mohamedmou/moroccan-darija-asr-dataset-split
Viewer • Updated • 16.8k • 184 • 1
Fine-tuning de omarxadel/hubert-large-arabic-egyptian pour la
reconnaissance automatique de la parole (ASR) en darija (dialecte marocain),
sur un ensemble combiné de datasets publics et privés.
🧪 Please test this model on your own audio samples before relying on it in production. If it works well for you, a ❤️ like or a comment on this repo is very appreciated — it helps me know it's useful and guides future improvements!
| Métrique | Valeur |
|---|---|
| WER (Word Error Rate) | 0.2386 (23.86%) |
| CER (Character Error Rate) | 0.0813 (8.13%) |
| Train loss (moyenne sur tout l'entraînement) | 0.41482063716298273 |
| Ressource | Détail |
|---|---|
| GPU | NVIDIA GeForce RTX 3090 (24 Go VRAM) |
| Driver / CUDA | Driver 595.84 / CUDA 13.2 |
| Mémoire GPU utilisée | ~10.2 Go / 24 Go |
| Hyperparamètre | Valeur |
|---|---|
| Modèle de base | omarxadel/hubert-large-arabic-egyptian |
| Époques | 50.0 |
| Batch size (par device) | 2 |
| Gradient accumulation | 16 |
| Batch effectif | 32 |
| Learning rate | 5e-5 |
| Weight decay | 0.005 |
| Optimiseur | AdamW (par défaut Trainer) |
| Précision | fp16 |
| Gradient checkpointing | Oui (use_reentrant=False) |
| Feature encoder | gelé (freeze_feature_encoder) |
| Attention dropout | 0.1 |
| Hidden dropout | 0.1 |
| Feat proj dropout | 0.0 |
| Mask time prob | 0.05 |
| LayerDrop | 0.1 |
| CTC loss reduction | mean |
| Durée d'entraînement | 1 day(s), 01:51:47 |
| Vitesse | 21.547 samples/s, 0.673 steps/s |
adiren7/darija_speech_to_textmohamedmou/moroccan-darija-asr-dataset-splitafyfbadreddine77/darija-asr-datasetntariklk/darija-merged-asratlasia/Moroccan-Darija-Wiki-Audio-DatasetDatasmartly/moroccan_darija_audioRHEZLOUNE/darija-restaurant-audioanaszil/Segmented-Moroccan-Darija-Wiki-Audio-Datasetamineouaki/dialect_maroccan_data_cleanedFiltre appliqué : durée audio ≤ 15s.
from transformers import HubertForCTC, Wav2Vec2Processor
import librosa
import torch
processor = Wav2Vec2Processor.from_pretrained("amineouaki/hubert-darija-combined")
model = HubertForCTC.from_pretrained("amineouaki/hubert-darija-combined")
speech, _ = librosa.load("audio.wav", sr=16000)
inputs = processor(speech, sampling_rate=16000, return_tensors="pt", padding=True)
with torch.no_grad():
logits = model(**inputs).logits
pred_ids = torch.argmax(logits, dim=-1)
print(processor.decode(pred_ids[0]))
Amine Ouakib
Base model
omarxadel/hubert-large-arabic-egyptian