---
license: apache-2.0
datasets:
- galsenai/wolof-audio-data
language:
- wo
- fr
metrics:
- wer
- cer
base_model:
- openai/whisper-large-v2
pipeline_tag: automatic-speech-recognition
---
# Kiriku ASR: Breaking the Silence for Wolof

**The largest open-source Automatic Speech Recognition (ASR) model for the Wolof language.**

Kiriku-ASR is a state-of-the-art fine-tune of `whisper-large-v2`. Developed by **IA Hub Senegal**, this model represents a massive leap forward for West African speech technology.

## Why This Matters
Wolof is a low-resource language with **zero native support** in the original Whisper vocabulary. Kiriku ASR doesn't just fine-tune weights; it expands the model's fundamental understanding:

* **Custom Vocabulary:** Manually injected Wolof-specific characters (`ñ`, `ë`, `ŋ`, `ɗ`, `ɓ`, `ƴ`) into the tokenizer.
* **Smart Initialization:** Used "Warm-Starting" from French `<|fr|>` embeddings to give the model a linguistic head-start.
* **Massive Expert Corpus:** Trained on **88.27 hours** of expert-verified speech—the largest open-source effort for Wolof to date.
* **Free-Transcription Mode:** Optimized to operate without a forced language token, avoiding the common "hallucination" artifacts found in base models.

## Performance & Results

Through rigorous optimization and data augmentation, Kiriku ASR achieves a benchmark-setting Word Error Rate (WER) for the region.

| Metric | Score | Note |
| :--- | :--- | :--- |
| **Best WER** | **20.70%** | New Open Source SOTA for Wolof |
| **Training Data** | **88.27h** | Multi-domain (Agriculture, Banking, Urban) |
| **Base Model** | **Whisper Large-v2** | Enhanced with custom Wolof diacritics |

### Training Configuration
* **Learning Rate:** 1e-5 (Cosine schedule)
* **Batch Size:** 16 (Effective)
* **Regularization:** Dropout 0.1 to prevent overfitting on low-resource data.
* **Augmentation:** Gaussian noise, time stretching, and pitch shifting for real-world robustness.
* 
## Quick Start
You can start transcribing Wolof immediately using the Hugging Face `pipeline` API:

```python
from transformers import pipeline
# Load Kiriku ASR
asr = pipeline(
    "automatic-speech-recognition", 
    model="AIHubSN/kiriku-ASR",
    device=0 
)
# Transcribe (Ensure audio is sampled at 16kHz)
result = asr("audio_sample_wolof.wav")
print(result["text"])
```
---
## Produced By

This project is a testament to Senegalese AI excellence:

* **AI Hub Senegal** — Coordination, community mobilization, and project support  
* **Contact:** contact@aihubsenegal.com
  
## Citation

If this work helps your research or application, please cite our efforts:

```bibtex
@misc{iahubsn2026kiriku_asr,
  title        = {Kiriku ASR: Whisper Large-v2 Fine-tuned for Wolof},
  author       = {{IA Hub Senegal},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/AIHubSN/kiriku-ASR}},
  note         = {Fine-tuned on 88.27 hours of expert-verified Wolof speech}
}
```
**Developed with 🤍 in Dakar for the African AI Renaissance.**