Kacem Abida

A novel template matching approach to speaker-independent arabic spoken digit recognition

In this paper we propose a quantized time series algorithm for spoken word recognition. In particular, we apply the algorithm to the task of spoken Arabic digit recognition. The quantized time series algorithm falls into the category of template matching approach, but with two important extensions. The first is that instead of selecting some typical templates from a set of training data, all the data is processed through vector quantization. The second extension consists of a built-in temporal structure within the quantized time series to facilitate the direct matching, instead of relying on time warping techniques. Experimental results have shown that the proposed approach outperforms the time warping pattern matching schemes in terms of accuracy and processing time.

Systems combination in large vocabulary continuous speech recognition

We present a review of the most significant advances in the field of system combination towards reducing word error rates (WER) in large vocabulary continuous speech recognition (LVCSR). We have mainly focused on Recognizer Output Voting Error Reduction (ROVER), confusion network (CN) and minimum frame word error rate (fWER) based combination along with the latest improvements. Despite lot of progress witnessed in this field, some challenges still remain in enhancing the performance of LVCSR. We suggest in this paper some directions that may lead to a lower WER within the framework of system combination.

Combination of error detection techniques in automatic speech transcription

Speech recognition technology has been around for several decades now, and a considerable amount of applications have been developed around this technology. However, the current state of the art of speech recognition systems still generate errors in the recognizers output. Techniques to automatically detect and even correct speech transcription errors have emerged. Due to the complexity of the problem, these error detection approaches have failed to ensure both a high recall and a precision ratio. The goal of this paper is to present an approach that combines several error detection techniques to ensure a better classification rate. Experimental results have proven that such an approach can indeed improve on the current state of the art of automatic error detection in speech transcription.

Comparison of GMM and fuzzy-GMM applied to phoneme classification

The increasing need for more natural human machine interfaces has generated intensive research work directed toward designing and implementing natural speech enabled systems. Because it is very hard to constrain a speaker when expressing a voice-based request, speech recognition systems have to be able to handle out of vocabulary words in the users speech utterance. In this paper, we investigate an approach that can be deployed in keyword spotting systems. We propose a phoneme classifier that will be ultimately used to provide confidence values to be compared against existing Automatic Speech Recognizer word confidences. The end goal is to build a keyword spotting system for natural language speech. The presented approach is based on fuzzy gaussian mixture modeling to carry out the English phonemes classification.

Features' Weight Learning towards Improved Query Classification

This paper is an attempt to enhance query classification in call routing applications. We have introduced a new method to learn weights from training data by means of regression model. In this work, we have tested our method with tf-idf weighting scheme, but the approach can be applied to any weighting scheme. Empirical evaluations with several classifiers including Support Vector Machines (SVM), Maximum Entropy, Naive Bayes, and K-Nearest Neighbor (KNN) show substantial improvement in both macro and micro F1 measure.