Respair/Arabic_Vocal_Dialect_Classifier

Model Card for Model ID

Accuracy : ~91%

Note: The model may sometimes confuse closely related dialects.

Inference

# load the model

import torch
from transformers import AutoModelForAudioClassification
from transformers import AutoFeatureExtractor
import librosa

ckpt = "Respair/Arabic_Vocal_Dialect_Classifier"
feature_extractor = AutoFeatureExtractor.from_pretrained(ckpt)
model = AutoModelForAudioClassification.from_pretrained(ckpt).to('cuda').to(torch.bfloat16) # if your hardware supports it

# inference function
def classify_dialect(audio):
    
    wav = librosa.load(audio, sr=16_000)[0]
    inputs = feature_extractor(wav, sampling_rate=16_000, return_tensors="pt").to(torch.bfloat16).to('cuda')
    with torch.no_grad():
        logits = model(**inputs).logits
    

    predicted_class_ids = torch.argmax(logits).item()

    predicted_class_id = predicted_class_ids.item() if hasattr(predicted_class_ids, 'item') else int(predicted_class_ids)
    
    predicted_label = model.config.id2label[predicted_class_id]
        
    return predicted_label


# run the function

classify_dialect("path/to/audio.mp3") # output something like -> "egyptian"