README.md

---
library_name: peft
license: apache-2.0
base_model: openai/whisper-large-v2
tags:
- automatic-speech-recognition
- whisper
- asr
- bambara
- bm
- Mali
- MALIBA-AI
- lora
- fine-tuned
- code-switching
language:
- bm
language_bcp47:
- bm-ML

metrics:
- name: WER
  type: wer
  value: 0.22
- name: CER
  type: cer
  value: 0.10
  
model-index:
- name: maliba-asr-v1
  results:
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: oza75/bambara-asr
      type: oza75/bambara-asr
      split: test
      subset: clean-combined
      args:
        language: bm
    metrics:
    - name: WER
      type: wer
      value: 0.2264
    - name: CER
      type: cer
      value: 0.1094
pipeline_tag: automatic-speech-recognition
---
[![Model architecture](https://img.shields.io/badge/Model_Arch-Whisper--LoRA-lightgrey)](#model-architecture)
 [![WER](https://img.shields.io/badge/WER-0.22-green)](#performance)
 [![CER](https://img.shields.io/badge/WER-0.10-green)](#performance)
 [![Language](https://img.shields.io/badge/Language-Bambara-lightgrey)](#datasets)
 [![License](https://img.shields.io/badge/License-Apache--2.0-blue)](#license)


# MALIBA-ASR-v1: Revolutionizing Bambara Speech Recognition

MALIBA-ASR-v1 represents a breakthrough in African language technology, setting a new for Bambara speech recognition. Developed by MALIBA-AI, this model significantly outperforms all existing open-source solutions for Bambara ASR, bringing unprecedented quality speech technology to Mali's most widely spoken language.

## Bridging the Digital Language Divide

Despite being spoken by over 22 million people, Bambara has remained severely underrepresented in speech technology. MALIBA-ASR-v1 directly addresses this critical gap, achieving performance levels that make digital voice interfaces accessible to Bambara speakers. This work represents a crucial step toward digital language equality and demonstrates that high-quality speech technology is possible for African languages.

## Performance Metrics

MALIBA-ASR-v1 achieves breakthrough results on the oza75/bambara-asr benchmark:
Here's the metrics table showing only the WER and CER values for your model:


|  Metric     |  Value   |   
|:-----:|:---:|
| WER | 0.22 |
| CER | 0.10 | 





## Exceptional Code-Switching Capabilities

One of the most significant advantages of MALIBA-ASR-v1 is its capability of code-switching – the natural mixing of Bambara with French or other languages that characterizes everyday speech in Mali. MALIBA-ASR-v1 accurately transcribes multi-lingual content, making it practical for real-world applications.


## Transforming Access to Technology in Mali

MALIBA-ASR-v1 enables numerous applications previously unavailable to Bambara speakers:

- **Healthcare**: Voice interfaces for medical information and services
- **Education**: Audio-based learning tools for literacy and education
- **News & Media**: Automated transcription of Bambara broadcasts and podcasts
- **Preservation**: Documentation of oral histories and traditional knowledge
- **Accessibility**: Voice technologies for visually impaired Bambara speakers
- **Mobile Access**: Voice commands for smartphone users with limited literacy


## Training Details

### Dataset and Evaluation
The model was trained on the [coming soon] dataset, representing diverse speakers, dialects, and recording conditions.

### Training Procedure
- **Base Model**: openai/whisper-large-v2
- **Adaptation Method**: LoRA (PEFT)
- **Training Duration**: 6 epochs
- **Batch Size**: 128 (32 per device with gradient accumulation steps of 4)
- **Learning Rate**: 0.001 with linear scheduler and 50 warmup steps
- **Mixed Precision**: Native AMP
- **Optimizer**: AdamW (betas=(0.9, 0.999), epsilon=1e-08)

### Training Results

| Training Loss | Epoch | Step | Validation Loss |
|:-------------:|:-----:|:----:|:---------------:|
| 0.3265        | 1.0   | 531  | 0.4117          |
| 0.2711        | 2.0   | 1062 | 0.3612          |
| 0.223         | 3.0   | 1593 | 0.3397          |
| 0.1802        | 4.0   | 2124 | 0.3330          |
| 0.1268        | 5.0   | 2655 | 0.3339          |
| 0.0932        | 6.0   | 3186 | 0.3491          |

## Usage Examples

```bash
  pip install git+https://github.com/sudoping01/whosper.git
```

```python
from whosper import WhosperTranscriber

# Initialize the transcriber
transcriber = WhosperTranscriber(model_id = "MALIBA-AI/maliba-asr-v1") 

# Transcribe an audio file
result = transcriber.transcribe_audio("path/to/your/audio.wav")
print(result)
```

## The MALIBA-AI Impact

MALIBA-ASR-v1 is part of MALIBA-AI's broader mission to ensure "No Malian Language Left Behind." This initiative is actively transforming Mali's digital landscape by:

1. **Breaking Language Barriers**: Providing technology in languages that Malians actually speak
2. **Enabling Local Innovation**: Allowing Malian developers to build voice-based applications
3. **Preserving Cultural Heritage**: Digitizing and preserving Mali's rich oral traditions
4. **Democratizing AI**: Making cutting-edge technology accessible to all Malians regardless of literacy level
5. **Building Local Expertise**: Training Malian AI practitioners and researchers

## Future Development

MALIBA-AI is committed to continuing this work with:

- Extension to other Malian languages (Songhoy, Pular, Tamasheq, etc.)


## Join Our Mission

MALIBA-ASR-v1 embodies our commitment to open science and the advancement of African language technologies. We believe that by making cutting-edge speech recognition models freely available, we can accelerate NLP development across Africa.

Join our mission to democratize AI technology:

- **Open Science**: Use and build upon our research - all code, models, and documentation are open source
- **Data Contribution**: Share your Bambara speech datasets to help improve model performance
- **Research Collaboration**: Integrate our models into your research projects and share your findings
- **Application Development**: Build tools that serve Malian communities using our models
- **Educational Impact**: Use our models in educational settings to train the next generation of African AI researchers

## License

This model is released under the Apache 2.0 license to encourage research, commercial use, and innovation in African language technologies while ensuring proper attribution and patent protection.

## Citation

```bibtex
@misc{maliba-asr-v1,
  author = {MALIBA-AI},
  title = {MALIBA-ASR-v1:  Bambara Automatic Speech Recognition},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/MALIBA-AI/maliba-asr-v1}}
}
```


---

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

*"No Malian Language Left Behind"*