An Improved Method for Late-Reverberant Suppression Based on Statistical Model

• In this study, we proposed an improved method based on Late reverberant spectral variance (LRSV) estimator.
• At first, we extend the LRSV estimator to the noisy and reverberant environment, which also means the LRSV estimator is derived under this environment.
• A new method that estimator much smoother shape parameter (leading the LRSV expression) is used to demonstrate that a smoother parameter can give better performance.
• A novel method that combines a LRSV estimator with maximum likelihood (ML) algorithm in spectral domain is proposed.
• A new spectral process system that estimates the late reverberation before the background noise is developed.

Model-based late reverberant spectral variance (LRSV) estimator is considered as an effective approach for speech dereverberation, which can construct a simple expression for the LRSV according to the past spectral variance of the reverberant signal. In this paper, we develop a new LRSV estimator based on the time-varying room impulse responses (RIRs) with the assumption that the background noise is comprised of reverberant noise and direct-path noise in a noisy and reverberant environment. In the LRSV estimator, more than one item of past spectral variance of the reverberant signals are used to obtain a smoother shape parameter, which can lead to a better performance for dereverberation compared to the classic methods. Since this shape parameter affected by the estimation error of LRSV may in turn affect the subsequent LRSV estimation, we combine this smoother shape parameter based LRSV estimator with maximum likelihood (ML) algorithm in spectral domain in order to get a more reliable estimation of LRSV. Furthermore, we use the proposed LRSV estimator prior rather than posterior to speech enhancement in noisy and reverberant environment. Experimental results demonstrate our new LRSV estimator is more effective for both noise-free and noisy reverberant speech.