🇷🇸 sr-cv-vits

Serbian Text-to-Speech model based on VITS architecture.

---

📌 Model Overview

sr-cv-vits is a high-quality single-speaker Serbian TTS model trained using the VITS architecture and the Coqui TTS framework.

The model generates natural-sounding speech in Serbian (Cyrillic script).

• 🗣️ Language: Serbian (Ćirilica)
• 👩 Speaker: Female (single-speaker)
• 🧠 Architecture: VITS
• ⚙️ Framework: Coqui TTS
• 🔤 Text processing: multilingual_cleaners
• 📚 Dataset formatter: common_voice
• 🔊 Sampling rate: 22050 Hz

---

🎧 Audio Samples

Below are example outputs generated by the model:

▶️ Sample 1

▶️ Sample 2

📌 Place test1.wav and test2.wav in the root of the repository.

---

📊 Training Details

• Epochs: 1000
• Dataset size: ~8 hours of speech
• Speaker: Female (volunteer)
• Script: Cyrillic only
• Numbers: Not supported (removed from dataset)

---

📚 Dataset

The model is trained on a dataset formatted using the common_voice formatter. The dataset contains approximately 8 hours of curated Serbian speech data.

⚠️ Note: Numerical expressions were excluded during training.

---

⚙️ Usage

This model can be used via Coqui TTS.

🐍 Python Example

from TTS.api import TTS

# Load model from Hugging Face
tts = TTS(model_name="daremc86/sr-cv-vits")

# Generate speech
tts.tts_to_file(
    text="Здраво, како си?",
    file_path="output.wav"
)

---

📦 Installation

Install Coqui TTS:

pip install coqui-tts

---

🔊 Example Text

Здраво, како си?
Ово је пример синтезе говора.

---

⚠️ Limitations

• ❌ Does not support numbers (e.g. "123")

• 🔤 Works best with Cyrillic script only

• 📉 May struggle with:

  • long or complex sentences
  • uncommon words
  • punctuation-heavy text

---

📜 License

This model is released under the CC-BY-4.0 License.

---

👤 Author

Darko Milošević 📧 daremc86@gmail.com

---

🤝 Acknowledgements

This model is based on the VITS architecture and trained using the Coqui TTS framework.

---

🚀 Future Improvements

• Support for numbers
• Latin script support
• Larger dataset
• Multi-speaker version

---

🧩 Notes

For best results:

• Use clean, well-structured sentences
• Avoid numbers and special symbols
• Prefer standard Serbian grammar and punctuation