A Non-Iterative Kalman Filtering Algorithm with Dynamic Gain Adjustment for Single-Channel Speech Enhancement

Abstract

In this paper, we present a non-iterative Kalman filtering algorithm that applies a dynamic adjustment factor on the Kalman filter gain to alleviate the negative effects of estimating speech model parameters from noise-corrupted speech. These poor estimates introduce a bias in the first component of the Kalman gain vector, particularly during the silent (non-speech) regions, resulting in a significant level of residual noise in the enhanced speech. The proposed dynamic gain adjustment algorithm utilises a recently developed metric for quantifying the level of robustness in the Kalman filter. Objective and human subjective listening tests on the NOIZEUS speech database were performed. The results showed that the output speech from the proposed algorithm has improved quality over the noniterative Kalman filter that uses noisy model estimates and is competitive with the MMSE-STSA method. 

Robustness and Sensitivity Tuning of the Kalman Filter for Speech Enhancement

“…The KF-based SEAs in [19,20] address tuning of K 0 (n) using J 2 (n) and J 1 (n) metrics 107 for speech enhancement in WGN condition as described next. We analyze the shortcomings of existing KF-based SEAs [19,20] in terms of biased interpretation of K 0 (n).…”

“…The KF-based SEAs in [19,20] address tuning of K 0 (n) using J 2 (n) and J 1 (n) metrics 107 for speech enhancement in WGN condition as described next. We analyze the shortcomings of existing KF-based SEAs [19,20] in terms of biased interpretation of K 0 (n). For this purpose, we conduct an experiment with the utterance sp05 (Wipe the grease off his dirty face) of NOIZEUS corpus [1,Chapter 12] (sampled at 8 kHz) corrupted with 5 dB WGN noise [23].…”

“…For this purpose, we conduct an experiment with the utterance sp05 (Wipe the grease off his dirty face) of NOIZEUS corpus [1,Chapter 12] (sampled at 8 kHz) corrupted with 5 dB WGN noise [23]. In [19,20], a 20 ms non-overlapped rectangular window was considered for converting y(n) into frames as: (e)) to the clean speech (Figure 1 (a)).…”

“…In [19,20], So et all assumed that the impact of WGN in {ã i } is negligible. Thus,σ 2 w could be approximately estimated as:…”

“…In non-oracle case, it is also observed that J 2 (n) ≈ 1 typically during speech pauses of y(n, k) (e.g., 0-0.15 s or 1.8-2.19 s in Figure 1 (c)). Therefore, J 2 (n) metric is found to be useful in tuning biased K 0 (n) as [19]:…”