Orpheus-Urdu-TTS-Zia-Mohyeddin

This model is a fine-tuned version of mahwizzzz/orpheus-urdu-tts, specifically adapted to replicate the iconic voice and oratorical style of the legendary Pakistani orator, Zia Mohyeddin.

The model was fine-tuned on the muhammadsaadgondal/urdu-tts dataset, which contains high-quality audio recordings of Zia Mohyeddin. The fine-tuning process was performed efficiently using Unsloth, which enables faster training with significantly less memory.

This model allows for the generation of clear, natural-sounding Urdu speech that captures the distinct cadence, tone, and gravitas of Zia Mohyeddin's delivery, making it ideal for high-quality TTS applications.

Model Details

• Base Model: mahwizzzz/orpheus-urdu-tts
• Language: Urdu (ur)
• Core Technology: Orpheus 3B
• Fine-tuning Framework: Unsloth

Fine-tuning Details

• Dataset: muhammadsaadgondal/urdu-tts - A high-quality dataset of Zia Mohyeddin's speech, originally compiled for ASR tasks.
• Training Framework: The model was fine-tuned with Unsloth, enabling efficient training on consumer-grade hardware.
• Training Configuration:
  • Epochs: 2
  • Batch Size: 4
  • Learning Rate: 1e-4
  • Hardware: 1x NVIDIA L4 GPU on Google Colab
  • Quantization: LoRA (16-bit)

Limitations

The model cannot pronounce many Urdu words properly, although it has adapted the voice of Zia Mohyeddin.

Citation

If you use this model in your work, please cite the original dataset and acknowledge this fine-tuned version.

@dataset{gondal_2024_urdutts,
  author       = {Muhammad Saad Gondal},
  title        = {urdu-tts},
  year         = {2024},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/muhammadsaadgondal/urdu-tts}
}

@software{Aasher_2025_orpheus_zia,
  author       = {Aasher Kamal},
  title        = {Orpheus-Urdu-TTS-Zia-Mohiuddin},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/Aasher/Orpheus-Urdu-TTS-Zia-Mohyeddin}
}