Abstract

Abstract: Speech signals are complex in nature with relation to other sorts of communication media like text or image.Different types of noises (e.g., additive noise, channel noise, babble noise) interfere with the speech signals and drasticallyhamper standard clarity of the speech. Enhancement of speech signals could be a daunting task considering multiple kinds ofnoises while denoising a speech signal. Certain analog noise eliminator models are studied over the years for this purpose.Researchers have also delved into some machine learning techniques (e.g., artificial neural network) to reinforce speechsignals. During this study, a speech enhancement system is investigated using Convolutional Denoising Autoencoder (CDAE).Convolutional neural network (CNN) may be a special quite deep neural network which is suitable for 2D structured input(e.g., image) and CDAE could be a CNN based special quite Denoising Autoencoder. CDAE takes advantage from the 2Dstructured inputs of the features extracted from speech signals and also considers the local temporal relationship among thefeatures. within the proposed system, CDAE is trained considering features from noisy speech signal as input and clean speechfeatures as desired output. The CDAE method is tested on a standard dataset, called Speech Command Dataset, and attained80% similarity between denoised speech and actual clean speech. The system successfully achieved perceptual evaluation ofspeech quality (PESQ) value of 2.43 which succeeded other related existing methodsKeywords: Autoencoder, Automatic Speech Recognition, Convolutional Denoising Autoencoder, Denoising Autoencoder, DeepNeural Network, Fully Convolutional Network, Feed forward Neural Network, Mel Frequency Cepstral Coefficient