Research

We publish what we learn.

We recorded, cleaned and aligned 10.5 hours of studio-quality Tunisian Arabic speech — then open-sourced every checkpoint.

A cross-dialect NER benchmark covering MSA, Tunisian, Egyptian and Moroccan Arabic, achieving state-of-the-art F1 on low-resource dialects.

Techniques for training competitive TTS models with fewer than 2 hours of data on commodity hardware — presented at Interspeech 2023.

RAG architecture that injects compliance rules at inference time, keeping LLM outputs inside legal and regulatory boundaries.

Ready to start?Reply · < 1 business day

Tell us about your challenge. We'll reply within one business day with a concrete starting point.

ExploreBuildDeploy