A Dilated CNN and BiLSTM-GRU Based on Multi-Noise Classifier for Single Speech enhancement
Aliouat Mahfoud, Mohamed Djendi · 2024
In this study, we propose a new acoustic noise reduction system based on deep learning. To enhance speech quality n many fields such as medical, mobile telephony and free-hand applications, various techniques, including adaptive filtering, have been utilized. Advances in AI research and data availability have been leveraged in many recent works to enhance speech and reduce acoustic noise. In this paper, we propose a novel two-stage noise reduction system. Our system includes a BiLSTM-based classification model to identify the type of noise, followed by several noise reduction models based on dilated Inception CNNs and another architecture based on BiLSTM-GRU. The system was evaluated using both objective metrics, such as PESQ and STOI, and subjective listening tests. Our proposed ID-CNN model achieved a PESQ score of 2.23 and a STOI score of 85% for Babble noise, outperforming the baseline HSD model by 18%. Additionally, the model significantly improved performance for babble, car, and engine noises, demonstrating robust noise reduction across a wide range of environments