Noise robust speech encoding system in challenging acoustic conditions

Wireless communication systems are facing challenges in maintaining high-quality speech transmission due to the growing amount of data being transmitted. Linear predictive coding (LPC) is by far the most popular speech coding technique. A well-known shortcoming of LPC-based methods is their sensitivity to noise. In recent research, slight performance improvement is observed for LPC-based speech coding system under noisy environments. In this paper, we propose a pre-processing speech enhancement technique based on deep neural networks (DNN) to improve speech coding quality in negative signal-to-noise ratio (SNR) conditions. Furthermore, this study emphasizes the significance of phase information and underscores the potential benefits of employing DNN-based pre-processing modules to improve speech coding quality. We evaluate our proposed method using two databases, NOIZEUS and Kannada, which contain a wide range of noise levels. Experimental results show that our method outperforms traditional LPC, as demonstrated by the speech quality metrics.