel signals. Voice Activity Detector (VAD) is ignored in this proposed system, because a-priori information about the noise is assumed. The research has been conducted using Gaussian White Noise and Color Noise. The experimental result shows a remarkable improvement in SNR for the generalized version and it is noticed that the result is very much satisfactory when white noises are added but the addition of color noise produces a comparatively poor improvement report. The system has been tested with eight different datasets and on an average, 65.27% improvement in SNR (Signal to Noise Ratio) for White Noise using Generalized Spectral Subtraction Method is achieved comparing with Traditional Spectral Subtraction Method. The average improvement in SNR for Color Noise recorded is 53.31%. The Generalized Spectral Subtraction method is shown to improve the speech quality and to improve SNR as well.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.