Robust Features with Convolutional Autoencoder Speech Command Recognition

Today, deep learning (DL) has proven to be a prominent method in many recognition tasks. However, deep learning-based Automatic Speech Recognition (ASR) has degraded performance when dealing with noisy data. In this study, we investigate the effectiveness of DL architectures using the autoencoder convolutional neural network (AECNN) combined with traditional feature extraction methods such as Mel-Filter Bank (MFB), Mel Frequency Cepstral Coefficients (MFCC) and Log Mel-Filter Bank (LMFB) to generate robust features for speech command recognition. We implement RMSprop as an optimizer to get optimal models. Based on our experiment results, we found that MFB features perform best in both clean and noisy conditions.