Noise Reduction for Multi-Channel Speech Enhancement System

Abstract – This paper introduces a deep learning-based system for environmental noise reduction and speech enhancement, designed to improve audio clarity in applications like hearing aids and voice-activated devices. The system combines Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) to classify and filter different types of environmental noise. It operates in two stages: first, the audio input is pre-processed into a time-frequency representation (such as a spectrogram), and then noise reduction is performed using a deep denoising autoencoder (DDAE). Implemented on the STM32746G-Discovery embedded platform, the system is capable of real-time processing, making it suitable for low-latency applications. Experimental results show that the system achieves a 75% noise classification accuracy and significantly enhances the Signal-to-Noise Ratio (SNR) of speech. Additionally, the open-source nature of the project encourages further development and customization for various practical uses. The system operates in two stages. In the first stage, the audio input is pre-processed into a time-frequency representation, such as a spectrogram, which captures both temporal and frequency-based information from the audio signal. This detailed representation is crucial for feeding structured data to the deep learning models. In the second stage, the system applies a deep denoising autoencoder (DDAE) to classify and reduce the noise while preserving speech. By combining CNNs for feature extraction and RNNs for temporal context, the system is able to distinguish between speech and noise with greater accuracy than traditional methods. Keywords: Noise Reduction, Speech Enhancement, Deep Learning, Convolutional Neural Networks, Recurrent Neural Networks, Embedded Systems.