Robustness metric-based tuning of the augmented Kalman filter for the enhancement of speech corrupted with coloured noise

Abstract

In this paper, we describe a tuning method based on a robustness metric and extended to work with the augmented Kalman filter for enhancing coloured-noise-corrupted speech. The method proposed within utilises the robustness metric to provide dynamic and adaptive tuning of the Kalman filter gain in order to reduce the residual noise that results from poor speech model estimates. An analysis of the Kalman filter recursion equations is presented that augments the robustness metric equations to include coloured noise model parameters. Objective and blind AB subjective listening tests were performed on the NOIZEUS speech corpus for both white and coloured noises with the results being compared with the MMSE method. In the blind AB subjective testing, the 15 English-speaking listeners showed preference for the proposed method over both the MMSE and oracle Kalman filter methods (where clean speech parameters were used). These results imply that the proposed tuned Kalman filter produces more perceptibly-acceptable enhanced speech than the oracle Kalman filter, which is considered the ideal for this enhancement technique.

Robustness and Sensitivity Tuning of the Kalman Filter for Speech Enhancement

“…stationary noise conditions, we repeat the experiment inFigure 2. It can be seen from Comparing K 0 (n) obtained using KF-Oracle, Proposed, and AKF-RMBT[21] methods from the utterance sp05 corrupted with 5 dB: (a) non-stationary (babble) and (b) colored (f16) noises.238 (a) that K 0 (n) is closely similar to the oracle K 0 (n). Specifically, it maintains a 239 smooth transition at the edges and the temporal changes in speech regions are closely 240 matched to the oracle K 0 (n).…”

“…that the biasing effect is reduced significantly in K 0 (n) and closely similar to the oracle 246 K 0 (n). However, the AKF-RMBT method [21] still produced under-estimated K 0 (n) in 247 speech regions. In light of the comparative study, it is evident to say that the proposed 248 method adequately addresses the tuning of biased K 0 (n) both in real-life non-stationary 249 and colored noise conditions.…”

“…To examine the performance of the proposed tuning algorithm in real-life non- proposed method is more appropriate to mitigate the risks of distortion in the enhanced 243 speech than that of AKF-RMBT method [21]. We also repeat the experiment in Figure 2 244 except the utterance sp05 is corrupted by 5 dB colored (f16) noise.…”

“…Among the benchmark methods, the AKF-RMBT [21] method is found to be the best 382 preferred method (64%) followed by the AKF-IT method [13] (around 48%) by the…”

“…The 170 whitening filter is also constructed with the constant noise LPCs to pre-whiten each 171 noisy speech frame prior to compute speech LPC parameters. As a result, the tuning of 172 K 0 (n) [21] becomes irrespective in conditions having time-varying amplitudes, such as 173 babble noise.174Motivated by the shortcomings of[19][20][21], we propose J 2 (n) and J 1 (n) metrics-175 based tuning of KF gain, K 0 (n) for speech enhancement in real-life noise conditions.…”