Update app.py

app.py

@@ -7,55 +7,60 @@ import logging
 from transformers import AutoModelForAudioClassification
 import random
 
+# Configure logging
 logging.basicConfig(level=logging.INFO)
 
+# Load the model
 model_path = "./"
 model = AutoModelForAudioClassification.from_pretrained(model_path)
 
 def custom_feature_extraction(waveform, sr, n_mels=128, n_fft=2048, hop_length=512, target_length=1024):
+    # Generate Mel spectrogram
     S = librosa.feature.melspectrogram(y=waveform, sr=sr, n_mels=n_mels, n_fft=n_fft, hop_length=hop_length)
     S_DB = librosa.power_to_db(S, ref=np.max)
     
+    # Pitch feature (using piptrack to estimate pitches and then averaging)
     pitches, _ = librosa.piptrack(y=waveform, sr=sr, n_fft=n_fft, hop_length=hop_length)
-    pitches = pitches.mean(axis=0, keepdims=True)
+    pitch_mean = np.mean(pitches, axis=0, keepdims=True)
+
+    # Spectral centroid
     spectral_centroids = librosa.feature.spectral_centroid(y=waveform, sr=sr, n_fft=n_fft, hop_length=hop_length)
 
-    features = np.concatenate([S_DB, pitches, spectral_centroids], axis=0)
+    # Concatenate features and normalize
+    features = np.concatenate([S_DB, pitch_mean, spectral_centroids], axis=0)
     features_tensor = torch.tensor(features).float()
 
-    current_length = features_tensor.shape[1]
-    if current_length > target_length:
+    # Adjust the tensor's length
+    if features_tensor.shape[1] > target_length:
         features_tensor = features_tensor[:, :target_length]
-    elif current_length < target_length:
-        padding = target_length - current_length
-        features_tensor = F.pad(features_tensor, (0, padding), "constant", 0)
+    elif features_tensor.shape[1] < target_length:
+        padding = target_length - features_tensor.shape[1]
+        features_tensor = F.pad(features_tensor, (0, padding), 'constant', 0)
 
     return features_tensor.unsqueeze(0)
 
 def apply_time_shift(waveform, max_shift_fraction=0.1):
-    shift = int(max_shift_fraction * waveform.size)
-    shift = random.randint(-shift, shift)
+    shift_amount = int(max_shift_fraction * len(waveform))
+    shift = random.randint(-shift_amount, shift_amount)
     return np.roll(waveform, shift)
 
-def predict_voice(audio_file_path):
+def predict_voice(audio_file):
     try:
-        waveform, sample_rate = librosa.load(audio_file_path, sr=None)
+        waveform, sample_rate = librosa.load(audio_file, sr=None)
         augmented_waveform = apply_time_shift(waveform)
-        
+
         original_features = custom_feature_extraction(waveform, sample_rate)
         augmented_features = custom_feature_extraction(augmented_waveform, sample_rate)
 
         with torch.no_grad():
             outputs_original = model(original_features)
             outputs_augmented = model(augmented_features)
-
             logits = (outputs_original.logits + outputs_augmented.logits) / 2
             predicted_index = logits.argmax()
             label = model.config.id2label[predicted_index.item()]
             confidence = torch.softmax(logits, dim=1).max().item() * 100
 
         result = f"The voice is classified as '{label}' with a confidence of {confidence:.2f}%."
-        logging.info("Prediction successful.")
     except Exception as e:
         result = f"Error during processing: {e}"
         logging.error(result)
@@ -64,8 +69,8 @@ def predict_voice(audio_file_path):
 
 iface = gr.Interface(
     fn=predict_voice,
-    inputs=gr.inputs.Audio(label="Upload Audio File", type="filepath"),
-    outputs=gr.outputs.Textbox(label="Prediction"),
+    inputs=gr.Audio(label="Upload Audio File", type="filepath"),
+    outputs=gr.Textbox(label="Prediction"),
     title="Voice Authenticity Detection",
     description="Detects whether a voice is real or AI-generated. Upload an audio file to see the results."
 )