AhmedEladl/Magpie-TTS-Saudi-Arabic

Magpie TTS - Saudi Arabic (Fine-Tuned)

This model is a fine-tuned version of the NVIDIA Magpie TTS Multilingual (357M) model, specifically adapted for Saudi Arabic text-to-speech generation.

Audio Comparison: Base vs. Fine-Tuned

Below are raw, unedited audio samples demonstrating the dialect adaptation.

Text Prompt	Base Model magpie_tts_multilingual_357m	Fine-Tuned Model
ترى الطريق جهة الرياض اليوم مرة زحمة، فإذا مستعجلين أحسن نطلع من الحين عشان نوصل قبل الموعد وما نتأخر.
حياكم الله في أي وقت، وإذا وصلتوا علموني، وبجهز لكم القهوة والحلا ونجلس كلنا مع بعض ونسولف لين وقت متأخر.
من كثر ما يتغير الجو هالأيام، الواحد ما يدري وش يلبس، الصبح يكون الجو بارد شوي والعصر يرجع الحر مرة ثانية.
أنا أشوف نخلي المشوار بعد المغرب، لأن الجو بيكون أحسن والناس تخف من الشوارع، ونقدر ناخذ راحتنا في الطلعة.

---

Model Details

• Architecture: Magpie TTS (Encoder-Decoder Transformer with Audio Codec)
• Base Model: nvidia/magpie_tts_multilingual_357m
• Language: Arabic (Saudi Arabia) ar-SA
• Framework: NVIDIA NeMo (v2.8.0rc0)
• Codec: nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps

Training Dataset

The model was fine-tuned using the AhmedEladl/saudi-voice-dataset dataset.

• Audio Processing: Resampled to 22,050 Hz to match the NanoCodec requirements.
• Filtering: Utterances under 0.5 seconds were removed to ensure training stability.
• Context Selection: Randomized context reference audio from within the dataset to provide target voice conditioning.

Training Configuration

• Hardware: NVIDIA L4 / T4 GPU
• Epochs: 10
• Batch Size: 2 (with gradient accumulation steps = 2)
• Learning Rate: 5e-6 (AdamW)
• Precision: 32-bit float
• Context Duration: 2.0s to 12.0s

📦 Dependencies & Installation

To run inference with this model, ensure you have the correct versions of the NeMo toolkit and audio handling libraries installed.

pip install torch>=2.1.0
pip install soundfile>=0.12.1
pip install librosa>=0.10.1
pip install huggingface_hub>=0.23.0
pip install nemo_toolkit[all]==2.8.0rc0

Inference Code Example

import soundfile as sf
from huggingface_hub import hf_hub_download
from nemo.collections.tts.modules.magpietts_inference.utils import ModelLoadConfig, load_magpie_model

# 1. Download Model & Codec from Hugging Face Hub
print("Downloading models...")
model_path = hf_hub_download(repo_id="AhmedEladl/Magpie-TTS-Saudi-Arabic", filename="Magpie-TTS-Saudi-Female.nemo")
codec_path = hf_hub_download(repo_id="nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps", filename="nemo-nano-codec-22khz-1.89kbps-21.5fps.nemo")

# 2. Load the fine-tuned model and codec
config = ModelLoadConfig(nemo_file=model_path, codecmodel_path=codec_path)
model, _ = load_magpie_model(config)
model.eval().cuda()

# 3. Generate Audio
prompt = "الذكاء الاصطناعي صار جزء أساسي من حياتنا اليومية، وتطوير نماذج تدعم لهجتنا خطوة جداً مهمة."
print("Generating audio...")
res = model.do_tts(transcript=prompt, language="ar-SA", apply_TN=False)

# 4. Save the output
audio = res[0].cpu().numpy()
if len(audio.shape) == 2:
    audio = audio[0]
    
sf.write("output.wav", audio, 22050)
print("✅ Audio saved successfully to output.wav")

📜 Citation

If you use this model in your research or projects, please cite it as follows:

@misc{magpie_tts_sa-ar_arabic,
  author       = {Ahmed Eladl},
  title        = {Magpie TTS - Saudi Arabic (Fine-Tuned)},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/AhmedEladl/Magpie-TTS-Saudi-Arabic}
}

📧 Contact

For any questions, issues, or inquiries regarding this model, please reach out:

• Email: aeladl.dev@gmail.com