whisper-medium-darija-combined

Fine-tuning de openai/whisper-medium 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!

Résultats finaux

Métrique	Valeur
WER (Word Error Rate)	0.0346 (3.46%)
CER (Character Error Rate)	0.0135 (1.35%)

Stats d'entraînement

Stat	Valeur
Époques	10.0
Train loss (moyenne)	0.07550942687588065
Durée d'entraînement	20:20:27
Vitesse	5.479 samples/s, 0.171 steps/s

Matériel d'entraînement

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

Datasets utilisés

• adiren7/darija_speech_to_text
• mohamedmou/moroccan-darija-asr-dataset-split
• afyfbadreddine77/darija-asr-dataset
• ntariklk/darija-merged-asr
• atlasia/Moroccan-Darija-Wiki-Audio-Dataset
• Datasmartly/moroccan_darija_audio
• RHEZLOUNE/darija-restaurant-audio
• anaszil/Segmented-Moroccan-Darija-Wiki-Audio-Dataset
• amineouaki/dialect_maroccan_data_cleaned

Filtre appliqué : durée audio ≤ 15s.

Utilisation

from transformers import WhisperForConditionalGeneration, WhisperProcessor
import librosa
import torch

processor = WhisperProcessor.from_pretrained("amineouaki/whisper-medium-darija-combined")
model = WhisperForConditionalGeneration.from_pretrained("amineouaki/whisper-medium-darija-combined")

speech, _ = librosa.load("audio.wav", sr=16000)
inputs = processor(speech, sampling_rate=16000, return_tensors="pt")

with torch.no_grad():
    generated_ids = model.generate(inputs.input_features, max_length=225)

print(processor.batch_decode(generated_ids, skip_special_tokens=True)[0])

Auteur

Amine Ouakib

• Email : amineouakib3@gmail.com
• GitHub : ouakibamine
• LinkedIn : amine-ouakib