Spoken Arabic Algerian Dialect Identification

Dialect identification is a challenging task and this becomes more complicated when dealing with under-resourced dialects. In this paper, we propose a system based on prosodic speech information, namely intonation and rhythm for identification of Intra-country dialects. The speech features are extracted after a coarse-grained consonant/vowel segmentation. Dialect models are built using both Deep Neural Networks (DNNs) and SVM. The hyper-parameters for the DNNs topology are tuned using a genetic algorithm. Our framework is implemented and evaluated on KALAM'DZ, a Web-based corpus dedicated to Algerian Arabic Dialectal varieties, with more than 42 h encompassing the four major Algerian subdialects: Hili, Sulaymite, Ma'qilian, and Algiers-blanks. The results show that the DNNs implementation of Algerian Arabic Dialect IDentification system (A2DID) reaches the same results when compared to SVM modeling. In addition, we concluded that a contrastive baseline acoustic-based classification system can serve as a complementary system to our A2DID. The overall results reveal the suitability of our prosody-based A2DID for speaker-independent dialect identification when utterances size are short. A requirement for real-time applications.