A strong trend in robotics is the investigation of adaptable machine learning algorithms and frameworks that enhance the skills and application of artificial systems during the interaction with humans. The use of language is one of the most convenient human methods to communicate with artificial agents. In the last decades, the introduction of automatic speech recognition (ASR) systems achieved important advances in the field, however the simple and natural style of how parents teach speech to their children is still an open question under investigation. With the goal of improving the interaction and learning process with artificial agents, these must be able to increase their vocabulary (acquire novel terms) in a satisfying manner (rapid and adequate) for the user/tutor. In this work, we introduce an incremental word learning system1 to enhance speech acquisition in artificial agents. Here, different word-models are successively learned in a framework that possesses little prior knowledge and is inspired by the human infants language acquisition process. In order to build a user-friendly system that requires a low tutoring time our approaches are built to cope with a small number of training samples. Relative to the used architecture, we employ a hidden Markov model (HMM) framework similar to most ASR systems. Although HMMs are a powerful tool, for obtaining good recognition scores, special attention is required for the quantity of samples employed, the bootstrapping method and the performance of the discriminative training techniques integrated in the framework. Therefore, we present several procedures to overcome these challenges. A main drawback of employing few training data samples is the overspecialization of the learned models, which complicates the recognition of unseen items. In this context, we propose a novel computation of a parameter, which is adapted according to the amount of provided data samples, and analyze different influences for limited learning data. Afterwards, we describe the proposed initialization technique introduced in the system to properly bootstrap the estimates of the newly created model. In this approach, we combine unsupervised and supervised training methods with the following re-building of the model through a multiple sequence alignment method, which arranges and incorporates the succession of hidden states obtained by the Viterbi decoding algorithm. Next, several large margin (LM) discriminative training methods are analyzed to increase the generalization performance of the previously created models, i.e. improving the classification of the models. Here, we propose different procedures appropriate for employing limited data in discriminative training. Finally, the proposed methods are compared against state-of-the-art techniques during the experimental phase. Similarly, each individual contribution of the introduced approaches is measured in relation to the global yielded improvement of the whole framework. Additionally, we examine a potential decrease of the amount of training samples with the purpose of decreasing the time employed to teach a robot. The evaluation of our approaches is realized on different recognition tasks containing isolated and continuous digits. After all, we demonstrate that the introduced collection of techniques achieve important improvements so that they represent a significant step towards efficient incremental word learning with limited data. 1Based on the previous works enumerated in Appendix D.

Towards natural speech acquisition: incremental word learning with limited data

Single-class support vector machines(SC-SVMs) are different from the standard two-class SVM techniques. This algorithm is used to allow the pass through of the only positive data by treating the origin as the only member of the second class. Here, we present an algorithm for rejecting out-of-vocabulary (OOV) phrases. Our method uses SC-SVM mechanism to detect and reject the phrases of OOV. We also depict the trends in in-vocabulary(IV) correct recognition and OOV rejection with the number of command increasing and give an analysis.

Single-class support vector machine for an out-of-vocabulary rejection of isolated words

The cChromic shaped bricks have been used as the refractory lining in alumina kiln for long time, whose service life is about 120~250 days. Besides the short working time, the environmental pollution led to by the chromic component is also serious. In this thesispaper (thesis 是硕士论文), some materials are were chosen based on the actual situation of the usage of the refractory lining in the current alumina kiln. The choosing principals are were the extensive source and the relatively low price. Finally, a new formula of the refractory castable is was acquired based on the orthogonal test and has passed the properties testing.

Preparation of Refractory Castable for High-Temperature Zone of the Alumina Kiln

With the increased use of speech recognition outside of the laboratory environment, the need for better out-of- vocabulary (OOV) rejection techniques is critical for the continued success of this user interface. In this paper, we propose a novel rejection method for isolated word recognition. Unlike the traditional method which is based on the use of garbage models, our proposed method does not need using garbage models explicitly. Therefore it does not need additional memory requirement and there is no increase in computation. So the proposed method is especially suitable for resource limited applications. Results are presented for a database consisting of 33 speakers (13 speakers for training and 20 speakers for testing) and 50 short words (10 in vocabulary words and 40 out of vocabulary words). Performance shows that the proposed method is superior to the traditional rejection method which based on the use of garbage models.

A novel out-of-vocabulary rejection method for isolated word recognition

In this paper, we present a class of embedding methods for nonsmooth equations. Under suitable conditions, we prove that there exists a homotopy solution curve, which is unique and continuous. We also prove that the solution curve is singlevalue-d with respect to the homotopy parameter. Then we construct an efficient algorithm for this class of equations and prove its convergence. Finally, we apply the algorithm to the nonlinear complementarity problem. The numerical results show that the algorithm is satisfacotry.

The embedding method for nonsmooth equations

In this paper, an effective two-phase approach is proposed for removing salt-and-pepper noise in color images. In the first phase, an adaptive median filter is used to identify color pixels which are likely to be contaminated by noise, and in the second phase, the color image is restored by using an edge-preserving regularization method that applies only to those identified color pixels. Experimental results indicate that our proposed algorithm have good performance, and provide some improvements over many existing methods in the literature.

An effective algorithm for salt-and-pepper noise removal in color images

Restoration of bi-level images is generally performed in two steps, i.e., restoration and thresholding seg- mentation, which generally does not produce ideal results. This paper proposes a nonlinear regularization method for bi-level image restoration by explicitly using the knowledge of the bi-level image. Unlike the Tikhonov regularization method which eventually results in a linear system of equations, the proposed method leads to a nonlinear optimiza- tion problem, solved with the global Barzilai and Borwein gradient method. Simulation results show feasibility and eectiveness of the method.

Zhang Jian-jun

Papers

A novel out-of-vocabulary rejection method for isolated word recognition

The embedding method for nonsmooth equations

An effective algorithm for salt-and-pepper noise removal in color images

Nonlinear Regularization for Bi-level Image Restoration