Efficient frequency-domain realization of robust generalized, sidelobe cancellers

This paper deals with hands-free acoustical front-ends for human/machine interaction in competing talker situations using efcient frequency-domain implementations of a robust Generalized Sidelobe Canceller (GSC). A new scheme is devised which leads to a factor of three of computational savings. The GSC tracking behavior is not impaired as the block lengths are kept suÆciently small. INTRODUCTION With the need for natural and comfortable means of human-machine interactions gaining critical signi cance, speech-driven application control, videoconferencing, and many other multimedia services call for high-quality handsfree acoustical interface technologies that allow the user to move freely without wearing or holding any speech capture device. Even with rapidly increasing computational power, computational complexity remains an important aspect, especially when no additional hardware is used and the speech processing algorithms run directly on the processor in the background. For optimum quality, the signals of interest should be free from any kind of impairment, i.e. noise, reverberation, and interferences. Microphone arrays in conjunction with robust adaptive beamforming techniques [1, 4] promise to meet these demands. Here, the robust Generalized Sidelobe Canceller (GSC) after [2] provides both high rejection of local interferences and suÆcient quality of the desired signals. Thus far, only time-domain implementations of that adaptive beamformer have been considered, even though more eÆcient block convolution-based frequency-domain adaptive lters (FDAFs) are known [3]. This contribution examines constrained and unconstrained frequency domain versions of the robust GSC. First, the block-based beamforming algorithm is described. It is shown how the adaptive modules of the GSC can be arranged for minimum computational complexity. The second part compares this arrangement to the conventional time-domain implementation. The nal part illustrates the performance by simulations. This work was supported by a grant from Intel Corp., Hillsboro, OR. ROBUST GSC IN THE FREQUENCY DOMAIN To prevent circular convolution e ects [3], the frequency-domain GSC uses the overlap-savemethod for partitioning and reassembling the data. Although both constrained and unconstrained FDAF algorithms are considered in later sections, we only describe the unconstrained adaptation algorithm here. The constrained version can be obtained by implementing additional constraints in the lter update equations [3]. Fig. 1 shows the conventional robust time-domain GSC, which is transformed into the frequency domain in the following. The description starts with the xed beamformer (FBF) that is implemented in the time domain. We then discuss the frequency-domain adaptive sidelobe cancelling path, i.e. the Adaptive Blocking Matrix (ABM) and the Adaptive Interference Canceller (AIC), and show that the GSC modules can be eÆciently combined, which reduces the computational complexity considerably.