With advancements in deep learning, Text-to-Speech has achieved remarkable results for high-resource languages; however, low-resource languages, such as Moroccan Darija, remain underexplored. This work addresses this gap by implementing speech synthesis for Moroccan Darija by finetuning the state-of-the-art model FastSpeech 2, which is a non-autoregressive transformer-based model. We used a 2-h Darija speech dataset featuring a female voice. To enable effective model training, we designed an elaborated a data pipeline encompassing cleaning, diacritizing, transliterating to Buckwalter scripts, phoneticizing, and aligning the scripts with corresponding speech files. The synthesized speech was evaluated using the Mean Opinion Score (MOS) metric, which assesses the correctness and naturalness of the output. FastSpeech 2 achieved a MOS score of 3.905 on simple and short sentences. These results highlight the feasibility of leveraging deep learning to advance speech synthesis systems for Moroccan Darija, contributing also to broader linguistic inclusivity of low-resource languages in Natural Language Processing.