Subband spectral histogram feature for improved sound recognition in low SNR conditions

In this work, we use the subband intensity histogram values extracted from the spectrogram image of sound signals to form the feature vector for sound classification in an audio surveillance application. We propose two features based on this approach. Firstly, we extract the histogram features from the short time Fourier transform spectrogram image of sound signals, which we refer as the spectral histogram feature (SHF). Secondly, we apply the mel-filter to the spectrogram image before histogram feature extraction which we refer as the mel-spectral histogram feature (MSHF). When compared to baseline features from similar work, the SHF was shown to give significantly improved results in low SNR conditions with a higher overall classification performance. In addition, the MSHF produced even better results than the SHF with the added advantage of a lower feature dimension.