mirror of
https://github.com/w-okada/voice-changer.git
synced 2025-01-23 21:45:00 +03:00
bugfix: rvc quality, gui improve: index default to 0.0
This commit is contained in:
wataru
parent
d7e541ffa7
commit
40eccd653f
@ -65,9 +65,7 @@ class RVC:
|
||||
needSwitch: bool = False
|
||||
|
||||
def __init__(self, params: VoiceChangerParams):
|
||||
self.pitchExtractor = PitchExtractorManager.getPitchExtractor(
|
||||
self.settings.f0Detector
|
||||
)
|
||||
self.pitchExtractor = PitchExtractorManager.getPitchExtractor(self.settings.f0Detector)
|
||||
self.params = params
|
||||
EmbedderManager.initialize(params)
|
||||
self.loadSlots()
|
||||
@ -75,9 +73,7 @@ class RVC:
|
||||
|
||||
# サンプルカタログ作成
|
||||
sampleJsons: list[str] = []
|
||||
sampleJsonUrls, _sampleModels = getRVCSampleJsonAndModelIds(
|
||||
params.rvc_sample_mode
|
||||
)
|
||||
sampleJsonUrls, _sampleModels = getRVCSampleJsonAndModelIds(params.rvc_sample_mode)
|
||||
for url in sampleJsonUrls:
|
||||
filename = os.path.basename(url)
|
||||
sampleJsons.append(filename)
|
||||
@ -143,9 +139,7 @@ class RVC:
|
||||
# slotInfo.f0 = sampleInfo.f0
|
||||
else:
|
||||
slotInfo.modelFile = params["files"]["rvcModel"]
|
||||
slotInfo.indexFile = (
|
||||
params["files"]["rvcIndex"] if "rvcIndex" in params["files"] else None
|
||||
)
|
||||
slotInfo.indexFile = params["files"]["rvcIndex"] if "rvcIndex" in params["files"] else None
|
||||
|
||||
slotInfo.defaultTune = params["defaultTune"]
|
||||
slotInfo.defaultIndexRatio = params["defaultIndexRatio"]
|
||||
@ -159,9 +153,7 @@ class RVC:
|
||||
|
||||
# メタデータを見て、永続化モデルフォルダに移動させる
|
||||
# その際に、メタデータのファイル格納場所も書き換える
|
||||
slotDir = os.path.join(
|
||||
self.params.model_dir, RVC_MODEL_DIRNAME, str(target_slot_idx)
|
||||
)
|
||||
slotDir = os.path.join(self.params.model_dir, RVC_MODEL_DIRNAME, str(target_slot_idx))
|
||||
os.makedirs(slotDir, exist_ok=True)
|
||||
slotInfo.modelFile = self.moveToModelDir(slotInfo.modelFile, slotDir)
|
||||
if slotInfo.indexFile is not None and len(slotInfo.indexFile) > 0:
|
||||
@ -189,9 +181,7 @@ class RVC:
|
||||
|
||||
modelSlots: list[ModelSlot] = []
|
||||
for slot_idx in range(len(self.settings.modelSlots)):
|
||||
slotDir = os.path.join(
|
||||
self.params.model_dir, RVC_MODEL_DIRNAME, str(slot_idx)
|
||||
)
|
||||
slotDir = os.path.join(self.params.model_dir, RVC_MODEL_DIRNAME, str(slot_idx))
|
||||
jsonDict = os.path.join(slotDir, "params.json")
|
||||
if os.path.exists(jsonDict):
|
||||
jsonDict = json.load(open(os.path.join(slotDir, "params.json")))
|
||||
@ -209,10 +199,7 @@ class RVC:
|
||||
if val < 0:
|
||||
return True
|
||||
val = val % 1000 # Quick hack for same slot is selected
|
||||
if (
|
||||
self.settings.modelSlots[val].modelFile is None
|
||||
or self.settings.modelSlots[val].modelFile == ""
|
||||
):
|
||||
if self.settings.modelSlots[val].modelFile is None or self.settings.modelSlots[val].modelFile == "":
|
||||
print("[Voice Changer] slot does not have model.")
