Inter-floor noise classification using convolutional neural network

Abstract

In apartment houses, noise between floors can disturb pleasant living environments and cause disputes between neighbors. As a means of resolving disputes caused by inter-floor noise, noises are recorded for 24 hours in a household to verify whether the inter-floor noise exceeded the legal standards. If the noise exceeds the legal standards, the recorded sound is listened to, and it is checked whether the noise comes from neighboring households. When done manually, this process requires time and is costly, and there is a problem of whether the listener’s judgments of the sound source are consistent. This study aims to classify inter-floor noise according to noise sources by using a convolutional neural network model. A total of 1,515 sound sources of data recorded for 24 h from three households were annotated, and 40 4s audio clips of six noise sources, including “Footsteps,” “Dragging furniture,” “Hammering,” “Instant impact (dropping a heavy item),” “Vacuum cleaner,” and “Public announcement system” were identified. Moreover, datasets of 16 classes using ESC50’s urban sound category audio were used to distinguish the inter-floor noise heard indoors from the external noise. Although DenseNet, ResNet, Inception, and EfficientNet are models that use images as their domains, they showed an accuracy of 91.43–95.27% when classifying the inter-floor noise dataset. Among the reviewed models, ResNet showed an accuracy of 95.27±2.30% as well as a highest performance level in the F1 score, precision, and recall metrics. Additionally, ResNet showed the shortest inference time. This paper concludes by suggesting that the present findings can be extended in future research for monitoring acoustic elements of indoor soundscape.

AUCO ResNet: an end-to-end network for Covid-19 pre-screening from cough and breath

“…It is not clear why data augmentation decreases the accuracy of the model. In fact, results on UrbanSound 8K dataset are lower when compared to techniques such as [28] that make use of data augmentation. Our intuition is that AUCO ResNet is very sensitive to the input audio quality.…”

“…NO means not present and YES means the presence of the pre-training or data augmentation technique. Model Accuracy Precision Recall F1 score AUC ROC Pre Trained Data Augmentation AUCO ResNet 0.7783 0.7851 0.7783 0.7709 0.9677 NO NO Chong et al [27] 0.751 NA NA NA NA NO NO Salamon et al [25] 0.75 NA NA NA NA NO NO Giannakopoulos et al [29] 0.731 NA NA NA NA NO NO Salamon et al [23] 0.73 NA NA NA NA NO NO Piczac et al [24] 0.73 NA NA NA NA NO NO Jin et al [26] 0.705 NA NA NA NA NO NO Salamon et al [22] 0.70 NA NA NA NA NO NO Shin et al [28] 0.8514 NA NA NA NA YES YES Shin et al [28] …”

“…On the other hand, the approach does not perform well in the case of data augmentation. State of art accuracy is achieved by Shin et al in [28] , where both data augmentation and transfer learning were adopted. Confusion matrices for AUCO-ResNet are reported in Figure 8 .…”

“…A dilated CNN-based Environmental Sound Classification Tasks D-CNN-ESC system and LeakyReLU activation function were adopted with the aim to evaluate ESC particularly on the UrbanSound8k dataset [27] . Authors of the study presented in [28] aim to classify inter-floor noise according to noise sources by using a convolutional neural network model, in [29] a similar task is performed on recognizing audio events in urban environments adopting a deep learning architectural scheme (Convolutional Neural Networks, CNNs), which has been trained to distinguish between different audio context classes.…”

“…Later the dot product (the angular parentheses) with a learned vector u is performed for inferring an initial attention mechanism which intuitively can be interpreted as a learning weighting factor for common patterns among instances of the same class. This attention is derived from the alignment model of attentions used in [49] which led to transformers in natural language processing tasks [28] .…”