Developed an experimental AI speech-generation

Developed an experimental AI speech-generation pipeline focused on Libyan Arabic, addressing the lack of suitable off-the-shelf voices and public speech datasets for the dialect. The project aimed to enable more natural Libyan Arabic voice interactions for applications such as AI voice assistants and automated calling systems.

The project included a dialect-specific speech dataset, transcript alignment and normalization, a Libyan Arabic lexicon, recording-script validation, and fine-tuning an open-source text-to-speech model. It also incorporated speech-quality filtering and voice-similarity evaluation to improve output quality while avoiding imitation of real speakers. The result was a research prototype with measurable progress and further improvements planned.