LyngualLabs/mms-300m-yoruba-english

Model Description

• Developed by: LyngualLabs
• Model type: Fine-tuned Meta MMS-300M (CTC)
• Languages: Yoruba, English (Code-Switched)
• Finetuned from model: facebook/mms-300m

Intended Uses & Limitations

This model is designed for transcribing speech that frequently alternates between Yoruba and English.

🛠 Training Procedure

Training Logs

The model was trained for approximately 5 epochs on an NVIDIA A100. The training reached stable convergence as shown below:

Step	Training Loss	Validation Loss	Wer
500	3.1712	2.9054	1.0000
1000	0.8191	0.6543	0.5025
3000	0.4789	0.4079	0.3294
6000	0.3450	0.3391	0.2823
9000	0.2958	0.3087	0.2536
11000	0.2609	0.3076	0.2433

Hyperparameters

• Learning Rate: 1e-4
• Epochs: 5
• Optimizer: AdamW
• LR Scheduler: Linear
• Batch Size: 32 (Train) / 32 (Eval)
• Mixed Precision: FP16
• Hardware: NVIDIA A100-SXM4-40GB

Evaluation Results

The model was evaluated on a held-out test set of 9,949 samples.

• Test WER: 0.2911

Metric	Value
Word Error Rate (WER)	0.2911

💻 How to Use

You can use this model directly with the Hugging Face pipeline for high-speed inference.

from transformers import pipeline

# Load the model and processor
pipe = pipeline("automatic-speech-recognition", model="Professor/mms-300m-yoruba-english")

# Run transcription
result = pipe("path_to_your_audio.wav")
print(result["text"])

---

⚠️ Limitations & Bias

• Audio Quality: While model is robust in noisy environment, performance may degrade in extreme noise.
• Dialects: While optimized for standard Yoruba, variations in regional dialects may affect transcription accuracy.
• Tone Marks: The model attempts to predict Yoruba tone marks, but secondary verification is recommended for formal text.

📜 Citation

If you use this model in your research, please cite it as follows for attribution, proper citation coming soon:

@misc{mms-300m-yo-en-cs,
  author = {LyngualLabs},
  title = {MMS-300M Fine-tuned for Yoruba-English Code-Switching},
  year = {2026},
  publisher = {Hugging Face}
}