Chapter 2 - Recursive noise estimation-based Wiener filtering for monaural speech enhancement

Abstract

In this chapter, we propose a recursive noise estimation-based Wiener filtering (WF-RANS) approach for monaural speech enhancement. The proposed approach describes the conventional spectral subtraction method in the Wiener filtering framework using a novel noise reduction scheme and proceeds in two steps. First, we estimate the noise-power spectrum using the recursive approach. Second, we estimate the speech spectrum from the noisy signal by Wiener filtering. The recursive noise estimation estimates the noise from the present and past frames of the noisy speech continuously, using a smoothing parameter with a value between 0 and 1. The Wiener filtering depends on the filter transfer function from sample to sample based on the speech signal statistics (i.e., mean and variance). To measure the performance of the proposed approach, we used both objective and subjective evaluations on speech sentences pronounced by both male and female speakers and corrupted by white Gaussian and pink noises at varying levels of signal-to-noise ratio (SNR) from 0 to 15dB. For objective evaluations, we used global SNR, segmental SNR (SNRseg), and perceptual evaluation of speech quality (PESQ) score. We then conducted spectrograms and informal listening tests for subjective evaluations. Finally, we compared the performance of the proposed approach with conventional speech enhancement methods, and findings show that the proposed approach results are more pleasant and intelligible to the human listener for both noise types.