Deep Neural Network-Based Classification of Body-Conducted Speech Using MFCC Features

This paper presents a deep learning-based system for body-conducted speech classification using signals from five classes, namely forehead,head set,rigid in-ear,temple vibration and soft in-ear. The system is being built by the Vibravox dataset, which comprise of 38 hours of speech from 188 participants recorded under different types of acoustic environment. Feature extraction using Mel Frequency Cepstral Coefficients(MFCCs) has been done after necessary prepossessing and normalizations. We propose a Deep Neural Network based model for classification of different speech signals. Experimental results portray the effectiveness of the proposed system in handling different types of variations in speech signal quality and acoustic environments, which results in comparable classification accuracy.