🎙️ Mmeslay

An Automatic Speech Recognition System for the Kabyle language with a Flutter frontend.

• Live Demo Available at https://huggingface.co/spaces/g1ya/Mmeslay

---

📦

The backend Code has been moved to this repository : https://github.com/G1ya777/Mmeslay_backend-CLI

📌 Overview

• This was created by training the Squeezeformer-XS model using the Common Voice dataset (Kabyle subset).
• The model was trained, validated, and tested on a custom split of the dataset.
• A language model was also trained on a text corpus composed of sentences collected from various sources, such as:
  • Tatoeba
  • https://github.com/MohammedBelkacem/Kabyletexts
    using KenLM.
• The system was tested using various configurations of the CTC decoder:

---

📚 Cite Us

If you use our project in your work, please cite:

@mastersthesis{Mmeslay,
  author  = {Aomer Gaya Ouldali},
  school  = {Université A. Mira de Béjaïa},
  title   = {Système de reconnaissance de la parole appliqué à la langue Tamazight},
  year    = {2023}
}