Recognizing Command Words using Deep Recurrent Neural Network for Both Acoustic and Throat Speech

The importance of speech command recognition in a human-machine interaction system is increased recent years. In this study, we propose deep neural network-based for acoustic and throat recognition. We apply preprocessed pipeline to create the input learning model. Firstly, commands are decomposed into components using well-known signal decomposition techniques. Mel-frequency cepstral coefficients (MFCC) feature extraction method applied each component obtain inputs system. At stage, compare performance different techniques such as wavelet packet (WPD), continuous transform (CWT), empirical mode (EMD) order find out best technique our observe that WPD shows terms classification accuracy. This paper investigates long short-term memory (LSTM)-based recurrent network (RNN), which trained extracted MFCC features. proposed tested commands. Moreover, also train test model mic. well. Lastly, transfer employed increase accuracy weights with used initialize Overall, have found significant both speech. LSTM much better than GMM-HMM model, convolutional networks CNN-tpool2 residual res15 res26 an score over 97% on Google’s Speech Commands dataset achieve 95.35% data set technique.