Two-stage Adaptive Robust Noise Cancellation System for a Hindi Speech Database

Abstract

This paper presents a two-stage adaptive noise cancellation system for noise cancellation of Hindi speech signals corrupted by babble and factory noise, under different input signal-to-noise ratio (SNR) levels. The corrupted signals with different SNR levels are passed through three different noise cancellation systems: an adaptive noise cancellation system with a fast block least mean squares algorithm (FBLMS) algorithm; a fully connected convolutional deep neural network (FCCDNN); and a two-stage system with an adaptive noise cancellation system consisting of the FBLMS algorithm followed by FCCDNN. The performance of each noise cancellation system is evaluated and compared. The parameters such as output SNR, perceptual evaluation of speech quality and shorttime objective intelligibility are used for testing the intelligibility and quality of the recovered speech signals. The proposed system shows a maximum SNR improvement of 13.997 and 15.2965 dB at an input SNR of −5 dB for factory noise and babble noise, respectively, with significant increments in both of the evaluation metrics. The proposed system outperforms speech enhancement systems based on FCCDNN and FBLMS in terms of all performance parameters.