Beamforming With a Maximum Negentropy Criterion

Abstract

Abstract-In this paper, we address a beamforming application based on the capture of far-field speech data from a single speaker in a real meeting room. After the position of the speaker is estimated by a speaker tracking system, we construct a subband-domain beamformer in generalized sidelobe canceller (GSC) configuration. In contrast to conventional practice, we then optimize the active weight vectors of the GSC so as to obtain an output signal with maximum negentropy (MN). This implies the beamformer output should be as non-Gaussian as possible. For calculating negentropy, we consider the Γ and the generalized Gaussian (GG) pdfs. After MN beamforming, Zelinski postfiltering is performed to further enhance the speech by removing residual noise. Our beamforming algorithm can suppress noise and reverberation without the signal cancellation problems encountered in the conventional beamforming algorithms. We demonstrate this fact through a set of acoustic simulations. Moreover, we show the effectiveness of our proposed technique through a series of far-field automatic speech recognition experiments on the Multi-Channel Wall Street Journal Audio Visual Corpus (MC-WSJ-AV), a corpus of data captured with real far-field sensors, in a realistic acoustic environment, and spoken by real speakers. On the MC-WSJ-AV evaluation data, the delay-and-sum beamformer with post-filtering achieved a word error rate (WER) of 16.5%. MN beamforming with the Γ pdf achieved a 15.8% WER, which was further reduced to 13.2% with the GG pdf, whereas the simple delay-and-sum beamformer provided a WER of 17.8%. To the best of our knowledge, no lower error rates at present have been reported in the literature on this ASR task.

Bayesian Feature Enhancement for Reverberation and Noise Robust Speech Recognition

“…There, we also showed how the distribution of the observation error in the presence of reverberation and noise can be modeled by a GAUSSIAN with time-variant moments, eventually leading to the observation mappings (23) and (35) with the associated error statistics given in (26) and (37), respectively.…”

“…Further, (25) is the sequence of clean feature vectors of length . Finally, is an error term assumed to be a realization of a vector-valued GAUSSIAN stochastic process according to (26) In [12], the mean vector and the (diagonal) covariance matrix have been found to depend on the instanta-neous reverberant-to-noise ratio (IRNR) , which is defined by (27) rendering the error statistics time-variant. In particular, the mean and the variance of the observation error in the th mel band at time instant are given by (28) …”

“…The RNR at the input of a microphone (we employed the microphone denoted by "Array1-1" in [25]) of the circular array was about 15-20 dB on average. Estimates of the room reverberation time range from 380 ms [26] to 700 ms [25].…”

“…Several authors have published recognition results on this database [9], [25], [26]. While they mostly experimented with microphone array techniques to reduce the impact of reverberation, only [9] has been found to also work on the single-channel setup considered in this contribution.…”

Interspeech 2016

“…Compared to using recordings from single distant microphone only, beamforming is reported to reduce word error rate (WER) by 6-10% relative in large vocabulary conversational speech recognition tasks [1][2][3], and up to 60% relative in specific tasks [4,5]. Multi-channel dereverberation brings an extra 20% relative WER reduction over single channel dereverberation [6].…”