On the performance of ensemble-based classifiers for Arabic speech recognition

Speech recognition continues to be a challenging research problem for diverse applications. The challenge is to develop better recognition systems, more robust, computationally more efficient, and versatile in nature. While western and eastern languages have attracted a lot of interest among researchers, the Arabic language, unfortunately, did not get an appropriate share of this interest. The Arabic language exhibits richness in semantics rarely found in other languages. To contribute to this field of work, we explore, in this paper, the aspect of combining evidences from multiple classifiers to improve accuracy of individual speech classification algorithms. The analysis covers fusion of evidence taken from different angles (perspectives) from statistical, to leaning, to evidence perspectives. Our experiments showed that ensemble-based classifiers achieve, on the average, an improvement in recognition accuracy of 4% or more, leading to overall recognition accuracies in the case of Arabic digits to more than 90%.