Update app.py

app.py

@@ -8,12 +8,11 @@ import random
 
 model = AutoModelForAudioClassification.from_pretrained("./")
 
-def custom_feature_extraction(audio_file_path, sr=16000, n_mels=128, n_fft=2048, hop_length=512, target_length=1024):
-    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)
+def custom_feature_extraction(audio, sr=16000, n_mels=128, n_fft=2048, hop_length=512, target_length=1024):
+    S = librosa.feature.melspectrogram(y=audio, sr=sr, n_mels=n_mels, n_fft=n_fft, hop_length=hop_length)
     S_DB = librosa.power_to_db(S, ref=np.max)
-    pitches, _ = librosa.piptrack(y=waveform, sr=sample_rate, n_fft=n_fft, hop_length=hop_length)
-    spectral_centroids = librosa.feature.spectral_centroid(y=waveform, sr=sample_rate, n_fft=n_fft, hop_length=hop_length)
+    pitches, _ = librosa.piptrack(y=audio, sr=sr, n_fft=n_fft, hop_length=hop_length)
+    spectral_centroids = librosa.feature.spectral_centroid(y=audio, sr=sr, n_fft=n_fft, hop_length=hop_length)
     features = np.concatenate([S_DB, pitches, spectral_centroids], axis=0)
     features_tensor = torch.tensor(features).float()
     if features_tensor.shape[1] > target_length:
@@ -30,7 +29,7 @@ def predict_voice(audio_file_path):
     try:
         waveform, sample_rate = librosa.load(audio_file_path, sr=None)
         augmented_waveform = apply_time_shift(waveform)
-        original_features = custom_feature_extraction(audio_file_path, sr=sample_rate)
+        original_features = custom_feature_extraction(waveform, sr=sample_rate)
         augmented_features = custom_feature_extraction(augmented_waveform, sr=sample_rate)
         with torch.no_grad():
             outputs_original = model(original_features)