Diffusion-Based Convolutional Recurrent Neural Network for Improving Sound Event Detection

The latent diffusion model and convolutional recurrent neural network in this article combine benefits of diverse networks to recognize sound events. The system was trained using large datasets and then used transfer learning to adapt to the intended job to overcome data shortages. The latent diffusion model compresses the audio mel-spectrogram into a discrete representation. A convolutional neural network (CNN) is then coupled as the RNN front-end to build a feature map. The feature map is then used to forecast temporal-spectral attention maps via an attention module. For adaptive feature refinement, the input spectrogram is multiplied by the created attention maps. Finally, trainable scalar weights combine back-end RNN fine-tuned features. We found that the recommended systems outperform the baseline and state-of-the-art utilizing the DCASE 2016 and DCASE 2017 challenges. Experiments on DCASE 2016 revealed a peak in system performance with F1 of 66.2% and ER of 0.42. DCASE 2017 dataset findings show F1 and ER at 68.1% and 0.40, respectively.