Eigen-MLLR environment/speaker compensation for robust speech recognition

However, it is generally difficult if not impossible to prepare a complete set of a priori noisy environment knowledge. Especially, the noisy environments which are not seen in the training phase (aka unseen noisy environment) may become potential sources of serious performance degradation for those non-blind methods. In other words, it is crucial how to well organize and efficiently utilize the a priori noisy environment knowledge. To efficiently take the advantage of a priori noisy environment knowledge and, at the same time, alleviate the problem of unseen noisy environments and speakers, an Eigen-MLLR method originally proposed only for speaker adaptation [7-8] is adopted in this paper. In this paper an eigen-maximum likelihood linear regression (Eigen-MLLR) method is proposed to utilize a set of a priori noisy environment/speaker knowledge to online compensate the characteristics of unknown test environment/speaker. This idea is straightforward but is motivated from our recent findings that both the characteristics of different kinds of noisy environments and speakers could be simultaneously well organized in a PCA-constructed Eigen-MLLR subspace. Especially, the first three dimensions of the constructed EigenMLLR subspace are highly related to the SNR value, gender and type of noise. The proposed Eigen-MLLR was evaluated on Aurora 2 multi-condition training task. Experimental results showed that average word error rate (WER) of 6.14% was achieved. Moreover, Eigen-MLLR not only outperformed the multi-condition training baseline (Multi-Con., 13.72%) but also the blind ETSI advanced DSR front-end (ETSI-Adv., 8.65%), the histogram equalization (HEQ, 8.66%) and the non-blind reference model weighting (RMW, 7.29%) approaches. This idea is straightforward (from the viewpoint of speaker adaptation) but is motivated from our recent finding [5] that both the characteristics of different kinds of noisy environments (represented by a set of MLLR super-matrices) could be simultaneously well organized in a PCA-constructed Eigen-MLLR subspace. Especially, the first three dimensions of the constructed Eigen-MLLR subspace are highly related to the SNR value, gender and type of noise. It is therefore possible to (1) analyze a set of environment/speaker characteristics collected from all seen noisy environments/speakers in the training phase to construct an Eigen-MLLR environment/speaker subspace, and to then (2) optimally estimate (in the sense of maximum likelihood) the characteristics of the unknown test noisy environment/speaker in the test phase to compensate the HMMs of the ASR engine.