bugfix WIP: server device mode

server/voice_changer/DiffusionSVC/DiffusionSVC.py

@@ -42,7 +42,7 @@ class DiffusionSVC(VoiceChangerModel):
         # その他の設定
         self.settings.tran = self.slotInfo.defaultTune
         self.settings.dstId = self.slotInfo.dstId
-        self.settings.kstep = self.slotInfo.defaultKstep
+        self.settings.kStep = self.slotInfo.defaultKstep
 
         print("[Voice Changer] [DiffusionSVC] Initializing... done")
 
@@ -86,8 +86,8 @@ class DiffusionSVC(VoiceChangerModel):
         solaSearchFrame: int = 0,
     ):
         newData = newData.astype(np.float32) / 32768.0  # DiffusionSVCのモデルのサンプリングレートで入ってきている。（extraDataLength, Crossfade等も同じSRで処理）(★１)
-
-        new_feature_length = newData.shape[0] * 100 // self.slotInfo.samplingRate  # 100 は hubertのhosizeから (16000 / 160)
+        new_feature_length = int(((newData.shape[0] / self.inputSampleRate) * self.slotInfo.samplingRate) / 512)  # 100 は hubertのhosizeから (16000 / 160).
+        # ↑newData.shape[0]//sampleRate でデータ秒数。これに16000かけてhubertの世界でのデータ長。これにhop数(160)でわるとfeatsのデータサイズになる。
         if self.audio_buffer is not None:
             # 過去のデータに連結
             self.audio_buffer = np.concatenate([self.audio_buffer, newData], 0)
@@ -104,13 +104,14 @@ class DiffusionSVC(VoiceChangerModel):
             convertSize = convertSize + (128 - (convertSize % 128))
 
         # バッファがたまっていない場合はzeroで補う
+        generateFeatureLength = int(((convertSize / self.inputSampleRate) * self.slotInfo.samplingRate) / 512) + 1        
         if self.audio_buffer.shape[0] < convertSize:
             self.audio_buffer = np.concatenate([np.zeros([convertSize]), self.audio_buffer])
-            self.pitchf_buffer = np.concatenate([np.zeros([convertSize * 100 // self.slotInfo.samplingRate]), self.pitchf_buffer])
-            self.feature_buffer = np.concatenate([np.zeros([convertSize * 100 // self.slotInfo.samplingRate, self.slotInfo.embChannels]), self.feature_buffer])
+            self.pitchf_buffer = np.concatenate([np.zeros(generateFeatureLength), self.pitchf_buffer])
+            self.feature_buffer = np.concatenate([np.zeros([generateFeatureLength, self.slotInfo.embChannels]), self.feature_buffer])
 
         convertOffset = -1 * convertSize
-        featureOffset = -convertSize * 100 // self.slotInfo.samplingRate
+        featureOffset = -1 * generateFeatureLength
         self.audio_buffer = self.audio_buffer[convertOffset:]  # 変換対象の部分だけ抽出
         self.pitchf_buffer = self.pitchf_buffer[featureOffset:]
         self.feature_buffer = self.feature_buffer[featureOffset:]

server/voice_changer/DiffusionSVC/pipeline/Pipeline.py

@@ -143,8 +143,6 @@ class Pipeline(object):
                     f0_up_key,
                     silence_front=silence_front,
                 )
-#    def extract(self, audio: AudioInOut, sr: int, block_size: int, model_sr: int, pitch, f0_up_key, silence_front=0):
-
                 pitch = torch.tensor(pitch[-n_frames:], device=self.device).unsqueeze(0).long()
             except IndexError as e:  # NOQA
                 raise NotEnoughDataExtimateF0()

server/voice_changer/RVC/RVC.py

@@ -101,7 +101,7 @@ class RVC(VoiceChangerModel):
         solaSearchFrame: int = 0,
     ):
         newData = newData.astype(np.float32) / 32768.0  # RVCのモデルのサンプリングレートで入ってきている。（extraDataLength, Crossfade等も同じSRで処理）(★１)
-
+        # ↑newData.shape[0]//sampleRate でデータ秒数。これに16000かけてhubertの世界でのデータ長。これにhop数(160)でわるとfeatsのデータサイズになる。
         new_feature_length = newData.shape[0] * 100 // self.slotInfo.samplingRate
         if self.audio_buffer is not None:
             # 過去のデータに連結

server/voice_changer/VoiceChanger.py

@@ -84,6 +84,12 @@ class VoiceChanger(VoiceChangerIF):
     def setModel(self, model: Any):
         self.voiceChanger = model
 
+    def setInputSampleRate(self, sr: int):
+        self.settings.inputSampleRate = sr
+
+    def setOutputSampleRate(self, sr: int):
+        self.settings.outputSampleRate = sr
+
     def get_info(self):
         data = asdict(self.settings)
         if self.voiceChanger is not None:

server/voice_changer/VoiceChangerManager.py

@@ -53,10 +53,10 @@ class VoiceChangerManager(ServerDeviceCallbacks):
         return self.voiceChanger.get_processing_sampling_rate()
 
     def setInputSamplingRate(self, sr: int):
-        self.voiceChanger.settings.inputSampleRate = sr
+        self.voiceChanger.setInputSampleRate(sr)
 
     def setOutputSamplingRate(self, sr: int):
-        self.voiceChanger.settings.outputSampleRate = sr
+        self.voiceChanger.setOutputSampleRate(sr)
 
     ############################
     # VoiceChangerManager

server/voice_changer/VoiceChangerV2.py

@@ -95,6 +95,14 @@ class VoiceChangerV2(VoiceChangerIF):
         self.voiceChanger = model
         self.voiceChanger.setSamplingRate(self.settings.inputSampleRate, self.settings.outputSampleRate)
 
+    def setInputSampleRate(self, sr: int):
+        self.settings.inputSampleRate = sr
+        self.voiceChanger.setSamplingRate(self.settings.inputSampleRate, self.settings.outputSampleRate)
+
+    def setOutputSampleRate(self, sr: int):
+        self.settings.outputSampleRate = sr
+        self.voiceChanger.setSamplingRate(self.settings.inputSampleRate, self.settings.outputSampleRate)
+
     def get_info(self):
         data = asdict(self.settings)
         if self.voiceChanger is not None:

server/voice_changer/utils/VoiceChangerIF.py

@@ -25,3 +25,9 @@ class VoiceChangerIF(Protocol):
 
     def export2onnx() -> Any:
         ...
+
+    def setInputSampleRate(self, sr: int):
+        ...
+
+    def setOutputSampleRate(self, sr: int):
+        ...