1berry/wokada-voice-changer
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

WIP: so-vits-svc 40v2, alpha (refactoring)

Browse Source
This commit is contained in:
wataru 2023-03-13 01:36:14 +09:00
parent ca4d454d1a
commit ddcae23f20
2 changed files with 28 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
server/voice_changer/SoVitsSvc40v2/SoVitsSvc40v2.py
View File

@ -163,7 +163,7 @@ class SoVitsSvc40v2:
c = c.unsqueeze(0) c = c.unsqueeze(0)
return c, f0, uv return c, f0, uv
def generate_input(self, newData: any, convertSize: int, cropRange): def generate_input(self, newData: any, inputSize: int, crossfadeSize: int):
newData = newData.astype(np.float32) / self.hps.data.max_wav_value newData = newData.astype(np.float32) / self.hps.data.max_wav_value
if hasattr(self, "audio_buffer"): if hasattr(self, "audio_buffer"):
@ -171,13 +171,14 @@ class SoVitsSvc40v2:
else: else:
self.audio_buffer = newData self.audio_buffer = newData
# self.audio_buffer = self.audio_buffer[-(convertSize):] # 変換対象の部分だけ抽出 convertSize = inputSize + crossfadeSize + self.settings.processingLength
# self.audio_buffer = self.audio_buffer[-1024 * 32:] # 変換対象の部分だけ抽出
# self.audio_buffer = self.audio_buffer[-1024 * 128:] # 変換対象の部分だけ抽出
# self.audio_buffer = self.audio_buffer[(-1 * 1024 * 32) + (-1 * convertSize):] # 変換対象の部分だけ抽出
self.audio_buffer = self.audio_buffer[-1 * self.settings.processingLength + (-1 * convertSize):] # 変換対象の部分だけ抽出
crop = self.audio_buffer[cropRange[0]:cropRange[1]] if convertSize % self.hps.data.hop_length != 0: # モデルの出力のホップサイズで切り捨てが発生するので補う。
convertSize = convertSize + (self.hps.data.hop_length - (convertSize % self.hps.data.hop_length))
self.audio_buffer = self.audio_buffer[-1 * convertSize:] # 変換対象の部分だけ抽出
crop = self.audio_buffer[-1 * (inputSize + crossfadeSize):-1 * (crossfadeSize)]
rms = np.sqrt(np.square(crop).mean(axis=0)) rms = np.sqrt(np.square(crop).mean(axis=0))
vol = max(rms, self.prevVol * 0.0) vol = max(rms, self.prevVol * 0.0)

35
server/voice_changer/VoiceChanger.py
View File

@ -159,24 +159,28 @@ class VoiceChanger():
newData = receivedData newData = receivedData
inputSize = newData.shape[0] inputSize = newData.shape[0]
convertSize = inputSize + min(self.settings.crossFadeOverlapSize, inputSize) crossfadeSize = self.settings.crossFadeOverlapSize if self.settings.crossFadeOverlapSize > 0 else inputSize
# convertSize = inputSize + min(self.settings.crossFadeOverlapSize, inputSize)
print_convert_processing( print_convert_processing(
f" Input data size of {receivedData.shape[0]}/{self.settings.inputSampleRate}hz {inputSize}/{processing_sampling_rate}hz") f" Input data size: {receivedData.shape[0]}/{self.settings.inputSampleRate}hz {inputSize}/{processing_sampling_rate}hz")
print_convert_processing(
f" Crossfade data size: crossfade:{crossfadeSize}, crossfade setting:{self.settings.crossFadeOverlapSize}, input size:{inputSize}")
if convertSize < 8192: # if convertSize < 8192:
convertSize = 8192 # convertSize = 8192
if convertSize % processing_hop_length != 0: # モデルの出力のホップサイズで切り捨てが発生するので補う。 # if convertSize % processing_hop_length != 0: # モデルの出力のホップサイズで切り捨てが発生するので補う。
convertSize = convertSize + (processing_hop_length - (convertSize % processing_hop_length)) # convertSize = convertSize + (processing_hop_length - (convertSize % processing_hop_length))
overlapSize = min(self.settings.crossFadeOverlapSize, inputSize) # overlapSize = min(self.settings.crossFadeOverlapSize, inputSize)
cropRange = (-1 * (inputSize + overlapSize), -1 * overlapSize) # cropRange = (-1 * (inputSize + overlapSize), -1 * overlapSize)
print_convert_processing(f" Convert input data size of {convertSize}") print_convert_processing(f" Convert data size of {inputSize + crossfadeSize} (+ extra size)")
print_convert_processing(f" overlap:{overlapSize}, cropRange:{cropRange}") print_convert_processing(f" will be cropped:{-1 * (inputSize + crossfadeSize)}, {-1 * (crossfadeSize)}")
self._generate_strength(inputSize) self._generate_strength(inputSize)
data = self.voiceChanger.generate_input(newData, convertSize, cropRange) data = self.voiceChanger.generate_input(newData, inputSize, crossfadeSize)
preprocess_time = t.secs preprocess_time = t.secs
# 変換処理 # 変換処理
@ -187,8 +191,9 @@ class VoiceChanger():
if hasattr(self, 'np_prev_audio1') == True: if hasattr(self, 'np_prev_audio1') == True:
np.set_printoptions(threshold=10000) np.set_printoptions(threshold=10000)
prev_overlap = self.np_prev_audio1[-1 * overlapSize:] prev_overlap_start = -1 * crossfadeSize
cur_overlap_start = -1 * (inputSize + overlapSize) prev_overlap = self.np_prev_audio1[prev_overlap_start:]
cur_overlap_start = -1 * (inputSize + crossfadeSize)
cur_overlap_end = -1 * inputSize cur_overlap_end = -1 * inputSize
cur_overlap = audio[cur_overlap_start:cur_overlap_end] cur_overlap = audio[cur_overlap_start:cur_overlap_end]
# cur_overlap = audio[-1 * (inputSize + overlapSize):-1 * inputSize] # cur_overlap = audio[-1 * (inputSize + overlapSize):-1 * inputSize]
@ -199,10 +204,10 @@ class VoiceChanger():
powered_cur = cur_overlap * self.np_cur_strength powered_cur = cur_overlap * self.np_cur_strength
powered_result = powered_prev + powered_cur powered_result = powered_prev + powered_cur
cur = audio[-1 * inputSize:-1 * overlapSize] cur = audio[-1 * inputSize:-1 * crossfadeSize]
result = np.concatenate([powered_result, cur], axis=0) result = np.concatenate([powered_result, cur], axis=0)
print_convert_processing( print_convert_processing(
f" overlap:{overlapSize}, current:{cur.shape[0]}, result:{result.shape[0]}... result should be same as input") f" overlap:{crossfadeSize}, current:{cur.shape[0]}, result:{result.shape[0]}... result should be same as input")
if cur.shape[0] != result.shape[0]: if cur.shape[0] != result.shape[0]:
print_convert_processing(f" current and result should be same as input") print_convert_processing(f" current and result should be same as input")