README.md

---
library_name: transformers
base_model: SparkAudio/Spark-TTS-0.5B

tags:
  - text-to-speech
  - tts
  - spark-tts
  - llm-based-tts
  - bambara
  - african-languages
  - open-source
  - mali
  - maliba-ai
  - text-generation-inference
  - transformers
  - unsloth

extra_gated_fields:
  Name: text
  Official Email (organization or academic email): text
  Affiliation (University, Research Lab, etc): text
  I confirm I am a researcher, student, or member of a non-profit organization: checkbox
  I confirm I am NOT affiliated with a for-profit company and will not use this model on behalf of one: checkbox
  I have read and agree to the MALIBA-AI Research License (Non-Commercial, Non-Profit Use Only): checkbox
  I agree to cite the MALIBA-AI paper and all original references when using this model: checkbox


language:
  - bm
language_bcp47:
  - bm-ML

model-index:
  - name: bambara-tts
    results:
      - task:
          name: text-to-speech
          type: speech-synthesis
        metrics:
          - name: Subjective Quality (MOS)
            type: mos
            value: 4.2
          - name: Speaker Similarity
            type: similarity
            value: High
          - name: Naturalness
            type: naturalness
            value: 4.1

pipeline_tag: text-to-speech
license: cc-by-nc-sa-4.0

---





# MALIBA-AI Bambara TTS 🇲🇱

<style>
img {
 display: inline;
}
</style>

[![Model architecture](https://img.shields.io/badge/Model_Arch-Spark--TTS-lightgrey)](#model-architecture)
| [![Model size](https://img.shields.io/badge/Params-500M-lightgrey)](#model-architecture)
| [![Language](https://img.shields.io/badge/Language-bm-lightgrey)](#datasets)
| [![License](https://img.shields.io/badge/License-CC--BY--NC--SA--4.0-blue)](#license)

## Model Overview

This model provides neural text-to-speech synthesis for Bambara (Bamanankan), the most widely spoken language in Mali. The model supports 10 authentic Bambara speakers and produces high-fidelity audio without requiring separate vocoder models. It serves over 14 million Bambara speakers across West Africa with native-level pronunciation and cultural authenticity.

<!-- - Try our live demo on [Hugging Face Spaces](https://huggingface.co/spaces/MALIBA-AI/BambaraText2Speech) -->
- **Available Speakers:** Adama, Moussa, Bourama, Modibo, Seydou, Amadou, Bakary, Ngolo, Ibrahima, Amara

> **⚠️ Non-commercial use only.** One of the training data of the base model `spark-tts` carries a license that restricts commercial use.


## Quick Start

### Installation


<!-- ```bash
  pip install maliba-ai==1.1.1b0
``` -->

<!-- For development installations: -->

```bash
pip install git+https://github.com/MALIBA-AI/bambara-tts.git
```
with uv (faster)

<!-- ```bash
    uv pip install maliba-ai==1.1.1b0
``` -->

```bash
    uv pip install git+https://github.com/MALIBA-AI/bambara-tts.git
```
Note : if you are in colab  please install those additional dependencies : 

```
    !pip install --no-deps bitsandbytes accelerate xformers==0.0.29.post3 peft trl triton cut_cross_entropy unsloth_zoo
    !pip install sentencepiece protobuf huggingface_hub hf_transfer
    !pip install --no-deps unsloth
```

### Basic Usage

```python
from maliba_ai.tts.inference import BambaraTTSInference
from maliba_ai.config.settings import Speakers

tts = BambaraTTSInference()

text = "Aw ni ce. I ka kɛnɛ wa?"  
audio = tts.generate_speech(text=text, speaker_id=Speakers.Bourama, output_filename="greeting.wav")

```

Note: More detail : https://github.com/sudoping01/bambara-tts/blob/main/README.md

A notebook is available on [this link](https://colab.research.google.com/drive/1rJy-mV4Zte33xOWSpkzmY-jBp0T71AFk?usp=sharing), enabling you to test the model quickly.
The inference server: [server](https://github.com/MALIBA-AI/maliba-framework/tree/main/server)
## Technical Specifications

### Architecture
- **Base Model**: Spark-TTS (LLM-based TTS)
- **Foundation**: Qwen2.5-based language model
- **Parameters**: ~500M
- **Audio Format**: 16kHz, 16-bit PCM mono
- **Language Support**: Bambara (bm-ML)


## Model Input/Output

### Input
- **Text**: Bambara text in standard orthography
- **Speaker ID**: Choice of 10 available speakers
- **Parameters**: Temperature, top-k, top-p (optional)

### Output
- **Audio**: 16kHz mono WAV format
- **Quality**: Professional-grade speech synthesis

## ⚠️ Known Limitations

### Language Mixing
- **Issue**: Poor performance with French-Bambara code-switching
- **Recommendation**: Use pure Bambara text for optimal results

### Numeric Content
- **Update**: We now use **[Bambara Text Normalizer](https://pypi.org/project/bambara-text-normalizer/)** to preprocess input text. This allows the TTS model to generate speech for **digits, dates, and times** automatically.  
- **Integration**: This normalization is already **built into the MALIBA-AI Bambara TTS Inference framework**, so users do not need to manually convert numbers or dates.  
- **Recommendation**: For best results, provide text in standard Bambara orthography, but numeric content and dates are handled automatically.


## ⚠️ Disclaimer

This model provides high-fidelity Bambara speech synthesis intended for research, education, and community applications. The following uses are **strictly forbidden**:

- **Voice Impersonation**: Do not clone voices without explicit consent
- **Deceptive Content**: Do not generate misleading or fraudulent audio
- **Illegal Activities**: Do not use for any unlawful purposes

By using this model, you agree to uphold ethical standards and legal responsibilities. We **are not responsible** for any misuse and firmly oppose unethical usage of this technology.

If you have concerns about potential misuse or need guidance on ethical applications, please contact us at ml.maliba.ai@gmail.com


## License

**[MALIBA-AI Research Licence](https://huggingface.co/MALIBA-AI/bambara-tts/blob/main/LICENCE.md)** - Non-commercial use due to some data used in base model training.

### Key Terms (Non-Commercial License)

* ✅ **Permitted:** research, academic, educational, and personal use
* ✅ **Required:** clear attribution to the original authors (MALIBA-AI)
* ✅ **Allowed:** modifications and derivative works, **provided they are shared under the same non-commercial, share-alike terms**
* ❌ **Prohibited:** any commercial use, monetization, or revenue-generating activity
* ❌ **Prohibited:** use by companies, startups, or for-profit organizations without prior written permission
* ❌ **Prohibited:** selling, sublicensing, hosting as a paid service, or integrating into commercial products

If you have any questions, please contact us at: `ml[dot]maliba[dot]ai[at]gmail.com`


## Citation

```bibtex
@software{maliba_ai_bambara_tts,
  title={MALIBA-AI Bambara Text-to-Speech: Open-Source High-Quality TTS for Bambara Language},
  author={MALIBA-AI},
  year={2025},
  url={https://huggingface.co/MALIBA-AI/bambara-tts}
}
```

---

**MALIBA-AI: Empowering Mali's Future Through Community-Driven AI Innovation**

*"No Malian Language Left Behind"*
---

**Contact Information:**
- Website: [maliba-ai.org](https://maliba-ai.org)
- Email: ml.maliba.ai@gmail.com
- GitHub: [MALIBA-AI](https://github.com/MALIBA-AI)
- HuggingFace: [MALIBA-AI](https://huggingface.co/MALIBA-AI)