Update pitch/inference.py

pitch/inference.py

@@ -4,7 +4,24 @@ import librosa
 import argparse
 import numpy as np
 import parselmouth
-# pip install praat-parselmouth
+from pitch.rmvpe import RMVPE
+import torch
+import numpy as np
+
+def compute_f0_rmvpe(path, model_path='rmvpe.pt'):
+    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+    model = RMVPE(model_path, is_half=False, device=device)
+    x, sr = librosa.load(path, sr=16000)
+    assert sr == 16000
+    f0 = model.infer_from_audio(x)
+    # Adjust to 10 ms resolution by repeating (assuming 20 ms output)
+    f0 = np.repeat(f0, 2)
+    # Pad f0 to match other functions
+    lpad = 6
+    rpad = 6
+    f0 = np.pad(f0, [lpad, rpad], mode='constant')
+    return f0
+    
 
 def compute_f0_mouth(path):
     x, sr = librosa.load(path, sr=16000)