Automatic Multiple Sounds Detection with Recurrent Neural Networks (LSTM)

Sound event detection (SED) aims to recognize and distinguish various types of events, encompassing those associated with humans, nature, the environment, household machinery, and more. In real-world scenarios, it is common to encounter multiple sound sources co-occurring, leading to overlapping sound events. For instance, such events could include bird songs, footsteps, or the engine sounds of motorbikes in realistic environments. SED finds practical applications across several fields, including Home Automation and Security, Multimedia Organization and Retrieval, and Human–Computer Interaction. In our research, we propose a model that leverages LSTM and CNN algorithms to identify different sounds automatically. In this model, to train and evaluate the sound samples, we utilized a Kaggle audio dataset which contains 18,873 files that were classified into 41 distinct classes. We extracted various audio data features from the samples, such as Mel-spectrograms, spectral bandwidth, spectral centroid, and statistical features like spectrogram mean and median. By combining these features, we successfully trained our LSTM-RNN and CNN models. In our research, LSTM achieved 73.1% accuracy on the chosen dataset.