Audio Signal Classification Using DSP and LSTM Neural Network

Abstract

This paper presents a comprehensive audio signal classification system that integrates Digital Signal Processing (DSP) with deep neural network architectures for accurate multi-class audio categorization. Raw audio waveforms are preprocessed through a rigorous DSP pipeline encompassing pre-emphasis filtering, Short-Time Fourier Transform, Mel-Frequency Cepstral Coefficient extraction, and spectral features computation. The extracted feature representations are supplied to two deep learning models: a Convolutional Neural Network (CNN) operating on log-mel spectrograms and a "Bidirectional Long Short-Term Memory (BilSTM)" network with an attention mechanism processing MFCC sequences. Both architectures are evaluated on the UrbanSound8K benchmark dataset comprising 8,732 labelled clips across 10 urban sound categories. The CNN model achieves 88.4% test accuracy, outperforming both the BiLSTM (85.7%) and a Support Vector Machine baseline (76.2%). Mathematical derivations of key DSP transforms and information-theoretic analysis of feature discriminability are presented. Data augmentation and regularization strategies contribute to robust generalization. The hybrid DSP-deep learning pipeline establishes a strong foundation for real-time audio classification on resource-constrained embedded platforms.