KandirResearch/Whisper-Small-Darija

Whisper Small Darija - Lyte (Yassine Ennour)

This is a fine-tuned version of OpenAI's whisper-small model on the DarijaTTS-clean dataset. The goal of this project is to improve automatic speech recognition (ASR) for Moroccan Darija (ary).

Model Details

• Model Name: Lyte/Whisper-Small-Darija
• Base Model: openai/whisper-small
• Fine-Tuned On: Lyte/DarijaTTS-clean
• Language: Moroccan Darija (ary)
• Task: Automatic Speech Recognition (ASR)
• Dataset Size: 19.9k training samples, 1k test samples

Training Progress

Training was started but interrupted. The training will be resumed from step 600. Below is the progress so far:

Step	Training Loss	Validation Loss	WER
200	1.0142	1.0804	129.35
400	0.8288	0.9905	72.44
600	0.7618	0.9656	70.41

Usage

You can use this model with Hugging Face's transformers library:

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

model_id = "Lyte/whisper-small-darija"
processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(model_id)

# Load an audio file and preprocess
input_features = processor("path_to_audio.wav", return_tensors="pt").input_features

generated_ids = model.generate(input_features)
predicted_text = processor.batch_decode(generated_ids, skip_special_tokens=True)

print(predicted_text)

Next Steps

• Resume training from step 600 to further improve WER.
• Optimize hyperparameters to reduce validation loss.
• Expand dataset for better generalization.

Acknowledgments

Special thanks to OpenAI for Whisper and Hugging Face for their amazing platform. This model is built as part of my ongoing research in ASR for Darija.

---

For updates and more details, stay tuned to this repository!