Assessment of Mel-Filter Bank Features on Sound Classifications Using Deep Convolutional Neural Network

Sound classification has been widely used but some aspects such as best features for classification, presence of background noise, and short duration with different characteristics make sound classification based on audio to be a challenging task. This study aims to assess the Mel-filter bank features on sound classifications using a deep convolutional neural network. We collected the DCASE 2018 task 2 challenge training data, which contained music and non-music sound classes, and then we picked six categories with distinct frequencies, wave speeds, and durations. The audio noise was then filtered using a pre-emphasis filter. The 6,500 samples were considered to generate 3-second audio data using a random sampling approach with replacement and transform them into Mel-filter bank features to attain feature vectors. These features were further input to the deep convolutional neural network for classification. The model performance was measured using seven metrics. The results showed that the log-Mel (pow-dB) produced the highest accuracy of 95.37% followed by 92.82% of log-Mel (amp-dB). The least accuracy of 82.34% was found in Mel-spectrogram (amp-freq). Overall, the log-Mel (pow-dB) had an impressive performance in contrast with other features. All features were subjected to a human hearing in Mel-scale.