F5-TTS: Fine-Tuned Arabic Speech Synthesis Model
Improved version of the model can be found here https://huggingface.co/IbrahimSalah/Arabic-F5-TTS-v2
Update 1
the model can not produce high quality audio, the main goal of it is to provide a chekpoint to continue the fine tunning process from. when using please set use_ema=False as it affects the quality. Here is a notebook to try it quickly : https://colab.research.google.com/drive/1kX7HB05CouHa5A-4Wy0UPqMuW4APqDBr?usp=sharing
Update 0
three checkpoints has bee added to the repo , the 380000 is the last , more data is needed to get better results so I will stop the fine tunning till getting more data and then i will proceed the fine tunning.
Overview
This project fine-tunes the F5-TTS model for high-quality Arabic speech synthesis, incorporating regional diversity in pronunciation and accents. The fine-tuning process is ongoing, and temporary checkpoints are provided as progress updates. Future iterations will include improved models with enhanced accuracy and naturalness.
Samples for now
'''
1- "لكن على ما يبدو ان هناك تصاعد غير مسبوق للاحداث."
2- "لذلك يجب علينا الإتحاد فى وجه كل الصدامات التى قد تؤثر علينا."
3- "كان هناك الكثير من التحديات للوصول إلى الدقه المطلوبة."
''' 1-
2-
3-
License
This model is released under the CC BY-NC 4.0 license, which allows free usage, modification, and distribution for non-commercial purposes.
Datasets
Training is based on the MBZUAI/ClArTTS so basically the model support MSA
Model Information
- Base Model: SWivid/F5-TTS
- Current Status: Ongoing fine-tuning (Temporary Checkpoints Available)
- (Final training parameters will be updated upon completion of fine-tuning.)
Usage Instructions
To use the fine-tuned Arabic model, follow these steps:
Usage
- GitHub Repository: Follow the F5-TTS setup instructions, but replace the default model with the Arabic checkpoint and vocabulary files provided here.
Contributions & Collaboration
This model is a work in progress, and community contributions are highly encouraged! Suggestions, improvements, and dataset contributions are welcome to refine its performance across different Arabic dialects.
Recommendations for Better Results
- Use clear reference audio with minimal background noise.
- Ensure balanced audio levels for improved synthesis quality.
- Contributions in dataset expansion and model evaluation are highly valuable.
Acknowledgment
- This work is done using Zewail City of science and technology machine
If you have any questions or suggestions, feel free to reach out! 🚀