A Deep Learning Model for Western Neo-Aramaic Speech Recognition

The automatic speech recognition (ASR) task has recently become one of the crucial branches of the natural language processing field owing to the wide range of such models’ applications. Many companies develop voice assistants, audio translators, and transcribers for major languages. Still, there is no doubt that ASR systems are just as useful for low-resource languages, particularly for fieldwork purposes. In the framework of this project, we propose an ASR deep learning model for the Maaloula dialect of the Modern Western Aramaic language (MWA). MWA is a minor language spoken in the suburbs of Damascus, Syria. There are three villages where the language can be found: Bakh'a, Jubb'adin, and Maaloula, each having its dialectal differences. Current work focuses on the latter for it being best documented so far. However, we would also consider the other two in future work. Developing the system includes choosing an initial model to fine-tune and prepare for our language, collecting the data, the learning process, and evaluation. Our work involves data augmentation, i. e. artificial generation of additional data and training a language model to enhance the model’s performance.