A comprehensive suite of Text-to-Speech (TTS) models for Ugandan languages, supporting English, Luganda, Runyonkole, Tesso, and Acholi with multiple speaker voices for each language.
This project introduces a groundbreaking collection of Text-to-Speech models specifically designed for Ugandan languages. These models represent a significant advancement in African language technology, addressing the critical gap in speech synthesis capabilities for Uganda's diverse linguistic landscape.
The Uganda TTS model family consists of fine-tuned versions of the Orpheus 3B model, each specialized for different Ugandan languages while maintaining quality speech synthesis capabilities. Each language model supports multiple distinct speaker voices, enabling versatile applications in education, accessibility, content creation, and digital preservation of Ugandan languages.
These models are designed to serve researchers, developers, educators, and content creators who need high-quality speech synthesis in Ugandan languages, contributing to the broader goal of digital language preservation and accessibility in Africa.
Model Architecture
The Uganda TTS models utilize a sophisticated two-stage architecture:
Audio Token Generation: The models generate audio tokens based on the SNAC (Structured Neural Audio Codec) framework
Fine-tuned Processing: These audio tokens are then processed through specialized fine-tuned versions of the Orpheus 3B model, each optimized for specific Ugandan languages
This architecture enables efficient and high-quality speech synthesis while maintaining computational efficiency suitable for various deployment scenarios. More about the Orpheous Models here
Supported Voices: Michelle, Barbara, Jessica, Christopher, James, Daniel, Charles, Mark
Acholi
Supported Voices: James, Barbara, Michelle, Mark, Christopher
Audio Examples
English
Christopher Prompt: "Hello I can speak in English as christopher, one of the voices I can speak."
Barbara Prompt: "Or as barbra, this is one of my female voices. Pretty cool right?."
Mary Prompt: "I can also speak as Mary as well."
James Prompt: "Or I can speak as james as you can see."
Jessica Prompt: "This is my other voice called jessica, I have more voices of jennifer, suzan, linda, patricia and elizabeth. But I will be sharing these voices once I am fully done from baking."
Luganda
Christopher Prompt: "Nsobolla okwo'geranga Christopher nga wowulila kati."
Charles Prompt: "Oba neenjogela nga charles wenti."
All models are provided under an open-source license, encouraging collaboration and further development in African language technologies.
Limitations
While these models represent significant progress in Ugandan language TTS, there are some current limitations:
Non-English Language Quality: The non-English models may occasionally produce lower quality outputs compared to the English model. This is primarily due to the SNAC audio codec not being pre-trained on these languages, which affects the initial audio token generation quality.
Speaker Consistency: Non-English voices may sometimes generate speech that does not perfectly match the specified speaker characteristics due to limited training data for certain voice-language combinations.
Language Coverage: Current models focus on five major Ugandan languages, with plans to expand to additional languages based on data availability and community needs.
Note: We are actively working on an improved version that addresses these limitations, including training SNAC on a more diverse set of languages and expanding the training datasets for better speaker fidelity.
Future Work
Completed ✅
Train the models for each of the languages
Open source the models
In Progress 🔄
Develop a Python package to act as an API for the models
Write a comprehensive white paper detailing the training process and methodology
Improve SNAC training for better non-English language support
Expand training datasets for enhanced speaker consistency
Citation
If you use these models in your research or applications, please cite:
bibtex
1@misc{uganda_tts_2024,
2 author = {Kisejjere Rashid and Magala Reuben},
3 title = {Uganda Text-to-Speech (TTS) Models},
4 year = {2024},
5 howpublished = {\url{https://huggingface.co/Uganda-lang}},
6 note = {Fine-tuned versions of Orpheus 3B for Ugandan languages}
7}
Contributing: We welcome contributions, feedback, and collaboration from the community. Please feel free to open issues or submit pull requests to help improve these models.
Contact: For questions, collaborations, or support, please reach out through the HuggingFace model repository or create an issue in our GitHub repository.
