Loubna1101/Qwen3-TTS-Darija-LoRa

# Qwen3-TTS-Darija-LoRa A Darija Moroccan Arabic text-to-speech project based on **Qwen3-TTS-12Hz-1.7B-Base**, fine-tuned with **LoRA** on atlasia/DODa-audio-dataset F3. ## Overview This project fine-tunes Qwen3-TTS on a Darija dataset and provides inference through a simple Gradio interface. ## Base Model - `Qwen/Qwen3-TTS-12Hz-1.7B-Base` ## Hugging Face - 'loubna1101/Qwen3-TTS-Darija-LoRa' ## Fine-tuning Method - LoRA fine-tuning - Single-speaker custom voice - 24kHz audio - Speaker name used during training: `darija_speaker` ## Dataset Preparation The training pipeline: 1. Resamples audio to 24kHz 2. Builds `train_raw_24k.jsonl` 3. Extracts codec tokens with `prepare_data.py` 4. Fine-tunes Qwen3-TTS with LoRA ## Training Setup - Batch size: `1` - Learning rate: `5e-5` - Epochs: `10` ## Repository Contents - `finetuning.ipynb`: training workflow - `inference.ipynb`: inference workflow - `app.py`: Gradio demo - `requirements.txt`: dependencies ## Model Files The trained LoRA adapter and custom speaker embedding are hosted on Hugging Face. ## Inference The model is loaded as: - base model from Qwen - LoRA adapter from this project - custom speaker embedding from this project ## Authors - loubna haouach - chaimae haddouche ## Disclaimer This project is for research and educational purposes. Please ensure you have the right to use and publish the voice data used for fine-tuning. ## Credits - Qwen team for Qwen3-TTS - atlasIA for the Doda dataset - Hugging Face ecosystem