Adaptive Beamforming With a Minimum Mutual Information Criterion

Abstract

In this work, we consider an acoustic beamforming application where two speakers are simultaneously active. We construct one subband-domain beamformer in generalized sidelobe canceller (GSC) configuration for each source. In contrast to normal practice, we then jointly optimize the active weight vectors of both GSCs to obtain two output signals with minimum mutual information (MMI). Assuming that the subband snapshots are Gaussian-distributed, this MMI criterion reduces to the requirement that the cross-correlation coefficient of the subband outputs of the two GSCs vanishes. We also compare separation performance under the Gaussian assumption with that obtained from several super-Gaussian probability density functions (pdfs), namely, the Laplace, K0, and Γ pdfs. Our proposed technique provides effective nulling of the undesired source, but without the signal cancellation problems seen in conventional beamforming. Moreover, our technique does not suffer from the source permutation and scaling ambiguities encountered in conventional blind source separation algorithms. We demonstrate the effectiveness of our proposed technique through a series of far-field automatic speech recognition experiments on data from the PASCAL Speech Separation Challenge (SSC). On the SSC development data, the simple delay-and-sum beamformer achieves a word error rate (WER) of 70.4%. The MMI beamformer under a Gaussian assumption achieves a 55.2% WER, which is further reduced to 52.0% with a K0 pdf, whereas the WER for data recorded with a close-talking microphone is 21.6%.

Filter bank design based on minimization of individual aliasing terms for minimum mutual information subband adaptive beamforming

“…Based on the average speaker position estimated for each utterance, a beamformer was constructed. The active weights were estimated so as to achieve the minimum mutual information (MMI) of the outputs from the beamformers [4]. In this work, we assumed that subband snapshots were Gaussian-distributed.…”

“…We also discuss the performance limitation of our filter banks due to numerical problems caused by singular matrices, and propose an alternate solution for the special case which can eliminate not only the total response error but also residual aliasing distortion completely. The filter banks proposed here are applied to minimum mutual information (MMI) beamforming where the active weight vectors are estimated so that mutual information of two beamforming outputs is minimized [4]. After that, the separated speech is further processed with Zelinski post-filtering and binary masking [5] in order to remove diffuse noises and a residual interference signal.…”

A Neural Network Based Regression Approach for Recognizing Simultaneous Speech

“…The speaker location information is used in a GSC-configured beamformer with a minimum mutual information (MMI) criterion [13] to separate the speech of different speakers. Assuming there are two such beamformers aimed at different sources as shown in Figure 2, the output of the i-th beamformer for a given subband can be expressed as,…”

“…In keeping with the GSC formalism, w q,i is chosen to preserve a signal from the look direction and, at the same time, to suppress an The optimization procedure of finding that w a,i under a minimum mutual information (MMI) criterion is described in Kumatani et al [13].…”

“…However, in the case of overlapping speech (with coherent noise), the estimation of coherence matrix is far from trivial, and inaccurate estimations may consequently introduce artifacts into the reconstructed signal. Such disadvantages to conventional beamforming have spurred the development of approaches such as the MMI beamforming criterion for beamforming [13] which alleviates the signal cancellation problem by ensuring the orthogonality of desired and interfering signals.…”