Environmental Noise Reduction based on Deep Denoising Autoencoder

Abstract

Speech enhancement plays an important role in Automatic Speech Recognition (ASR) even though this task remains challenging in real-world scenarios of human-level performance. To cope with this challenge, an explicit denoising framework called Deep Denoising Autoencoder (DDAE) is introduced in this paper. The parameters of DDAE encoder and decoder are optimized based on the backpropagation criterion, where all denoising autoencoders are stacked up instead of recurrent connections. For better speech estimation in real and noisy environments, we include matched and mismatched noisy and clean pairs of speech data to train the DDAE. The DDAE has the ability to achieve optimal results even for a limited amount of training data. Our experimental results show that the proposed DDAE outperformed the baseline algorithms. The DDAE shows superior performances based on three-evaluation metrics in noisy and clean pairs of speech data compared to three baseline algorithms.