Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- ary
|
| 4 |
+
license: apache-2.0
|
| 5 |
+
datasets:
|
| 6 |
+
- Lyte/DarijaTTS-clean
|
| 7 |
+
metrics:
|
| 8 |
+
- wer
|
| 9 |
+
base_model:
|
| 10 |
+
- openai/whisper-small
|
| 11 |
+
pipeline_tag: automatic-speech-recognition
|
| 12 |
+
library_name: transformers
|
| 13 |
+
tags:
|
| 14 |
+
- asr
|
| 15 |
+
- darija
|
| 16 |
+
---
|
| 17 |
+
# Whisper Small Darija - Lyte (Yassine Ennour)
|
| 18 |
+
|
| 19 |
+
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).
|
| 20 |
+
|
| 21 |
+
## Model Details
|
| 22 |
+
|
| 23 |
+
- **Model Name**: [`Lyte/Whisper-Small-Darija`](https://huggingface.co/Lyte/Whisper-Small-Darija)
|
| 24 |
+
- **Base Model**: [`openai/whisper-small`](https://huggingface.co/openai/whisper-small)
|
| 25 |
+
- **Fine-Tuned On**: [`Lyte/DarijaTTS-clean`](https://huggingface.co/datasets/Lyte/DarijaTTS-clean)
|
| 26 |
+
- **Language**: Moroccan Darija (ary)
|
| 27 |
+
- **Task**: Automatic Speech Recognition (ASR)
|
| 28 |
+
- **Dataset Size**: 19.9k training samples, 1k test samples
|
| 29 |
+
|
| 30 |
+
## Training Progress
|
| 31 |
+
|
| 32 |
+
Training was started but interrupted. The training will be resumed from step **600**. Below is the progress so far:
|
| 33 |
+
|
| 34 |
+
| Step | Training Loss | Validation Loss | WER |
|
| 35 |
+
|------|--------------|----------------|-----|
|
| 36 |
+
| 200 | 1.0142 | 1.0804 | 129.35 |
|
| 37 |
+
| 400 | 0.8288 | 0.9905 | 72.44 |
|
| 38 |
+
| 600 | 0.7618 | 0.9656 | 70.41 |
|
| 39 |
+
|
| 40 |
+
## Usage
|
| 41 |
+
|
| 42 |
+
You can use this model with Hugging Face's `transformers` library:
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from transformers import WhisperProcessor, WhisperForConditionalGeneration
|
| 46 |
+
import torch
|
| 47 |
+
|
| 48 |
+
model_id = "Lyte/whisper-small-darija"
|
| 49 |
+
processor = WhisperProcessor.from_pretrained(model_id)
|
| 50 |
+
model = WhisperForConditionalGeneration.from_pretrained(model_id)
|
| 51 |
+
|
| 52 |
+
# Load an audio file and preprocess
|
| 53 |
+
input_features = processor("path_to_audio.wav", return_tensors="pt").input_features
|
| 54 |
+
|
| 55 |
+
generated_ids = model.generate(input_features)
|
| 56 |
+
predicted_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
|
| 57 |
+
|
| 58 |
+
print(predicted_text)
|
| 59 |
+
```
|
| 60 |
+
|
| 61 |
+
## Next Steps
|
| 62 |
+
|
| 63 |
+
- **Resume training** from step 600 to further improve WER.
|
| 64 |
+
- **Optimize hyperparameters** to reduce validation loss.
|
| 65 |
+
- **Expand dataset** for better generalization.
|
| 66 |
+
|
| 67 |
+
## Acknowledgments
|
| 68 |
+
|
| 69 |
+
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.
|
| 70 |
+
|
| 71 |
+
---
|
| 72 |
+
For updates and more details, stay tuned to this repository!
|