replace parselmouth with pyworld.dio

inference/infer_tool.py

@@ -10,6 +10,7 @@
 import numpy as np
 # import onnxruntime
 import parselmouth
+import pyworld
 import soundfile
 import torch
 import torchaudio
@@ -89,29 +90,20 @@ def resize2d_f0(x, target_len):
     res = np.nan_to_num(target)
     return res
 
-def get_f0(x, p_len,f0_up_key=0):
-
-    time_step = 160 / 16000 * 1000
-    f0_min = 50
-    f0_max = 1100
-    f0_mel_min = 1127 * np.log(1 + f0_min / 700)
-    f0_mel_max = 1127 * np.log(1 + f0_max / 700)
-
-    f0 = parselmouth.Sound(x, 16000).to_pitch_ac(
-        time_step=time_step / 1000, voicing_threshold=0.6,
-        pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency']
-
-    pad_size=(p_len - len(f0) + 1) // 2
-    if(pad_size>0 or p_len - len(f0) - pad_size>0):
-        f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
-
+def compute_f0(x, c_len,f0_up_key=0):
+    # x, sr = librosa.load(path, sr=48000)
+    f0, t = pyworld.dio(
+        x.astype(np.double),
+        fs=16000,
+        f0_ceil=800,
+        frame_period=1000 * 160 / 16000,
+    )
+    f0 = pyworld.stonemask(x.astype(np.double), f0, t, 16000)
+    for index, pitch in enumerate(f0):
+        f0[index] = round(pitch, 1)
+    assert abs(c_len - x.shape[0]//160) < 3, (c_len, f0.shape)
     f0 *= pow(2, f0_up_key / 12)
-    f0_mel = 1127 * np.log(1 + f0 / 700)
-    f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1
-    f0_mel[f0_mel <= 1] = 1
-    f0_mel[f0_mel > 255] = 255
-    f0_coarse = np.rint(f0_mel).astype(np.int)
-    return f0_coarse, f0
+    return None, resize2d_f0(f0, c_len)
 
 def clean_pitch(input_pitch):
     num_nan = np.sum(input_pitch == 1)
@@ -203,7 +195,7 @@ def get_unit_pitch(self, in_path, tran):
         if len(source.shape) == 2 and source.shape[1] >= 2:
             source = torch.mean(source, dim=0).unsqueeze(0)
         soft = self.get_units(source, sr).squeeze(0).cpu().numpy()
-        f0_coarse, f0 = get_f0(source.cpu().numpy()[0], soft.shape[0]*2, tran)
+        f0_coarse, f0 = compute_f0(source.cpu().numpy()[0], soft.shape[0]*2, tran)
         f0 = resize2d_f0(f0, soft.shape[0]*3)
         return soft, f0
 

preprocess_hubert_f0.py

@@ -108,7 +108,7 @@ def process(filename):
         c = torch.load(save_name)
     f0path = filename+".f0.npy"
     if not os.path.exists(f0path):
-        cf0, f0 = get_f0(filename, c.shape[-1] * 3)
+        cf0, f0 = compute_f0(filename, c.shape[-1] * 3)
         np.save(f0path, f0)