Dama12/dioula-tts-vits

🎙️ Dioula Text-to-Speech (VITS)

This is a Text-to-Speech (TTS) model trained for the Dioula (Jula) language, a Manding language spoken by over 20 million people in West Africa (Ivory Coast, Burkina Faso, Mali).

This model is built using the VITS architecture via the Coqui TTS framework.

👨‍💻 Developer Information

• Developed by: Soumana Dama — Full stack developer & AI engineer - Founder & Lead AI Engineer at Scoinvestigator AI
• LinkedIn: Soumana Dama
• GitHub: Damasoumana1

📊 Model Details

• Architecture: VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech)
• Framework: TTS==0.20.6 (Coqui TTS)
• Parameters: 83.05 Million (83,048,044)
• Language: Dioula (dyu)
• Sample Rate: 16,000 Hz
• Status: Work in Progress — v5 (Best model at step 193,972 | Last checkpoint: 229,500 steps)

📈 Training Metrics

v5 — Best Model (Step 193,972) · June 28, 2026

Metric	Value	vs v4	Description
Avg Loss Discriminator (avg_loss_disc)	2.45	✅ 2.59 → 2.45	Discriminator loss (real vs fake audio).
Avg Loss Generator (avg_loss_0)	2.45	✅ 2.59 → 2.45	Overall generator loss.
Avg Loss Mel/KL (avg_loss_1)	41.26	⚠️ 40.71 → 41.26	Mel spectrogram + KL divergence loss.
Avg Loss Duration (avg_loss_duration)	2.49	✅ 2.50 → 2.49	Phoneme/character duration prediction.
Avg Loss Feat (avg_loss_feat)	6.15	↗️ 5.36 → 6.15	Feature matching loss.
Epoch	8	—	Total training epochs completed.
Last Checkpoint	229,500 steps	—	Available for resuming training (v6).
Parameters	83,048,044	—	Total model parameters.

(Training curves, alignment plots and spectrograms are available in the TensorBoard logs in this repository.)

v4 — Best Model (Step 102,972) · Archived

Metric	Value	Description
Avg Loss Discriminator (avg_loss_disc)	2.59	Discriminator loss (real vs fake audio).
Avg Loss Generator (avg_loss_0)	2.59	Overall generator loss.
Avg Loss Mel/KL (avg_loss_1)	40.71	Mel spectrogram + KL divergence loss.
Avg Loss Duration (avg_loss_duration)	2.50	Phoneme/character duration prediction.
Avg Loss Feat (avg_loss_feat)	5.36	Feature matching loss.

🚀 How to Use (Inference)

You can use this model to generate Dioula speech from text using Python.

1. Install Dependencies

pip install TTS==0.20.6

2. Python Script

import os
from TTS.api import TTS

# Initialize the TTS API with the path to your downloaded model and config
model_path = "best_model.pth" # or checkpoint_xxxx.pth
config_path = "config.json"

tts = TTS(model_path=model_path, config_path=config_path, progress_bar=False, gpu=False)

# The text you want to synthesize in Dioula
text_to_speak = "I ni cɛ, i ka kɛnɛ wa ?"

# Generate the audio file
output_path = "output_dioula.wav"
tts.tts_to_file(text=text_to_speak, file_path=output_path)

print(f"Audio saved to {output_path}")

📚 Dataset

This model was trained on a custom Dioula dataset consisting of approximately 9,000 verified audio samples and transcripts, prepared specifically for speech processing tasks.

⚠️ Limitations

• Vocabulary: Some French-specific accented characters (é, è, ô, ç) might not be fully supported in this version's vocabulary and may be ignored during synthesis. It is recommended to use clean Dioula orthography.
• Intonation: As the model is still training, long sentences might lack perfect natural prosody.

⚖️ License & Citation

License

This model is built upon the VITS architecture via the Coqui TTS framework. The model weights are released under the CC BY-NC 4.0 License (free for academic and research purposes). The dataset, methodology, and this specific implementation are intended strictly for non-commercial research purposes. If you intend to use this model for commercial applications, please contact the author.

Citation

If you use this model or the Audio Dioula Project concept in your research, please cite it as follows:

@misc{soumanadama2026dioula_tts,
  author = {Soumana Dama},
  title = {Audio Dioula Project - Multilingual AI Assistant for Education in Burkina Faso},
  year = {2026},
  note = {Master's Project: Applied Artificial Intelligence in the African Context}
}