Lightweight Convolutional-iConformer for Sound Event Detection

The development of a sound event detection (SED) system is no trivial task where one has to consider both audio tagging and temporal localization concurrently. Often model ensembling is adopted to increase the overall detection accuracy. However, this can result in a large system that may face deployment issues in a resource-constrained environment. Subsequently, strongly labeled data was found to improve the audio classification performance in sound-related domains; this may indicate that such data may be required for SED model development. However, such data will inevitably contain a certain level of noise. In order to reduce the number of parameters, we proposed a lightweight system that utilized an improved depthwise separable convolution and an improved conformer layer. This lightweight system is then trained using an extension of the binary cross-entropy loss which considers the reverse binary cross-entropy to combat the noise that may be present in the training data. Based on the proposed framework, our lightweight system can obtain an event-based F1-score of 52%, and the ensemble of four systems through posterior averaging can further improve the event-based F1-score to 53.5%. Such results indicate a minimum margin of 16% against the Detection and Classification of Acoustic Scenes and events (DCASE) 2020 challenge task 4 baseline system. By comparing against the nonensembled system of the first-place submission in DCASE 2020 challenge task 4, our nonensembled system can achieve a higher event-based F1-score of 6% with 75% fewer parameters. In terms of the performance of the ensembled system, our approach remains competitive against the ensembled system of the first-place submission in DCASE 2020 challenge task 4 and has a winning margin of 2.9% with 88% fewer parameters. Comparison with other state-of-the-art also indicates that our system performance is better despite using a lightweight system.