Muhammed-OTP/Hassaniya-Arabic-Text-To-Speech-System-Using-Transfer-Learning

Hassaniya Arabic Text-To-Speech System Using Transfer Learning # Hassaniya Arabic Text-To-Speech System Using Transfer Learning **Development of a Hassaniya Dialect Speech Synthesis System** *Master M1 — Artificial Intelligence* *Module: NLP Dialects* **Mohamed Salem Ebnou Echvagha Oubeid** | ID: C34613 June 2026 --- ## Overview This project presents a **proof-of-concept Text-To-Speech (TTS) system** for the **Hassaniya Arabic dialect** — the primary Arabic dialect spoken in Mauritania and parts of Western Sahara, Mali, and Senegal. Hassaniya is a **low-resource dialect** with very limited digital presence and almost no existing speech synthesis tools. This project explores the feasibility of building a TTS pipeline using **transfer learning** from pretrained Arabic TTS models, rather than training from scratch. ## Objectives - Build an end-to-end TTS pipeline for Hassaniya Arabic - Demonstrate transfer learning from pretrained Arabic speech models - Create a reusable preprocessing and annotation pipeline - Document challenges of working with low-resource dialects - Provide a foundation for future Hassaniya speech technology research ## Dataset | Property | Value | |----------|-------| | **Samples** | 294 audio recordings | | **Format** | Audio bytes + text transcriptions | | **Language** | Hassaniya Arabic (Mauritanian dialect) | | **Avg. text length** | ~33 characters | | **Source** | Collected Hassaniya speech samples | ## Methodology ### Pipeline Architecture ```mermaid graph LR A[Raw Text Hassaniya Arabic] --> B[Text Preprocessing Normalization & Cleaning] B --> C[Tokenization Character/Phoneme Level] C --> D[TTS Model Pretrained + Fine-tuned] D --> E[Mel Spectrogram Generation] E --> F[Vocoder Waveform Synthesis] F --> G[Generated Speech Audio Output] ``` ### Transfer Learning Strategy ```mermaid graph TD A[Pretrained Arabic TTS Model] --> B[Feature Extraction Layers Frozen] A --> C[Output Layers Fine-tunable] D[Hassaniya Dataset 294 samples] --> E[Fine-tuning Process] C --> E E --> F[Hassaniya TTS Mo …