A Reconfigurable Gaussian Base Normalization Deep Neural Network Design for an Energy-Efficient Voice Activity Detector

Deep neural networks and approximation computation are offered in this research report to construct an energy-efficient voice activity detector. A potential energy-efficient adaptive GNDNN accelerator architecture is proposed for voice activity detection (VAD).The VAD method is divided into two sections: feature extraction and voice/noise classification using a deep neural network with Gaussian basis normalisation (GNDNN).Initial pre-processing of input data: The high-frequency components of the digitalized voice signal are pre-emphasized, and you try to put it together with a little later in the signal processing, you'll be less sensitive to finite precision effects. The feature extraction module extracts the input voice signals feature value using MFCC, or Time-Frequency Non-negative Matrix Factorization (TFNMF). The GNDNN speech prediction phase classifies the TFNMF, MFCC output from feature extraction, determining if the signal is a voice or noise. The proposed method can be adjusted dynamically to satisfy various computation accuracy requirements. The most interesting accuracy result of our proposed approach is 98.75 percent. Similar to CNN, which reach accuracy of 97.25 percent.