Acoustic-support vector machines approach to detect spoken Arabic language

Abstract

Spoken Language detection is the process of either accepting or rejecting a language identity from its sample speech. The process is essential as it represents the first phase for a complete multilingual-enabled speech processing applications. However, most efforts are focused on European languages and the research is relatively few for other languages such as Arabic. This is mainly due to the lack of tools and resources, e.g., Arabic speech corpora. Furthermore, the majority of the proposed approaches for Arabic detection are language-dependent rather than independent ones, in which the model uses only acoustic properties of speech signal. This paper describes an ongoing research to develop a language independent Modern Standard Arabic (MSA) detector, which is a binary Support Vector Machines (SVM) classifier that is based on speech acoustic features. In that context, the classifier is used to classify speech utterance into either classA, which represents the Arabic language or classNA to denote non-Arabic languages. As most currently available speech corpora are license restricted and their languages are selected based on population or geographical distribution, a new multilingual speech corpus with six languages is being created. Languages in this created corpus have some sort of similarity with MSA, e.g., Arabic and Hebrew. This property adds another dimension of complexity to the classification task, but it is essential as one of the major goal of this research is to measure whether the efficiency of the MSA model will be preserved on the same level when tested with other languages that have some sort of relationship with the MSA or other Arabic dialect. This will be referred to in this paper as stability-against-similarity of the model.