WIP: dio

server/voice_changer/DiffusionSVC/pitchExtractor/DioPitchExtractor.py

@@ -1,8 +1,9 @@
 import pyworld
 import numpy as np
 from const import PitchExtractorType
+import torch
 
-from voice_changer.RVC.pitchExtractor.PitchExtractor import PitchExtractor
+from voice_changer.DiffusionSVC.pitchExtractor.PitchExtractor import PitchExtractor
 
 
 class DioPitchExtractor(PitchExtractor):
@@ -10,40 +11,35 @@ class DioPitchExtractor(PitchExtractor):
     def __init__(self):
         super().__init__()
         self.pitchExtractorType: PitchExtractorType = "dio"
+        self.f0_min = 50
+        self.f0_max = 1100
+        self.sapmle_rate = 16000
+        self.uv_interp = True
 
-    def extract(self, audio, pitchf, f0_up_key, sr, window, silence_front=0):
+    def extract(self, audio: torch.Tensor, pitch, f0_up_key, sr, window, silence_front=0):
         audio = audio.detach().cpu().numpy()
-        n_frames = int(len(audio) // window) + 1  # NOQA
+        start_frame = int(silence_front * self.sapmle_rate / window)
-        start_frame = int(silence_front * sr / window)
+        real_silence_front = start_frame * window / self.sapmle_rate
-        real_silence_front = start_frame * window / sr
+        audio = audio[int(np.round(real_silence_front * self.sapmle_rate)):]
-
-        silence_front_offset = max(min(int(np.round(real_silence_front * sr)), len(audio) - 3000), 0)
-        audio = audio[silence_front_offset:]
-
-        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, t = pyworld.dio(
             audio.astype(np.double),
-            sr,
+            self.sapmle_rate,
-            f0_floor=f0_min,
+            f0_floor=self.f0_min,
-            f0_ceil=f0_max,
+            f0_ceil=self.f0_max,
             channels_in_octave=2,
-            frame_period=10,
+            frame_period=(1000 * window / self.sapmle_rate)
         )
-        f0 = pyworld.stonemask(audio.astype(np.double), _f0, t, sr)
+        f0 = pyworld.stonemask(audio.astype(np.double), _f0, t, self.sapmle_rate)
+        pitch[-f0.shape[0]:] = f0[:pitch.shape[0]]
+        f0 = pitch
 
-        f0 *= pow(2, f0_up_key / 12)
+        if self.uv_interp:
-        pitchf[-f0.shape[0]:] = f0[:pitchf.shape[0]]
+            uv = f0 == 0
-        f0bak = pitchf.copy()
+            if len(f0[~uv]) > 0:
-        f0_mel = 1127 * np.log(1 + f0bak / 700)
+                f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
-        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
+            f0[f0 < self.f0_min] = self.f0_min
-            f0_mel_max - f0_mel_min
-        ) + 1
-        f0_mel[f0_mel <= 1] = 1
-        f0_mel[f0_mel > 255] = 255
-        pitch_coarse = np.rint(f0_mel).astype(int)
 
-        return pitch_coarse, pitchf
+        f0 = f0 * 2 ** (float(f0_up_key) / 12)
+
+        return f0

server/voice_changer/DiffusionSVC/pitchExtractor/HarvestPitchExtractor.py

@@ -2,7 +2,8 @@ import pyworld
 import numpy as np
 from const import PitchExtractorType
 import torch
-from voice_changer.RVC.pitchExtractor.PitchExtractor import PitchExtractor
+
+from voice_changer.DiffusionSVC.pitchExtractor.PitchExtractor import PitchExtractor
 
 
 class HarvestPitchExtractor(PitchExtractor):
@@ -15,25 +16,25 @@ class HarvestPitchExtractor(PitchExtractor):
         self.sapmle_rate = 16000
         self.uv_interp = True
 
-    def extract(self, audio: torch.Tensor, pitchf, f0_up_key, sr, window, silence_front=0):
+    def extract(self, audio: torch.Tensor, pitch, f0_up_key, sr, window, silence_front=0):
         audio = audio.detach().cpu().numpy()
         start_frame = int(silence_front * self.sapmle_rate / window)
         real_silence_front = start_frame * window / self.sapmle_rate
         audio = audio[int(np.round(real_silence_front * self.sapmle_rate)):]
         f0, _ = pyworld.harvest(
             audio.astype('double'),
-            16000,
+            self.sapmle_rate,
-            f0_floor=50,
+            f0_floor=self.f0_min,
-            f0_ceil=1100,
+            f0_ceil=self.f0_max,
             frame_period=(1000 * window / self.sapmle_rate))
-        pitchf[-f0.shape[0]:] = f0[:pitchf.shape[0]]
+        pitch[-f0.shape[0]:] = f0[:pitch.shape[0]]
-        f0 = pitchf
+        f0 = pitch
 
         if self.uv_interp:
             uv = f0 == 0
             if len(f0[~uv]) > 0:
                 f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
-            f0[f0 < 50] = 50
+            f0[f0 < self.f0_min] = self.f0_min
 
         f0 = f0 * 2 ** (float(f0_up_key) / 12)
 

server/voice_changer/DiffusionSVC/pitchExtractor/PitchExtractorManager.py

@@ -1,5 +1,6 @@
 from typing import Protocol
 from const import PitchExtractorType
+from voice_changer.DiffusionSVC.pitchExtractor.DioPitchExtractor import DioPitchExtractor
 from voice_changer.DiffusionSVC.pitchExtractor.HarvestPitchExtractor import HarvestPitchExtractor
 from voice_changer.DiffusionSVC.pitchExtractor.PitchExtractor import PitchExtractor
 from voice_changer.utils.VoiceChangerParams import VoiceChangerParams
@@ -26,8 +27,8 @@ class PitchExtractorManager(Protocol):
     ) -> PitchExtractor:
         if pitchExtractorType == "harvest":
             return HarvestPitchExtractor()
-        # elif pitchExtractorType == "dio":
+        elif pitchExtractorType == "dio":
-        #     return DioPitchExtractor()
+            return DioPitchExtractor()
         # elif pitchExtractorType == "crepe":
         #     return CrepePitchExtractor()
         # elif pitchExtractorType == "crepe_tiny":