Investigations on speech recognition systems for low-resource dialectal Arabic–English code-switching speech

Code-switching (CS), defined as the mixing of languages in conversations, has become a worldwide phenomenon. The prevalence CS been recently met with growing demand and interest to build automatic speech recognition (ASR) systems. In this paper, we present our work on code-switched Egyptian Arabic–English ASR. We first contribute filling huge gap resources by collecting, analyzing publishing spontaneous corpus. ASR systems using DNN-based hybrid Transformer-based end-to-end models. thorough comparison between both approaches under setting low-resource, orthographically unstandardized, morphologically rich language pair. show that while achieve comparable overall results, have complementary strengths. can be improved combining outputs two propose several effective system combination approaches, where hypotheses are merged sentence- word-levels. Our result WER relative improvement 4.7%, over baseline performance 32.1% WER. case intra-sentential sentences, 4.8%. best performing achieves 30.6% ArzEn test set.