Research on speech recognitions were initiated by notions of HMIs (human machine interactions). ASR (Automatic voice Recognition) is method that employs implementable algorithms on computers to translate voice signals as strings of words. Systems can understand human speech inputs. Speech signals transmit two crucial forms of information, including speech contents and identities of speakers. Existing system have issues with speech recognition accuracies and feature extractions. To overcome these problems, in this work, ECNN (Enhanced Convolution Neural Networks) is proposed. The main modules are pre-processing, feature extraction and speech recognition. In pre-processing, noises are removed by the application of Wiener filters for obtaining cleaner speeches. Subsequently, MPCA (Modified Principal Component Analysis) is used for feature extractions where most informative features are extracted. Noise corrupt speech feature matrices are the focus of MPCA and it is demonstrated that the generated sparse partitions reveal speech dominant properties. The ECNN algorithm is subsequently used for speech recognitions and thus enhancing speech recognitions with reduced error rates. The experimental results demonstrate in the conclusion that the proposed MPCA+ECNN algorithm provides better values in comparison with other methods in terms of MSE (Mean Square Error) rates, accuracy, specificity and execution times.

Speech recognition, Enhanced Convolution Neural Network (ECNN), Modified Principal Component Analysis (MPCA)