"ENVIRONMENTAL SOUND CLASSIFICATION" (Using Deep Learning Models- LSTM, BiLSTM, GRU, CNN)

Abstract

This report suggests a powerful Environmental Sound Classification (ESC) system using Convolutional Recurrent Neural Networks (CRNN) that use convolutional layers to learn spatial features from the spectrogram representation of sounds and recurrent layers for modeling temporal patterns in sounds. UrbanSound8K dataset, which contains 8,732 sound samples in 10 urban sound classes, was augmented using data augmentation techniques such as shifting, time stretching, and noise addition to 15,000 samples. Mel-Frequency Cepstral Coefficients (MFCCs) were extracted to analyze sound data. Four deep learning architectures; Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (BiLSTM) were trained and evaluated. Of these, BiLSTM performed the best with 99.84% training accuracy, 93.75% validation accuracy, and a macro F1-score of 0.92. CNN followed with 99.89% and 92.83% training and validation accuracy, respectively. GRU (97.00%/91.00%) exhibited quick convergence and minimal overfitting, hence being perfect for light-weighted applications, while LSTM had 97.27%/90.58%. The results reveal the phenomenal ability of CRNN-based approaches, especially BiLSTM, in offering trustworthy, precise, and versatile sound perception in intelligent, sound-responsive systems.