WIP: RMVPE

server/voice_changer/DiffusionSVC/pipeline/Pipeline.py

@@ -9,7 +9,6 @@ from Exceptions import (
     NotEnoughDataExtimateF0,
 )
 from voice_changer.DiffusionSVC.inferencer.Inferencer import Inferencer
-from voice_changer.DiffusionSVC.inferencer.diffusion_svc_model.F0Extractor import F0_Extractor
 from voice_changer.DiffusionSVC.pitchExtractor.PitchExtractor import PitchExtractor
 
 from voice_changer.RVC.embedder.Embedder import Embedder

server/voice_changer/DiffusionSVC/pitchExtractor/RMVPEPitchExtractor.py

@@ -3,6 +3,7 @@ import numpy as np
 from const import PitchExtractorType
 from voice_changer.DiffusionSVC.pitchExtractor.PitchExtractor import PitchExtractor
 from voice_changer.DiffusionSVC.pitchExtractor.rmvpe.rmvpe import RMVPE
+from scipy.ndimage import zoom
 
 
 class RMVPEPitchExtractor(PitchExtractor):
@@ -24,7 +25,9 @@ class RMVPEPitchExtractor(PitchExtractor):
     def extract(self, audio: torch.Tensor, pitch, f0_up_key, window, silence_front=0):
         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)):]
+        silented_frames = int(audio.size(0) // window) + 1
 
         print("[RMVPE AUDI]", audio.device)
         print("[RMVPE RMVPE]", self.rmvpe.device)
@@ -47,6 +50,9 @@ class RMVPEPitchExtractor(PitchExtractor):
         # pd = torchcrepe.filter.median(pd, 3)
         # f0[pd < 0.1] = 0
         # f0 = f0.squeeze()
+        resize_factor = silented_frames / len(f0)
+        f0 = zoom(f0, resize_factor, order=0)
+
         pitch[-f0.shape[0]:] = f0[:pitch.shape[0]]
         f0 = pitch
 

server/voice_changer/DiffusionSVC/pitchExtractor/rmvpe/rmvpe.py

@@ -249,7 +249,7 @@ class E2E(nn.Module):
             )
         # else:
         #     self.fc = nn.Sequential(
-        #         nn.Linear(3 * N_MELS, N_CLASS), nn.Dropout(0.25), nn.Sigmoid() 
+        #         nn.Linear(3 * N_MELS, N_CLASS), nn.Dropout(0.25), nn.Sigmoid()
         #     )
 
     def forward(self, mel):
@@ -392,7 +392,7 @@ class RMVPE:
             hidden = hidden.astype("float32")
         f0 = self.decode(hidden, thred=thred)
         return f0
-    
+
     def to_local_average_cents(self, salience, thred=0.05):
         # t0 = ttime()
         center = np.argmax(salience, axis=1)  # 帧长#index

server/voice_changer/utils/VoiceChangerParams.py

@@ -15,4 +15,3 @@ class VoiceChangerParams:
     crepe_onnx_full: str
     crepe_onnx_tiny: str
     rmvpe: str
-