app.py

import gradio as gr
from transformers import pipeline
import os
import numpy as np
import torch

# Load the model
print("Loading model...")
model_id = "badrex/mms-300m-arabic-dialect-identifier"
classifier = pipeline("audio-classification", model=model_id)
print("Model loaded successfully")

# Define dialect mapping
dialect_mapping = {
    "MSA": "Modern Standard Arabic",
    "Egyptian": "Egyptian Arabic",
    "Gulf": "Gulf Arabic",
    "Levantine": "Levantine Arabic",
    "Maghrebi": "Maghrebi Arabic"
}

def predict_dialect(audio):
    if audio is None:
        return {"Error": 1.0}
    
    # The audio input from Gradio is a tuple of (sample_rate, audio_array)
    sr, audio_array = audio
    
    # Process the audio input
    if len(audio_array.shape) > 1:
        audio_array = audio_array.mean(axis=1)  # Convert stereo to mono

    # Convert audio to float32 if it's not already (fix for Chrome recording issue)
    if audio_array.dtype != np.float32:
        # Normalize to [-1, 1] range as expected by the model
        if audio_array.dtype == np.int16:
            audio_array = audio_array.astype(np.float32) / 32768.0
        else:
            audio_array = audio_array.astype(np.float32)
    
    print(f"Processing audio: sample rate={sr}, shape={audio_array.shape}")
    
    # Classify the dialect
    predictions = classifier({"sampling_rate": sr, "raw": audio_array})
    
    # Format results for display
    results = {}
    for pred in predictions:
        dialect_name = dialect_mapping.get(pred['label'], pred['label'])
        results[dialect_name] = float(pred['score'])
    
    return results

# Manually prepare example file paths without metadata
examples = []
examples_dir = "examples"
if os.path.exists(examples_dir):
    for filename in os.listdir(examples_dir):
        if filename.endswith((".wav", ".mp3", ".ogg")):
            examples.append([os.path.join(examples_dir, filename)])
    
    print(f"Found {len(examples)} example files")
else:
    print("Examples directory not found")



# Custom CSS for better styling
custom_css = """
<style>
.centered-content {
    text-align: center;
    max-width: 800px;
    margin: 0 auto;
    padding: 20px;
}

.logo-image {
    width: 200px;
    height: auto;
    margin: 20px auto;
    display: block;
}

.description-text {
    font-size: 16px;
    line-height: 1.6;
    margin-bottom: 20px;
}

.dialect-list {
    font-size: 15px;
    line-height: 1.8;
    text-align: left;
    max-width: 600px;
    margin: 0 auto;
}

.highlight-text {
    font-size: 16px;
    color: #2563eb;
    margin: 20px 0;
}

.footer-text {
    font-size: 13px;
    color: #6b7280;
    margin-top: 20px;
}
</style>
"""

"""
<p style="font-size: 15px; line-height: 1.8;">
                    <strong>The following Arabic language varieties are supported:</strong>
                    <br><br>
                    ✦ <strong>Modern Standard Arabic (MSA)</strong> - The formal language of media and education
                    <br>
                    ✦ <strong>Egyptian Arabic</strong> - The dialect of Cairo, Alexandria, and popular Arabic cinema
                    <br>
                    ✦ <strong>Gulf Arabic</strong> - Spoken across Saudi Arabia, UAE, Kuwait, Qatar, Bahrain, and Oman
                    <br>
                    ✦ <strong>Levantine Arabic</strong> - The dialect of Syria, Lebanon, Jordan, and Palestine
                    <br>
                    ✦ <strong>Maghrebi Arabic</strong> - The distinctive varieties of Morocco, Algeria, Tunisia, and Libya
                    </p>
                    <br>
"""

# Create the Gradio interface
demo = gr.Interface(
    fn=predict_dialect,
    inputs=gr.Audio(),
    outputs=gr.Label(num_top_classes=5, label="Predicted Dialect"),
    title="Tamyïz 🍉 Arabic Dialect Identification in Speech",
    description="""
        <div class="centered-content">
            <img src="https://huggingface.co/badrex/mms-300m-arabic-dialect-identifier/resolve/main/assets/logo.png" alt="Logo" style="width: 200px; height: auto; margin-bottom: 20px;">
            <br>
            <div>
                <p>
                This is a demo for the accurate and robust Transformer-based <a href="https://huggingface.co/badrex/mms-300m-arabic-dialect-identifier" style="color: #FF5349;">model</a> for Spoken Arabic Dialect Identification (ADI). 
                From just a short audio clip (5-10 seconds), the model can identify Modern Standard Arabic (<strong>MSA</strong>) as well as four major regional Arabic varieties: <strong>Egyptian</strong> Arabic, <strong>Gulf</strong> Arabic, <strong>Levantine</strong> Arabic, and <strong>Maghrebi</strong> Arabic. 
                <br>                   
                <p>
                Simply <strong>upload an audio file</strong> 📀 or <strong>record yourself speaking</strong> ⏯️⏺️ to try out the model!
                </p>
                <br>
                <p>
                Developed with ❤️🤍💚 by <a href="https://badrex.github.io/" style="color: #2563eb;">Badr Alabsi</a>
                </p>
            </div>
        </div>
        """,
    examples=examples if examples else None,
    cache_examples=False,  # Disable caching to avoid issues
    flagging_mode=None
)

# Launch the app
demo.launch(share=True)