Modern standard Arabic text to dialectal speech

Abstract

Artificial intelligence is all about imitating human behavior and logic. And perhaps speech is what represents the most natural human-like aspect to be integrated in machines. For this purpose, in my thesis, I dive deeper into dialect studies and develop a new Lebanese Arabic text-to-speech (TTS) system. The developed system consists mainly of two parts: diacritizing the input sentence and synthesizing the latter to a Lebanese speech using Tacotron, a TTS system. To satisfy this goal, two online datasets were collected and used in this project, one for diacritization and the other for the TTS. Since studies on dialects are rare, many issues and challenges were faced while developing the system especially while dealing with the datasets. Every part of the system was developed independently, and consequently, a model was generated for each one of the parts and later evaluated and improved as much as possible. Finally, the models were combined in order to obtain the desired final output. The obtained results were satisfactory given the fact that the datasets were limited in size and coverage.