Joint optimization on decoding graphs using minimum classification error criterion

Abstract

Motivated by the inherent correlation between the speech features and their lexical words, we propose in this paper a new framework for learning the parameters of the corresponding acoustic and language models jointly. The proposed framework is based on discriminative training of the models' parameters using minimum classification error criterion. To verify the effective- I . I N T R O D U C T I O NVarious approaches of statistical learning in the field of machine learning have been extensively studied [1,2]. In machine learning, there are two main categories of learning algorithms for building pattern classifiers, namely generative and discriminative training algorithms. In generative training, the probability distribution of data points in each class is estimated using density estimation methods. The parametric modeling approach [3] is usually adopted to make the density estimation problem more feasible. The parametric modeling is based on an assumption that unknown probability distributions belong to some families of computationally tractable functions, such as the family of exponential distributions [4]. The unknown parameters of the presumed distribution are then estimated from training data using the common maximum-likelihood estimation (MLE) approach. The estimated distributions are then used for pattern classification, such as speech decoding based on the maximum a posteriori (MAP) decision rule. There are many efficient algorithms for training the generative models, such as expectation maximization (EM) [5,6] and Baum-Welch [7] algorithms. The advantage of generative training approach is the ability to exploit the inherent dependency among the training data samples. However, the limitation of generative training is the assumption of the distribution of the data set, which is not the actual distribution and accordingly results in a suboptimal performance of the generated classifier.On the other hand, discriminative training approach does not explicitly attempt to model the data distribution, but instead, it directly optimizes a mapping function from input samples to output labels [8,9]. Therefore, the discriminative training approach aims to adjust only the decision boundary without constructing a data generator from the entire feature space. In the literature, considerable research effort is applied to discriminative training for improving the speech recognition performance [10][11][12][13][14][15][16]. The most popular discriminative training criteria include: minimum classification error (MCE) [17][18][19], maximum mutual information (MMI) [20][21][22], minimum error rate training (MERT) [23], minimum phone/word error (MPE/MWE) [24], and minimum Bayes risk (MBR) [25][26][27][28][29]. While the MMI method uses mutual information as the criterion for maximization, all other methods attempt to reduce the empirical error by optimizing error rate related objective functions. Although discriminative training is advantageous over generative training in terms of the avoidance of the data distribution assumpti...