FPGA-Based CNN-LSTM System for Audio Signal Classification

Convolutional Neural Networks (CNNs) combined with Long Short-Term Memory (LSTM) networks are strong machine learning models known for their high accuracy in decision-making. However, they require a large amount of computation and memory, which makes them difficult to deploy on edge devices, leading to higher latency and energy consumption. Field Programmable Gate Arrays (FPGAs) offer a compact and flexible hardware platform with good computing performance and low power consumption, making them suitable for edge computing applications. In this work, we implement an audio signal classification system based on a CNN–LSTM model on an FPGA platform and evaluate its performance. Three preprocessing methods are compared: Independent Component Analysis (ICA), Principal Component Analysis (PCA), and Non-Negative Matrix Factorization (NMF). These methods are integrated with the CNN–LSTM classifier during training and evaluated using FPGA-based inference. Mel-Frequency Cepstral Coefficients (MFCCs) are used as input features. The trained CNN–LSTM models are deployed on hardware using the Vitis AI tool. The results show that FPGA-based implementation provides significantly faster execution compared to CPU-based inference, with the ICA-based model achieving the best performance.