Abstract

Environmental noise in urban settings has several adverse effects on the population's health and quality of life. Detecting and classifying sound sources is a primary task to control this noise. Currently, advanced sound event detection (SED) is achieved through the use of deep learning (DL) models. However, one downside of DL is the large amount of labeled data required to train the models effectively. Labeling large datasets is costly and time-consuming, and synthetically generated datasets created so far have yet to be able to provide the spatio-temporality of the sound sources within the soundscape. To address these problems, this paper introduces a synthetic polyphonic ambient sound source (SPASS), a novel synthetic polyphonic dataset that features spatio-temporal labels of sound sources. The new dataset, created with acoustic virtual reality tools, is used to pre-train a DL model for polyphonic SED in urban soundscapes. The SPASS dataset was created by simulating five different soundscapes to ensure that DL models are not tailored to any specific soundscape type. The DL model was evaluated using a transfer learn-ing methodology on two real-world SED tasks involving urban sound events. The DL model pre-trained with SPASS significantly outperformed that trained solely on actual data and achieved similar results as one pre-trained using Google Au-dioset, a dataset that is 100 times larger than SPASS. Additionally, pre-training with the new SPASS dataset proved to be better at effectively detecting and classifying traffic-related sound sources in urban soundscapes.