Harmonic beamformers for speech enhancement and dereverberation in the time domain

Harmonic Beamformers for Non-Intrusive Speech Intelligibility Prediction

“…Even though the adaptive beamformer in theory should have a better performance this is not necessarily the case in practical performance, since it relies on estimates of the noise statistics. This is also supported by the findings in [18], where the adaptive beamformers provide a slightly lower SNR gain compared to the harmonic DSB. As such, due to being computational efficient and simple, i.e.…”

“…In the following, the signal model and the associated assumptions of the proposed methods are presented based on [18] in which a more thorough description of the theory behind the harmonic beamformers is available. In the proposed methods it is assumed that a uniform linear array (ULA) consisting of M microphones obtains the desired speech signal added to a mixture of interfering background noise and reverberation such that the samples of the mth microphone observations in a vector of frame length L is given by:…”

“…Furthermore, using the harmonic model to obtain the desired signal does not only reduce the interfering sources but also reverberation, since spectral and temporal smearing of the signal source due to reverberation is not included in the harmonic model. As such, the desired speech signal is modeled as [18,19]:…”

“…The harmonic delay-and-sum beamformer (DSB) is a fixed beamformer, which cannot adjust to the spatial characteristics of the interfering noise. It is advantageous for applications such as hearing aids, since it only requires low computational power and does not require estimates of the noise statistics but, at least in theory, comes at a cost in performance [18]. The DSB can be deduced by maximizing the white noise gain subject to the distortionless constraint:…”

“…The harmonic Wiener beamformer can be derived using the mean square error (MSE), which is given by [18]:…”