Subband Likelihood-Maximizing Beamforming for Speech Recognition in Reverberant Environments

Abstract

Abstract-Speech recognition performance degrades significantly in distant-talking environments, where the speech signals can be severely distorted by additive noise and reverberation. In such environments, the use of microphone arrays has been proposed as a means of improving the quality of captured speech signals. Currently, microphone-array-based speech recognition is performed in two independent stages: array processing and then recognition. Array processing algorithms, designed for signal enhancement, are applied in order to reduce the distortion in the speech waveform prior to feature extraction and recognition. This approach assumes that improving the quality of the speech waveform will necessarily result in improved recognition performance and ignores the manner in which speech recognition systems operate. In this paper a new approach to microphone-array processing is proposed in which the goal of the array processing is not to generate an enhanced output waveform but rather to generate a sequence of features which maximizes the likelihood of generating the correct hypothesis. In this approach, called likelihood-maximizing beamforming, information from the speech recognition system itself is used to optimize a filter-and-sum beamformer. Speech recognition experiments performed in a real distant-talking environment confirm the efficacy of the proposed approach.

Multichannel speech recognition using distributed microphone signal fusion strategies

“…The weighting factors represented by equations (5), (6), (7), (8), and (9) were independently implemented. These are denoted "equal weighting," "inverse distance weighting," "inverse distance squared weighting," "signal amplitude weighting," and "signal power weighting.…”

“…McCowan and Sridharan [7] performed sub-band processing of microphone array signals for speech recognition by integrating dynamically-weighted models trained on each sub-array frequency bands based on sub-band speech energy. By also using microphone arrays, Seltzer, Raj, and Stern formulated full-band [4] and sub-band [8] beamforming methods for optimally combining microphone array signals to generate the sequence of features that maximize the likelihood of producing the correct hypothesis. In contrast, Shimizu, Kajita, Takeda, and Itakura [9] developed methods using a fixed sound source that perform speech recognition for each microphone and selects the highest likelihood or equally weights and combines the feature vectors from the microphones.…”

Reverberant speech recognition exploiting clarity index estimation

“…Additionally, ASR techniques robust to reverberation can be also classified according to the number of microphones used to capture the signal such as single-channel methods [13,20,26,29] or multi-channel techniques [12,27,30,31].…”

Feature mapping using far-field microphones for distant speech recognition

“…In case of multi-channel ASR, there have been studies on designing a beamformer with the aim of optimizing ASR performance. A technique such as likelihood maximizing beamforming (LIMABEAM) [4,5] specifically optimizes array parameters using gradient descent to maximize the likelihood of the recognized hypothesis under an ASR speech model, given the filtered acoustic data. Recent research on LIMABEAM suggests no significant improvement using the standard LIMABEAM on large vocabulary distant speech recognition on the AMI meeting corpus and it is recommended to use a better optimization strategy for any LIMABEAM implementation [6].…”