A Multi-feature Fusion Based Method For Urban Sound Tagging
Jisheng Bai, Chen Chen, Jianfeng Chen · 2019
Noise pollution is one of the serious issues for citizens. Mapping urban noise is essential to improve the quality of life for residents and construction for smart cities. Yet, most cities lack effective classification or tagging methods to monitor urban noise. To tackle this challenge, we propose a multi-feature fusion based method for urban sound tagging (UST). This method combines various features and Convolutional Neural Networks (CNNs) to predict whether noise of pollution is present in a 10-second recording. Log-Mel, harmonic, short-time Fourier transform (STFT) and Mel Frequency Cepstral Coefficents (MFCC) spectrograms are fed into different CNN architectures. And a fusion method is applied to make the final outputs. The proposed method is evaluated on the DCASE2019 task5 dataset and achieves a macro-AUPRC score of 0.68, outperforming the baseline system of 0.54.