Update app.py

app.py

@@ -5,63 +5,47 @@ import torch
 import logging
 import soundfile as sf
 from transformers import AutoModelForAudioClassification
-from librosa.effects import pitch_shift, time_stretch
 
-# Enhanced logging for detailed debugging and information
 logging.basicConfig(level=logging.INFO)
 
-# Load your model (Ensure the model path is correct)
 model_path = "./"
 model = AutoModelForAudioClassification.from_pretrained(model_path)
 
-def augment_and_extract_features(audio_path, sr=16000, n_mfcc=40, pitch_shift_steps=4, stretch_rate=1.2):
-    """
-    Load, augment, and extract comprehensive features from an audio file.
-    Features: MFCC, Chroma, Mel Spectrogram, Spectral Contrast, Tonnetz.
-    Augmentations: Pitch Shift and Time Stretch for nuanced variability.
-    """
-    # Load audio
+def augment_and_extract_features(audio_path, sr=16000, n_mfcc=40, n_fft=2048, hop_length=512):
     y, sr = librosa.load(audio_path, sr=sr)
+    y_augmented = librosa.effects.pitch_shift(y, sr=sr, n_steps=4)
+    y_augmented = librosa.effects.time_stretch(y_augmented, rate=1.2)
 
-    # Audio Augmentation
-    y_pitched = pitch_shift(y, sr=sr, n_steps=pitch_shift_steps)  # Apply pitch shift
-    y_stretched = time_stretch(y_pitched, rate=stretch_rate)  # Apply time stretch
+    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)
 
-    # Feature Extraction
-    mfcc = librosa.feature.mfcc(y=y_stretched, sr=sr, n_mfcc=n_mfcc)
-    chroma = librosa.feature.chroma_stft(y=y_stretched, sr=sr)
-    mel = librosa.feature.melspectrogram(y=y_stretched, sr=sr)
-    contrast = librosa.feature.spectral_contrast(y=y_stretched, sr=sr)
-    tonnetz = librosa.feature.tonnetz(y=librosa.effects.harmonic(y_stretched), sr=sr)
-
-    # Aggregate features
-    features = np.concatenate([mfcc, chroma, mel, contrast, tonnetz], axis=0)
-    features_normalized = StandardScaler().fit_transform(features.T).T  # Normalize
-    features_tensor = torch.tensor(features_normalized).float().unsqueeze(0)  # Tensor conversion
+    features = np.concatenate((mfcc, chroma, mel, contrast, tonnetz), axis=0)
+    features = (features - np.mean(features, axis=1, keepdims=True)) / np.std(features, axis=1, keepdims=True)
 
+    features_tensor = torch.tensor(features).float().unsqueeze(0)
     return features_tensor
 
 def predict_voice(audio_file_path):
-    """
-    Make a voice authenticity prediction based on augmented and extracted audio features.
-    """
     try:
         features_tensor = augment_and_extract_features(audio_file_path)
         with torch.no_grad():
             outputs = model(features_tensor)
+
         logits = outputs.logits
         predicted_index = logits.argmax()
         label = model.config.id2label[predicted_index.item()]
         confidence = torch.softmax(logits, dim=1).max().item() * 100
-        result = f"Prediction: {label} with {confidence:.2f}% confidence."
-        logging.info(result)
-        return result
+
+        result = f"The voice is classified as '{label}' with a confidence of {confidence:.2f}%."
+        logging.info("Prediction successful.")
     except Exception as e:
-        error_msg = f"Error during processing: {e}"
-        logging.error(error_msg)
-        return error_msg
+        result = f"Error during processing: {e}"
+        logging.error(result)
+    return result
 
-# Gradio interface for easy and interactive model testing
 iface = gr.Interface(
     fn=predict_voice,
     inputs=gr.Audio(label="Upload Audio File", type="filepath"),