New Feature:

- contentvec onnx

server/downloader/SampleDownloader.py

@@ -169,5 +169,5 @@ def _downloadSamples(samples: list[ModelSamples], sampleModelIds: list[Tuple[str
                 slotInfo = RVCModelSlotGenerator._setInfoByONNX(slotInfo)
             else:
                 slotInfo = RVCModelSlotGenerator._setInfoByPytorch(slotInfo)
-            
+
             modelSlotManager.save_model_slot(targetSlotIndex, slotInfo)

server/downloader/WeightDownloader.py

@@ -7,6 +7,7 @@ from Exceptions import WeightDownladException
 
 
 def downloadWeight(voiceChangerParams: VoiceChangerParams):
+    content_vec_500_onnx = voiceChangerParams.content_vec_500_onnx
     hubert_base = voiceChangerParams.hubert_base
     hubert_base_jp = voiceChangerParams.hubert_base_jp
     hubert_soft = voiceChangerParams.hubert_soft
@@ -67,7 +68,6 @@ def downloadWeight(voiceChangerParams: VoiceChangerParams):
                 "position": 5,
             }
         )
-    
     if os.path.exists(crepe_onnx_tiny) is False:
         downloadParams.append(
             {
@@ -77,6 +77,15 @@ def downloadWeight(voiceChangerParams: VoiceChangerParams):
             }
         )
 
+    if os.path.exists(content_vec_500_onnx) is False:
+        downloadParams.append(
+            {
+                "url": "https://huggingface.co/wok000/weights_gpl/resolve/main/content-vec/contentvec-f.onnx",
+                "saveTo": content_vec_500_onnx,
+                "position": 7,
+            }
+        )
+
     with ThreadPoolExecutor() as pool:
         pool.map(download, downloadParams)
 

server/voice_changer/RVC/embedder/Embedder.py

@@ -1,18 +1,19 @@
-from typing import Any, Protocol
+from typing import Any
 
 import torch
 from torch import device
 
 from const import EmbedderType
+from voice_changer.RVC.embedder.EmbedderProtocol import EmbedderProtocol
 
 
-class Embedder(Protocol):
+class Embedder(EmbedderProtocol):
-    embedderType: EmbedderType = "hubert_base"
+    def __init__(self):
-    file: str
+        self.embedderType: EmbedderType = "hubert_base"
-    isHalf: bool = True
+        self.file: str
-    dev: device
+        self.dev: device
 
-    model: Any | None = None
+        self.model: Any | None = None
 
     def loadModel(self, file: str, dev: device, isHalf: bool = True):
         ...

server/voice_changer/RVC/embedder/EmbedderManager.py

@@ -5,6 +5,7 @@ from voice_changer.RVC.embedder.Embedder import Embedder
 from voice_changer.RVC.embedder.FairseqContentvec import FairseqContentvec
 from voice_changer.RVC.embedder.FairseqHubert import FairseqHubert
 from voice_changer.RVC.embedder.FairseqHubertJp import FairseqHubertJp
+from voice_changer.RVC.embedder.OnnxContentvec import OnnxContentvec
 from voice_changer.utils.VoiceChangerParams import VoiceChangerParams
 
 
@@ -27,10 +28,11 @@ class EmbedderManager:
             print("[Voice Changer] generate new embedder. (not match)")
             cls.currentEmbedder = cls.loadEmbedder(embederType, isHalf, dev)
         else:
-            cls.currentEmbedder.setDevice(dev)
+            print("[Voice Changer] generate new embedder. (anyway)")
-            cls.currentEmbedder.setHalf(isHalf)
+            cls.currentEmbedder = cls.loadEmbedder(embederType, isHalf, dev)
-            # print("[Voice Changer] generate new embedder. (ANYWAY)", isHalf)
+
-            # cls.currentEmbedder = cls.loadEmbedder(embederType, file, isHalf, dev)
+            # cls.currentEmbedder.setDevice(dev)
+            # cls.currentEmbedder.setHalf(isHalf)
         return cls.currentEmbedder
 
     @classmethod
@@ -38,13 +40,23 @@ class EmbedderManager:
         cls, embederType: EmbedderType, isHalf: bool, dev: device
     ) -> Embedder:
         if embederType == "hubert_base":
-            file = cls.params.hubert_base
+            try:
-            return FairseqHubert().loadModel(file, dev, isHalf)
+                file = cls.params.content_vec_500_onnx
+                return OnnxContentvec().loadModel(file, dev)
+            except Exception as e:
+                print(e)
+                file = cls.params.hubert_base
+                return FairseqHubert().loadModel(file, dev, isHalf)
         elif embederType == "hubert-base-japanese":
             file = cls.params.hubert_base_jp
             return FairseqHubertJp().loadModel(file, dev, isHalf)
         elif embederType == "contentvec":
-            file = cls.params.hubert_base
+            try:
-            return FairseqContentvec().loadModel(file, dev, isHalf)
+                file = cls.params.content_vec_500_onnx
+                return OnnxContentvec().loadModel(file, dev)
+            except Exception as e:
+                print(e)
+                file = cls.params.hubert_base
+                return FairseqContentvec().loadModel(file, dev, isHalf)
         else:
             return FairseqHubert().loadModel(file, dev, isHalf)

server/voice_changer/RVC/embedder/EmbedderProtocol.py

@@ -0,0 +1,15 @@
+from typing import Protocol
+
+import torch
+from torch import device
+
+
+class EmbedderProtocol(Protocol):
+
+    def loadModel(self, file: str, dev: device, isHalf: bool = True):
+        ...
+
+    def extractFeatures(
+        self, feats: torch.Tensor, embOutputLayer=9, useFinalProj=True
+    ) -> torch.Tensor:
+        ...

server/voice_changer/RVC/embedder/OnnxContentvec.py

@@ -0,0 +1,13 @@
+import torch
+from voice_changer.RVC.embedder.Embedder import Embedder
+
+
+class OnnxContentvec(Embedder):
+
+    def loadModel(self, file: str, dev: torch.device) -> Embedder:
+        raise Exception("Not implemented")
+
+    def extractFeatures(
+        self, feats: torch.Tensor, embOutputLayer=9, useFinalProj=True
+    ) -> torch.Tensor:
+        raise Exception("Not implemented")

server/voice_changer/RVC/pipeline/Pipeline.py

@@ -209,14 +209,13 @@ class Pipeline(object):
         # apply silent front for inference
         if type(self.inferencer) in [OnnxRVCInferencer, OnnxRVCInferencerNono]:
             npyOffset = math.floor(silence_front * 16000) // 360
-            feats = feats[:, npyOffset * 2 :, :]
+            feats = feats[:, npyOffset * 2 :, :]  # NOQA
 
         feats_len = feats.shape[1]
         if pitch is not None and pitchf is not None:
             pitch = pitch[:, -feats_len:]
             pitchf = pitchf[:, -feats_len:]
         p_len = torch.tensor([feats_len], device=self.device).long()
-        
 
         # 推論実行
         try: