Online Maximization of Subband Kurtosis for Blind Adaptive Beamforming in Realtime Speech Extraction

Abstract

This paper presents a method for blind beamforming with application in realtime speech extraction in a non-stationary environment. The blind beamforming is carried out using an online kurtosis maximization approach where the optimization is based on Newton's method. The main novelty of the paper lies in the formulation of the subband kurtosis approximation, where a locally quadratic criterion is solved at each iteration. Further, a real-time digital signal processor (DSP) implementation of the method is conducted and results with real data is presented.

Online blind speech extraction based on a locally quadratic kurtosis criteria and a preprocessing Automatic Gain Controller

“…This setting allows for comparison with the recently proposed method [3]. The realtime implementation is analyzed using short-term power estimates where the algorithm may be switched on and off.…”

“…The transformation approach reduces the number of computations in this DSP platform by 7.5 times for a typical vector operation [11]. The same hardware platform was used in [3] to realize the predecessor to the proposed blind beamformer.…”

“…The improvement proposed in this paper regards a pre-processing Automatic Gain Control (AGC) stage that normalizes the input signals prior to the adaptive beamformer's filter update equation, thus yielding a scale-invariant solution. The intended application is Blind Speech Extraction (BSE) where a dominant speech source (dominant in the kurtosis measure) is extracted from an observed convolutive mixture of sources [3,4,5,6]. The proposed approach stands in contrast to Blind Signal Separation (BSS) in which all dominant sources, or groups of sources, are separated, see for example [7].…”

“…The merging of an adaptive beamformer with a blind control algorithm results in a structure that continuously tracks sources in a changing environment [2,3]. This paper presents an improvement to a recently proposed approach of approximating subband kurtosis using a locally quadratic criteria [3]. The improvement proposed in this paper regards a pre-processing Automatic Gain Control (AGC) stage that normalizes the input signals prior to the adaptive beamformer's filter update equation, thus yielding a scale-invariant solution.…”

“…The inherent virtue of a blind control algorithm in beamforming is that no knowledge about the spatiotemporal environment is needed, such as the position of the sources relative to the microphone array, or knowledge regarding the physical dimension of the array itself. The merging of an adaptive beamformer with a blind control algorithm results in a structure that continuously tracks sources in a changing environment [2,3]. This paper presents an improvement to a recently proposed approach of approximating subband kurtosis using a locally quadratic criteria [3].…”