Convolutional Neural Network for Removal of Environmental Noises from Acoustic Signal
Shibani Kar, Vishwajeet Mukherjee · 2023
Environmental noises affects the quality and intelligibility of acoustic signals. Speech, audio, sound are common examples of acoustic signals. The noises present in our environment especially babble noise can affect the performances of applications such as speech and speaker recognition, hearing aids, mobile telephone etc. The research works presented earlier for the removal of these noises from the recorded signal were based on the estimation of noisy part from the signal to prepare a mask for the removal of the noises. But the recent developments in neural networks has provided a mapping based method that can directly map the noisy speech to clean form. In this paper, a comparative study of convolutional neural network architectures for the enhancement of acoustic signals such as speech is done. The NOIZEUS datasets used for the experimental work is an open source speech dataset. The evaluation metrics PESQ, improvement in SNR, STOI, Segmental SNR (SSNR), and Signal to Distortion Ratio (SDR) are used to compare the performances of these networks for the enhancement of noisy signals. The performance of Convolutional networks in auto-encoder framework performs better in terms of better PESQ, STOI, Segmental SNR, SNR Improvement, SDR for network trained with babble noise at 0dB SNR.