This model transcribes Wolof speech audio into Wolof text. It's the result of full fine-tuning (all 964M parameters) on 20,000 quality-filtered audio samples.

📊 Performance

Metric	Value
WER (Word Error Rate)	29.56%
Test set	500 examples
Training samples	20,000
Total audio	25.6 hours
Mode	Full fine-tuning (964M params)

🎯 Use Cases

🎙️ Transcription of Wolof audio (radio, podcasts, conversations)
🌍 First step in a Wolof → French translation pipeline
♿ Accessibility tools for Wolof-speaking communities
🔬 Research on low-resource African languages

🚀 Quick Start

python

from transformers import Wav2Vec2ForCTC, AutoProcessor
import torch
import torchaudio

# Load model and processor
model_id = "Sadou/mms-1b-wolof-finetuned"
processor = AutoProcessor.from_pretrained(model_id)
model = Wav2Vec2ForCTC.from_pretrained(model_id)

# Load audio (must be 16kHz, mono)
waveform, sr = torchaudio.load("your_audio.wav")
if sr != 16000:
    waveform = torchaudio.transforms.Resample(sr, 16000)(waveform)
if waveform.shape[0] > 1:
    waveform = waveform.mean(dim=0, keepdim=True)

# Transcribe
inputs = processor(waveform[0].numpy(), sampling_rate=16000, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)[0]
print(f"Transcription: {transcription}")

📈 Training Details

Dataset

Source: vonewman/wolof-audio-data
Combines: ALFFA + FLEURS + Urban Bus + Kallama
Quality filters applied:
- Audio duration: 2-15s
- Text: 3-30 words
- Speech rate: 1.0-4.5 words/sec
- Retention rate: 75.5%

Training Configuration

Base model: facebook/mms-1b-all
Mode: Full fine-tuning (all 964M parameters)
Precision: bf16 (Brain Float 16)
Effective batch size: 32 (16 × 2 gradient accumulation)
Gradient checkpointing: Enabled
Learning rate: 3e-5 (initial), then 5e-6 (final epochs)
Epochs: 6 (4 + 2 with LR decay)
Hardware: NVIDIA RTX PRO 6000 Blackwell

Training Progression

Step	WER	Phase
300	47.67%	Initial training
900	38.86%
1500	33.35%
2100	30.77%
2400	30.11%	Plateau detected
2700	30.56%	LR decay started
3300	29.56%
3600	29.56%	Final ✅

⚠️ Limitations

Code-switching: Difficulty with French words mixed in Wolof speech
Word segmentation: Sometimes merges or splits words incorrectly
Background noise: Performance degrades on noisy audio
Long audios: For audios > 30s, use chunking with stride

🔄 Long Audio Inference

For audios longer than 30 seconds, use HuggingFace pipeline with chunking:

python

from transformers import pipeline

pipe = pipeline(
    'automatic-speech-recognition',
    model="Sadou/mms-1b-wolof-finetuned",
    chunk_length_s=30,
    stride_length_s=(4, 2),
    device=0,
)

result = pipe("long_audio.wav")
print(result['text'])

🙏 Acknowledgments

Meta AI for MMS
vonewman for the Wolof audio dataset
GalsenAI and the Senegalese AI community

📚 Citation

bibtex

@misc{wolof-mms-2026,
  author = {Sadou Barry},
  title = {MMS-1B Wolof Fine-Tuned},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/Sadou/mms-1b-wolof-finetuned}
}