Automatic detection of code-switching in Arabic dialects

Abstract

Multilingual and multidialectal speakers commonly switch between languages and dialects while speaking, leading to the linguistic phenomenon known as code-switching. Most acoustic systems, such as automatic speech recognition systems, are unable to robustly handle input with unexpected language or dialect switching. Generally, this results from both a lack of available corpora and an increase in the difficulty of the task when applied to code-switching data. This thesis focuses on constructing an acoustic-based model to gather code-switching information from utterances containing Modern Standard Arabic and dialectal Arabic. We utilize the multidialectal GALE Arabic dataset to classify the code-switching style of an utterance and later to detect the location of code-switching within an utterance. We discuss the failed classification schemes and detection methods, providing analysis for why these approaches were unsuccessful. We also present an alignment-free classification scheme which is able to detect locations within an utterance where dialectal Arabic is likely being spoken. This method presents a marked improvement over the proposed baseline in average detection miss rate. By utilizing this information, Arabic acoustic systems will be more robust to dialectal shifts within a given input.