Memory-Efficient Sound Event Detection using Multi-Head Gated Recurrent Unit

Sound event detection is a critical task with widespread applications in industrial and scientific domains. However, most state-of-the-art methods rely on large neural networks with a substantial number of parameters, rendering them unsuitable for deployment in resource-constrained environments. Additionally, these approaches typically employ a single recurrent neural network as a decoder to predict the probabilities of event occurrences at specific timestamps. In contrast, this paper introduces a model featuring a multi-head gated recurrent unit decoder, which achieves competitive prediction accuracy while significantly reducing the number of trainable parameters. Our proposed model attains an F<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> -score of 62.53% and a polyphonic sound detection score of 0.4979 on the DESED evaluation dataset, requiring only 3.2 million trainable parameters, thereby achieving state-of-the-art performance among low-complexity models. Furthermore, we provide a comprehensive analysis of the proposed approach. First, we demonstrate that multi-heading enables the creation of models with comparable or superior prediction accuracy while utilizing significantly fewer trainable parameters. Second, we show that our approach is more effective for improving the model performance compared to simply increasing the hidden size of the network. Finally, we explore the limitations of the proposed technique in scenarios where auxiliary pretrained transformers are used to generate additional informative embeddings from input audio recordings.