Search

In this paper, we present a cross evaluation of different a priori SNR estimation methods as well as different time-frequency analysis processing using a subjective listening test. The noisy signal is corrupted by different types of background noise i.e. babble and pink, and varying levels of input SNR (0 dB and 10 dB). The signals are processed using Short Time Fourier Transform (STFT) or Critical Band (CB) processing. After estimating the clean speech signal, it was presented to 10 participants for evaluation using a subjective listening test according to (ITU-TP.835) methodology. The results demonstrate that the participants preferred the speech signal processed using CB for low SNR levels and non-stationary background noise, which means that critical band based frequency scale is more useful in adverse noisy conditions.

In speech enhancement, the optimal minimum mean square error (MMSE) short-time spectral amplitude estimator requires knowledge about the probability density functions of speech and noise in the short-time Fourier transform domain, for every signal-to-noise-ratio (SNR). However, both of these quantities are unknown and are usually non-stationary in real- world scenario. To tackle this problem, this paper proposes a speech enhancement approach based on a set of Neural Networks of which each Neural Network is developed particularly for a critical band and a predefined SNR. In this speech enhancement approach, the Neural Network simulates a set of gain functions which attempts to match human hearing and optimises a particular SNR. The Neural Networks are trained using one speech signal contaminated with pink noise. The trained Neural Networks are evaluated using a test set consisting of 28 noisy speech signals. The speech enhancement results are compared to a state of the art MMSE based speech enhancement technique in terms of four speech quality metrics namely noise reduction ratio (NRR), intelligibility frequency weighted segmental SNR (IFWSNRseg), perceptual evaluation of speech quality (PESQ) and short-time objective intelligibility (STOI). Through the evaluation, the effectiveness of the Neural Networks can be observed.

In this paper, a modified a priori SNR estimation method with application in speech enhancement is presented. The decision directed (DD) approach for the a priori SNR estimation has been the most popular method due to its ability to eliminate musical noise in speech enhancement. However, a limitation of this method is the slow tracking of speech onsets that generates a transient distortion to the speech. The proposed method modifies the DD a priori SNR estimator by employing an adaptive smoothing factor based on the instan-taneous a posteriori SNR and an adaptive sigmoid function to obtain an improved tracking of speech onsets. Experimental results using several instrumental measures show the ability of the proposed method to preserve weak speech components when compared to two reference methods while maintaining the advantage of the DD approach in eliminating the musical noise.

The paper proposes a convex combination fusion function based on a sigmoid function for the estimation of the a priori SNR in a speech enhancement framework with critical frequency band processing. The proposed method does not only eliminate the one frame delay generated by the well-known decision directed approach but also increases the adaptation speed during abrupt changes in the SNR estimation. As a result, the advantage of low musical noise has been maintained while more weak speech components have been preserved. Experimental results using instrumental and subjective measures also indicate improvement in speech quality compared to the reference methods.

In this letter, a hybrid descent method is used to determine a set of filter parameters for a sigmoid filter which attempts to work under various SNR conditions. It overcomes the limitations of the current sigmoid filters that performs effectively only at a single SNR. Results show that significant improvement in terms of better speech qualities can be achieved by the proposed sigmoid filter when working under various SNR conditions.

In this paper, a new approach is proposed to improve a sigmoid and conditional smoothing-based speech presence probability (SPP) method for noise power spectral density (PSD) estimation. In this approach, the a posteriori speech absence probability (SAP) is adapted with a sigmoid function mapped to the normalised spectral average variance in the consecutive frames that can effectively characterise noise variation. The adaptation is also employed in the conditional smoothing stage to characterise the a posteriori SPP, which is then utilised in the noise PSD estimation. Comparison with state of the art methods validates the effectiveness of the proposed method.

In industrial noise environments, the use of assistive listening headsets is a means to provide adequate access to voice communication while wearing hearing protection. This paper presents a performance evaluation and comparison of two different methods to provide the binaural speech enhancement in real industrial noise scenarios. The investigated binaural methods based on differential beamforming and multichannel Wiener filter show different strengths and weaknesses. A transient noise suppression algorithm is also proposed and evaluated. Performance evaluation shows that this algorithm, together with the binaural multi-channel Wiener filter approach, can successfully reduce the hammering noise. This can be observed from the PESQ scores and the signal characteristics.

This paper presents an adaptive coherence-based beamforming algorithm that enhances a target speech signal from an arbitrary direction of arrival (DOA) in the azimuthal plane, while localising and suppressing the interfering speech signal and noise from other directions. In these complex noisy acoustic environments, spatially selective speech enhancement schemes provide significant advantages over single channel enhancement methods. Experimental results demonstrate that the proposed method achieves considerable improvement in Perceptual Evaluation of Speech Quality (PESQ). Moreover, the proposed algorithm is directly suited for real-time implementation.

The real-time implementation of the existing multi-channel Wiener filter (MWF) algorithms suffer from performance degradation due to the lack of robustness against estimation errors of the second-order statistics. The reasons are twofold: one, the estimation of the statistics relies on real voice activity detector (VAD), which often fails in adverse environments. Second, the MWF solutions involve estimation of the second order clean speech statistics, which also exaggerates the errors. This paper presents an MWF algorithm that requires neither VAD nor clean speech statistics. Performance evaluation under real scenarios shows that the proposed method outperforms the conventional MWF solution in terms of the trade-off between noise reduction and speech distortion.

In this paper, a modified a priori SNR estimator is proposed for speech enhancement. The well-known decision-directed (DD) approach is modified by matching each gain function with the noisy speech spectrum at current frame rather than the previous one. The proposed algorithm eliminates the speech transient distortion and reduces the impact from the choice of the gain function towards the level of smoothing in the SNR estimate. An objective evaluation metric is employed to measure the trade-off between musical noise, noise reduction and speech distortion. Performance is evaluated and compared between a modified sigmoid gain function, the state-of-the-art log-spectral amplitude estimator and the Wiener filter. Simulation results show that the modified DD approach performs better in terms of the trade-off evaluation.