Performance Comparison among Different Wiener Filter Algorithms for Speech Enhancement

---

Abstract

Reconstruction of original speech signal from the noisy signal is still a difficult task as nature and characteristics of noise signal vary in time and also depends on application and context for which filtering has to be performed. There are different methods individually or also combined with other filters to get better performance in speech signal processing. Wiener filter is very popular for speech signal processing, it is a linear estimator and this makes it less complex and easy to handle. There are various algorithms to solve Wiener filtering problem. In the proposed paper, performance comparison among different wiener filter algorithms is performed in context of speech signal processing. The detailed analysis is presented for comparison among Implicit Wiener filter, Two Step Noise Reduction Wiener filter and Wiener filter with Harmonic Regeneration. The performance is evaluated for both male and female speech samples. Different non-stationary noise (airport, babble, car, station, street, train, exhibition, restaurant) and stationary noise (AWGN) generated for different SNR values are considered. Performance evaluation is done through objective quality measures such as Log likelihood Ratio (LLR), Cepstrum Distance, Weighted Spectral Slope (WSS), and frequency weighted Segmental SNR in MATLAB.

---

Keywords

Implicit Wiener Filter; Objective Measures; Two Step Noise Reduction; Harmonic Regeneration