Update app.py

app.py

@@ -3,41 +3,39 @@ import librosa
 import numpy as np
 import torch
 import logging
-from transformers import AutoModelForAudioClassification, Wav2Vec2Processor
-import soundfile as sf
-from scipy.signal import butter, lfilter
+from transformers import AutoModelForAudioClassification
 
 logging.basicConfig(level=logging.INFO)
 
 model_path = "./"
 model = AutoModelForAudioClassification.from_pretrained(model_path)
-processor = Wav2Vec2Processor.from_pretrained(model_path)
 
-def butter_bandpass(lowcut, highcut, fs, order=5):
-    nyq = 0.5 * fs
-    low = lowcut / nyq
-    high = highcut / nyq
-    b, a = butter(order, [low, high], btype='band')
-    return b, a
+def augment_and_extract_features(audio_path, sr=16000, n_mfcc=40, n_fft=2048, hop_length=512, target_length=512):
+    y, sr = librosa.load(audio_path, sr=sr)
+    y_pitch_shifted = librosa.effects.pitch_shift(y, sr=sr, n_steps=4)
+    y_time_stretched = librosa.effects.time_stretch(y_pitch_shifted, rate=1.2)
 
-def butter_bandpass_filter(data, lowcut, highcut, fs, order=5):
-    b, a = butter_bandpass(lowcut, highcut, fs, order=order)
-    y = lfilter(b, a, data)
-    return y
+    mfcc = librosa.feature.mfcc(y=y_time_stretched, sr=sr, n_mfcc=n_mfcc, n_fft=n_fft, hop_length=hop_length)
+    chroma = librosa.feature.chroma_stft(y=y_time_stretched, sr=sr, n_fft=n_fft, hop_length=hop_length)
+    mel = librosa.feature.melspectrogram(y=y_time_stretched, sr=sr, n_fft=n_fft, hop_length=hop_length)
+    contrast = librosa.feature.spectral_contrast(y=y_time_stretched, sr=sr, n_fft=n_fft, hop_length=hop_length)
+    tonnetz = librosa.feature.tonnetz(y=librosa.effects.harmonic(y_time_stretched), sr=sr)
 
-def quantum_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_filtered = butter_bandpass_filter(y, 300, 3400, sr)
-    y_quantum = np.fft.fft(y_filtered)
-    y_ifft = np.fft.ifft(y_quantum).real
+    features = np.concatenate((mfcc, chroma, mel, contrast, tonnetz), axis=0)
+    features_normalized = (features - np.mean(features, axis=1, keepdims=True)) / np.std(features, axis=1, keepdims=True)
 
-    features_processor = processor(y_ifft, sampling_rate=sr, return_tensors="pt", padding=True)
+    if features_normalized.shape[1] > target_length:
+        features_normalized = features_normalized[:, :target_length]
+    else:
+        padding = target_length - features_normalized.shape[1]
+        features_normalized = np.pad(features_normalized, ((0, 0), (0, padding)), 'constant')
 
-    return features_processor.input_values
+    features_tensor = torch.tensor(features_normalized).float().unsqueeze(0)  # Add batch dimension
+    return features_tensor
 
 def predict_voice(audio_file_path):
     try:
-        features_tensor = quantum_augment_and_extract_features(audio_file_path)
+        features_tensor = augment_and_extract_features(audio_file_path)
         with torch.no_grad():
             outputs = model(features_tensor)
 
@@ -62,4 +60,3 @@ iface = gr.Interface(
 )
 
 iface.launch()
-