Update app.py

app.py

@@ -2,58 +2,51 @@ import gradio as gr
 import librosa
 import numpy as np
 import torch
-import torch.nn.functional as F
 import logging
 from transformers import AutoModelForAudioClassification
+import soundfile as sf
 
 # Configure logging for debugging and information
 logging.basicConfig(level=logging.INFO)
 
-# Model loading from the specified local path
-local_model_path = "./"
-model = AutoModelForAudioClassification.from_pretrained(local_model_path)
+# Load the model
+model_path = "./"
+model = AutoModelForAudioClassification.from_pretrained(model_path)
 
-def custom_feature_extraction(audio_file_path, sr=16000, n_mels=128, n_fft=2048, hop_length=512, target_length=1024):
-    """
-    Custom feature extraction using Mel spectrogram, tailored for models trained on datasets like AudioSet.
-    Args:
-        audio_file_path: Path to the audio file for prediction.
-        sr: Target sampling rate for the audio file.
-        n_mels: Number of Mel bands to generate.
-        n_fft: Length of the FFT window.
-        hop_length: Number of samples between successive frames.
-        target_length: Expected length of the Mel spectrogram in the time dimension.
-    Returns:
-        A tensor representation of the Mel spectrogram features.
-    """
-    waveform, sample_rate = librosa.load(audio_file_path, sr=sr)
-    S = librosa.feature.melspectrogram(y=waveform, sr=sample_rate, n_mels=n_mels, n_fft=n_fft, hop_length=hop_length)
-    S_DB = librosa.power_to_db(S, ref=np.max)
-    mel_tensor = torch.tensor(S_DB).float()
-
-    # Ensure the tensor matches the expected sequence length
-    current_length = mel_tensor.shape[1]
-    if current_length > target_length:
-        mel_tensor = mel_tensor[:, :target_length]  # Truncate if longer
-    elif current_length < target_length:
-        padding = target_length - current_length
-        mel_tensor = F.pad(mel_tensor, (0, padding), "constant", 0)  # Pad if shorter
-
-    mel_tensor = mel_tensor.unsqueeze(0)  # Add batch dimension for compatibility with model
-    return mel_tensor
+def augment_and_extract_features(audio_path, output_path=None, sr=16000, n_mfcc=40, n_fft=2048, hop_length=512):
+    # Load and augment the audio file
+    y, sr = librosa.load(audio_path, sr=sr)
+    y_augmented = librosa.effects.pitch_shift(y, sr, n_steps=4)  # Pitch shifting
+    y_augmented = librosa.effects.time_stretch(y_augmented, rate=1.2)  # Time stretching
+    
+    # Save the augmented audio if an output path is provided
+    if output_path is not None:
+        sf.write(output_path, y_augmented, sr)
+    
+    # Extract features
+    mfcc = librosa.feature.mfcc(y=y_augmented, sr=sr, n_mfcc=n_mfcc, n_fft=n_fft, hop_length=hop_length)
+    chroma = librosa.feature.chroma_stft(y=y_augmented, sr=sr, n_fft=n_fft, hop_length=hop_length)
+    mel = librosa.feature.melspectrogram(y=y_augmented, sr=sr, n_fft=n_fft, hop_length=hop_length)
+    contrast = librosa.feature.spectral_contrast(y=y_augmented, sr=sr, n_fft=n_fft, hop_length=hop_length)
+    tonnetz = librosa.feature.tonnetz(y=librosa.effects.harmonic(y_augmented), sr=sr)
+    
+    # Combine all features
+    features = np.concatenate((mfcc, chroma, mel, contrast, tonnetz), axis=0)
+    
+    # Normalize features
+    features = (features - np.mean(features, axis=1, keepdims=True)) / np.std(features, axis=1, keepdims=True)
+    
+    # Convert to tensor
+    features_tensor = torch.tensor(features).float().unsqueeze(0)  # Add batch dimension
+    return features_tensor
 
 def predict_voice(audio_file_path):
-    """
-    Predicts the audio class using a pre-trained model and custom feature extraction.
-    Args:
-        audio_file_path: Path to the audio file for prediction.
-    Returns:
-        A string containing the predicted class and confidence level.
-    """
     try:
-        features = custom_feature_extraction(audio_file_path)
+        features_tensor = augment_and_extract_features(audio_file_path)
+        
         with torch.no_grad():
-            outputs = model(features)
+            outputs = model(features_tensor)
+        
         logits = outputs.logits
         predicted_index = logits.argmax()
         label = model.config.id2label[predicted_index.item()]
@@ -67,7 +60,6 @@ def predict_voice(audio_file_path):
     
     return result
 
-# Setting up the Gradio interface
 iface = gr.Interface(
     fn=predict_voice,
     inputs=gr.Audio(label="Upload Audio File", type="filepath"),
@@ -76,5 +68,4 @@ iface = gr.Interface(
     description="Detects whether a voice is real or AI-generated. Upload an audio file to see the results."
 )
 
-# Launching the interface
 iface.launch()