app.py

import gradio as gr
import numpy as np
import torch
import librosa
from transformers import AutoFeatureExtractor, AutoModelForAudioClassification
import matplotlib.pyplot as plt
from matplotlib.colors import Normalize

# Constants
SAMPLING_RATE = 16000
MODEL_NAME = "MIT/ast-finetuned-audioset-10-10-0.4593"
DEFAULT_THRESHOLD = 0.7

# Load model and feature extractor
feature_extractor = AutoFeatureExtractor.from_pretrained(MODEL_NAME)
model = AutoModelForAudioClassification.from_pretrained(MODEL_NAME)

def analyze_audio(audio_array, threshold=DEFAULT_THRESHOLD):
    """
    Process audio and detect anomalies
    Returns:
    - classification result
    - confidence score
    - spectrogram visualization
    """
    try:
        # Resample if needed and convert to mono
        if isinstance(audio_array, tuple):
            sr, audio = audio_array
            audio = librosa.resample(audio, orig_sr=sr, target_sr=SAMPLING_RATE)
        else:
            audio = audio_array
            
        if len(audio.shape) > 1:
            audio = librosa.to_mono(audio)
            
        # Extract features
        inputs = feature_extractor(
            audio, 
            sampling_rate=SAMPLING_RATE, 
            return_tensors="pt",
            padding=True,
            return_attention_mask=True
        )
        
        # Run inference
        with torch.no_grad():
            outputs = model(**inputs)
            logits = outputs.logits
            probs = torch.softmax(logits, dim=-1)
            
        # Get predicted class and confidence
        predicted_class = "Normal" if probs[0][0] > threshold else "Anomaly"
        confidence = probs[0][0].item() if predicted_class == "Normal" else 1 - probs[0][0].item()
        
        # Create spectrogram visualization
        spectrogram = librosa.feature.melspectrogram(
            y=audio, 
            sr=SAMPLING_RATE,
            n_mels=128,
            fmax=8000
        )
        db_spec = librosa.power_to_db(spectrogram, ref=np.max)
        
        plt.figure(figsize=(10, 4))
        plt.imshow(db_spec, aspect='auto', origin='lower', 
                  norm=Normalize(vmin=-80, vmax=0),
                  cmap='viridis')
        plt.colorbar(format='%+2.0f dB')
        plt.title('Mel Spectrogram')
        plt.tight_layout()
        plt.savefig('spec.png', bbox_inches='tight')
        plt.close()
        
        return (
            predicted_class,
            f"{confidence:.1%}",
            'spec.png',
            str(probs.tolist()[0])
        )
        
    except Exception as e:
        return f"Error: {str(e)}", "", "", ""

# Gradio interface
with gr.Blocks(title="Industrial Audio Analyzer", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🏭 Industrial Equipment Sound Analyzer
    ### Powered by Audio Spectrogram Transformer (AST)
    """)
    
    with gr.Row():
        with gr.Column():
            audio_input = gr.Audio(
                label="Upload Equipment Audio Recording",
                type="numpy",
                source="upload",
                show_download_button=True
            )
            threshold = gr.Slider(
                minimum=0.5,
                maximum=0.95,
                step=0.05,
                value=DEFAULT_THRESHOLD,
                label="Anomaly Detection Threshold",
                info="Higher values reduce false positives but may miss subtle anomalies"
            )
            analyze_btn = gr.Button("🔍 Analyze Sound", variant="primary")
            
            gr.Examples(
                examples=["examples/normal_machine.wav", "examples/anomalous_machine.wav"],
                inputs=audio_input,
                label="Sample Recordings"
            )
            
        with gr.Column():
            result_label = gr.Label(label="Detection Result")
            confidence = gr.Textbox(label="Confidence Score")
            spectrogram = gr.Image(label="Spectrogram Visualization")
            raw_probs = gr.Textbox(
                label="Model Output Probabilities",
                visible=False
            )
    
    analyze_btn.click(
        fn=analyze_audio,
        inputs=[audio_input, threshold],
        outputs=[result_label, confidence, spectrogram, raw_probs]
    )
    
    gr.Markdown("""
    ## How It Works
    - Upload audio recordings from industrial equipment
    - The AI analyzes sound patterns using spectrogram analysis
    - Detects anomalies indicating potential equipment issues
    
    **Tip**: For best results, use 5-10 second recordings of steady operation
    """)

if __name__ == "__main__":
    demo.launch()