Speech Command Recognition System using Deep Recurrent Neural Networks

Speech command recognition has become increasingly relevant in the past few years. This study proposes a new speech command recognition system based on the Mel-frequency cepstrum coefficients (MFCC) feature extraction method, a well-established approach for speech feature extraction and deep recurrent neural networks. In this work, we have compared different signal processing techniques such as wavelet packet decomposition (WPD), continuous wavelet transforms (CWT) and empirical mode decomposition (EMD) to decompose speech signals and extract MFCC features to train long short-term memory (LSTM) recurrent neural networks. Comprehensive studies on the performance of different network parameters and different signal processing strategies are presented in this paper. The proposed neural network model trained with MFCC features extracted from signals preprocessed using wavelet packet decomposition has performed better than raw speech data as well as both CWT and EMD preprocessing.