Towards One Model to Rule All: Multilingual Strategy for Dialectal Code-Switching Arabic ASR

With the advent of globalization, there is an increasing demand for\nmultilingual automatic speech recognition (ASR), handling language and\ndialectal variation of spoken content. Recent studies show its efficacy over\nmonolingual systems. In this study, we design a large multilingual end-to-end\nASR using self-attention based conformer architecture. We trained the system\nusing Arabic (Ar), English (En) and French (Fr) languages. We evaluate the\nsystem performance handling: (i) monolingual (Ar, En and Fr); (ii)\nmulti-dialectal (Modern Standard Arabic, along with dialectal variation such as\nEgyptian and Moroccan); (iii) code-switching -- cross-lingual (Ar-En/Fr) and\ndialectal (MSA-Egyptian dialect) test cases, and compare with current\nstate-of-the-art systems. Furthermore, we investigate the influence of\ndifferent embedding/character representations including character vs\nword-piece; shared vs distinct input symbol per language. Our findings\ndemonstrate the strength of such a model by outperforming state-of-the-art\nmonolingual dialectal Arabic and code-switching Arabic ASR.\n