A Multi-Resolution CRNN-Based Approach for Semi-Supervised Sound Event Detection in DCASE 2020 Challenge

Sound Event Detection is a task with rising relevance over the recent years in field of audio signal processing, due to creation specific datasets such as Google AudioSet or DESED (Domestic Environment Detection) and introduction competitive evaluations like DCASE Challenge (Detection Classification Acoustic Scenes Events). The different categories acoustic events can present diverse temporal spectral characteristics. However, most approaches use fixed time-frequency resolution represent segments. This work proposes multi-resolution analysis for feature extraction Detection, hypothesizing that resolutions be more adequate detection sound event categories, combining information provided by multiple could improve performance systems. Experiments are carried out dataset context 2020 Challenge, concluding combination up 5 allows neural network-based system obtain better results than single-resolution models terms event-based F1-score every category PSDS (Polyphonic Score). Furthermore, we analyze impact score thresholding computation results, finding standard value 0.5 suboptimal proposing an alternative strategy based threshold each category, which obtains further improvements performance.