ORPHEOUSHAUSA3

ORPHEOUSHAUSA3 is a fine-tuned version of Orpheus TTS designed for high-quality Hausa speech synthesis.

The model generates natural, fluent, and expressive Hausa speech directly from text while preserving pronunciation, rhythm, and prosody.

This model is intended for researchers, developers, and organizations building speech technologies for African languages.

---

Model Details

Property	Value
Model Name	ORPHEOUSHAUSA3
Base Model	Orpheus TTS
Language	Hausa
Task	Text-to-Speech
Framework	Transformers
Fine-tuned by	EYEDOL
License	Apache-2.0

---

Features

• Native Hausa speech generation
• Natural sounding voices
• Good pronunciation of Hausa words
• Fast autoregressive inference
• Compatible with Hugging Face Transformers
• Can be integrated into speech assistants, accessibility systems, education platforms, and conversational AI.

---

Intended Uses

This model is suitable for:

• Audiobook generation
• Voice assistants
• Accessibility applications
• Educational software
• Content creation
• Speech interfaces
• Interactive AI agents
• Hausa localization

---

Training

This model was fine-tuned from the Orpheus TTS base model using a curated Hausa speech dataset.

Training focused on improving:

• pronunciation accuracy
• speech naturalness
• prosody
• language fluency
• stability during long-form generation

---

Dataset

The training dataset consists of paired Hausa text and speech recordings.

Dataset characteristics include:

• native Hausa speakers
• high-quality audio
• cleaned transcripts
• normalized text
• multiple speaking styles

The dataset was processed into the format required by Orpheus TTS before fine-tuning.

---

Usage

Installation

pip install transformers accelerate torch soundfile

---

Load Model

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "EYEDOL/ORPHEOUSHAUSA3"

tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

---

Example

prompt = "Sannu! Barka da zuwa."

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output = model.generate(
    **inputs,
    max_new_tokens=1200
)

Refer to the original Orpheus documentation for complete inference examples.

---

Performance

The model has been qualitatively evaluated on:

• pronunciation accuracy
• intelligibility
• naturalness
• speech continuity

It performs well on standard Hausa text and conversational prompts.

---

Limitations

Performance is best on standard Hausa orthography.

---

Ethical Considerations

Users should ensure generated speech is used responsibly.

Potential misuse includes:

• impersonation
• misinformation
• deepfake generation

The authors discourage malicious use of synthetic speech.

---

Citation

@misc{orpheoushausa3,
  title={ORPHEOUSHAUSA3: Hausa Text-to-Speech Model},
  author={EYEDOL},
  year={2026},
  publisher={Hugging Face},
  howpublished={https://huggingface.co/EYEDOL/ORPHEOUSHAUSA3}
}

---

Acknowledgements

This work builds upon the excellent Orpheus Text-to-Speech model.

Thanks to the Orpheus developers and the Hugging Face community.

---

Contact

Maintained by EYEDOL

For issues, feature requests, or collaborations, please open an issue on the Hugging Face repository.

---

License

This model follows the license of the original Orpheus model unless otherwise specified.

Please ensure compliance with the base model's license before commercial deployment.