Audio Signal Classification using MFCC-based Deep Neural Network

Abstract

It has long been recognized that the accurate classification of audio signals is a major challenge in the field of speech and signal processing research. Applications range from speaker recognition, speech emotion detection, and environmental sound categorization. This is especially challenging due to the dynamic nature of audio signals. For example, the presence of ambient noise can make the recording unclear, different speakers have different vocal characteristics and recording setups are very different across datasets. To overcome these challenges, we propose a classification framework based on both Deep Neural Networks (DNNs) and Mel-Frequency Cepstral Coefficients (MFCCs). The raw audio signals are passed through our pre-processing pipeline which does multiple stages like rescaling signals for consistency, removing silent portions, and pre-emphasis filtering to improve the clarity of the signal. The MFCC extraction step captures the spectral features which are perceptually important. This is illustrated with MFCC heatmaps of different audio categories. The normalized features are fed into a fully connected DNN with dropout and early stopping to prevent overfitting. Experimental results show that the classification accuracy is 94.61%