With ever growing urbanization, the environmental noise is becoming hazardous. Vehicular traffic, locomotives, heavy machinery in industry, and construction processes are the major sources of noise pollution. It has adverse effects on the health of humans as well as that of the wild life. World Health Organization (WHO) puts noise pollution as the second major cause of illness due to environmental reasons. The effects of noise pollution on the quality of life are usually ignored. Due to this reason it is common, even in the first world countries, to have the WHO's peak noise standards violated in residential areas. Therefore, there is a need to have a real time, portable and easy to replicate, mechanism to monitor the noise sources. In this work, we propose a novel architecture of a deep neural network to classify a 10-class environmental noise data called URBANSOUND8K. This network is comprised of three components, (1) one dimensional two level Discrete Wavelet Transform (DWT) component, (2) branched component for feature extraction through auto-encoders, and (3) LSTM and fully-connected layers based classification component. With all components combined, we call this network DWTNet. By embedding the DWT component as a part of network, we eliminate the need of prior data conversion into spectral and/or spectro-temporal domains. The efficiency of DWTNet is comparable to the state of the art networks with significantly lower number of trainable parameters. We analyze the contribution of classification accuracy. We further study some of the classification results individually and show that some of the mis-classifications are actually multi-class classifications with distributed decision confidence.

中文翻译：

---

基于离散小波变换的分支深度混合网络用于环境噪声分类

随着城市化进程的不断发展，环境噪音变得越来越危险。车辆交通、机车、工业重型机械、施工过程是噪声污染的主要来源。它对人类以及野生动物的健康产生不利影响。世界卫生组织（WHO）将噪音污染列为环境原因导致疾病的第二大原因。噪音污染对生活质量的影响通常被忽视。因此，即使在第一世界国家，住宅区违反世界卫生组织峰值噪音标准的情况也很常见。因此，需要一种实时、便携且易于复制的机制来监测噪声源。在这项工作中，我们提出了一种新颖的深度神经网络架构，用于对 10 类环境噪声数据进行分类，称为 URBANSOUND8K。该网络由三个组件组成：(1) 一维两级离散小波变换 (DWT) 组件，(2) 通过自动编码器进行特征提取的分支组件，以及 (3) LSTM 和基于全连接层的分类组件。将所有组件组合在一起，我们将该网络称为 DWTNet。通过将 DWT 组件嵌入为网络的一部分，我们无需事先将数据转换为谱域和/或谱时域。DWTNet 的效率与可训练参数数量显着减少的最先进网络相当。我们分析分类准确性的贡献。