Sound Event Classification Based on Fusion of Multiple Pre-trained Models

Although audio event classification is a task that involves recognizing event categories in audio signals, it is well-known that a single model may possess limited generalization ability, making it unable to accurately predict audio events. Therefore, this paper proposes a non-mean ensemble method. Specifically, this approach concatenates multiple classification models in the channel dimension and employs non-mean modules to generate learnable parameters, which enables confidence operations to be carried out on each model. By doing so, the approach does not rely on mean information from model outputs, but instead adapts weight adjustments based on learnable parameters, thereby enhancing its adaptability to different classification models and improving the flexibility of the prediction model. Furthermore, our test results on the official dataset of the 2023 DCASE Task4 demonstrate that the proposed method is effective. The best F1 score achieved by a single pre-trained model is 87.6%, with a corresponding mAP of 93.4%. After applying the non-mean ensemble method, the F1 score and mAP are respectively improved to 88.4% and 93.5%.