|
||||
return True
|
||||
self.prepareModel(val)
|
||||
@ -229,9 +216,7 @@ class RVC:
|
||||
elif key in self.settings.strData:
|
||||
setattr(self.settings, key, str(val))
|
||||
if key == "f0Detector" and self.pipeline is not None:
|
||||
pitchExtractor = PitchExtractorManager.getPitchExtractor(
|
||||
self.settings.f0Detector
|
||||
)
|
||||
pitchExtractor = PitchExtractorManager.getPitchExtractor(self.settings.f0Detector)
|
||||
self.pipeline.setPitchExtractor(pitchExtractor)
|
||||
else:
|
||||
return False
|
||||
@ -246,9 +231,7 @@ class RVC:
|
||||
print("[Voice Changer] Prepare Model of slot:", slot)
|
||||
|
||||
# pipelineの生成
|
||||
self.next_pipeline = createPipeline(
|
||||
modelSlot, self.settings.gpu, self.settings.f0Detector
|
||||
)
|
||||
self.next_pipeline = createPipeline(modelSlot, self.settings.gpu, self.settings.f0Detector)
|
||||
|
||||
# その他の設定
|
||||
self.next_trans = modelSlot.defaultTune
|
||||
@ -293,9 +276,7 @@ class RVC:
|
||||
crossfadeSize: int,
|
||||
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 = 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)
|
||||
|
||||
if self.audio_buffer is not None:
|
||||
# 過去のデータに連結
|
||||
@ -303,18 +284,14 @@ class RVC:
|
||||
else:
|
||||
self.audio_buffer = newData
|
||||
|
||||
convertSize = (
|
||||
inputSize + crossfadeSize + solaSearchFrame + self.settings.extraConvertSize
|
||||
)
|
||||
convertSize = inputSize + crossfadeSize + solaSearchFrame + self.settings.extraConvertSize
|
||||
|
||||
if convertSize % 128 != 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 + (128 - (convertSize % 128))
|
||||
|
||||
# バッファがたまっていない場合はzeroで補う
|
||||
if self.audio_buffer.shape[0] < convertSize:
|
||||
self.audio_buffer = np.concatenate(
|
||||
[np.zeros([convertSize]), self.audio_buffer]
|
||||
)
|
||||
self.audio_buffer = np.concatenate([np.zeros([convertSize]), self.audio_buffer])
|
||||
|
||||
convertOffset = -1 * convertSize
|
||||
self.audio_buffer = self.audio_buffer[convertOffset:] # 変換対象の部分だけ抽出
|
||||
@ -324,9 +301,7 @@ class RVC:
|
||||
else:
|
||||
device = torch.device("cpu")
|
||||
|
||||
audio_buffer = torch.from_numpy(self.audio_buffer).to(
|
||||
device=device, dtype=torch.float32
|
||||
)
|
||||
audio_buffer = torch.from_numpy(self.audio_buffer).to(device=device, dtype=torch.float32)
|
||||
|
||||
# 出力部分だけ切り出して音量を確認。(TODO:段階的消音にする)
|
||||
cropOffset = -1 * (inputSize + crossfadeSize)
|
||||
@ -354,7 +329,7 @@ class RVC:
|
||||
# self.needSwitch = False
|
||||
|
||||
# half = self.deviceManager.halfPrecisionAvailable(self.settings.gpu)
|
||||
half = self.pipeline.isHalf
|
||||
# half = self.pipeline.isHalf
|
||||
|
||||
audio = data[0]
|
||||
convertSize = data[1]
|
||||
@ -363,11 +338,9 @@ class RVC:
|
||||
if vol < self.settings.silentThreshold:
|
||||
return np.zeros(convertSize).astype(np.int16)
|
||||
|
||||
audio = torchaudio.functional.resample(
|
||||
audio, self.settings.modelSamplingRate, 16000, rolloff=0.99
|
||||
)
|
||||
repeat = 3 if half else 1
|
||||
repeat *= self.settings.rvcQuality # 0 or 3
|
||||
audio = torchaudio.functional.resample(audio, self.settings.modelSamplingRate, 16000, rolloff=0.99)
|
||||
repeat = 1 if self.settings.rvcQuality else 0
|
||||
print()
|
||||
sid = 0
|
||||
f0_up_key = self.settings.tran
|
||||
index_rate = self.settings.indexRatio
|
||||
@ -382,8 +355,7 @@ class RVC:
|
||||
f0_up_key,
|
||||
index_rate,
|
||||
if_f0,
|
||||
self.settings.extraConvertSize
|
||||
/ self.settings.modelSamplingRate, # extaraDataSizeの秒数。RVCのモデルのサンプリングレートで処理(★１)。
|
||||
self.settings.extraConvertSize / self.settings.modelSamplingRate, # extaraDataSizeの秒数。RVCのモデルのサンプリングレートで処理(★１)。
|
||||
embOutputLayer,
|
||||
useFinalProj,
|
||||
repeat,
|
||||
@ -393,9 +365,7 @@ class RVC:
|
||||
|
||||
return result
|
||||
except DeviceCannotSupportHalfPrecisionException as e:
|
||||
print(
|
||||
"[Device Manager] Device cannot support half precision. Fallback to float...."
|
||||
)
|
||||
print("[Device Manager] Device cannot support half precision. Fallback to float....")
|
||||
self.deviceManager.setForceTensor(True)
|
||||
self.prepareModel(self.settings.modelSlotIndex)
|
||||
raise e
|
||||
@ -461,9 +431,7 @@ class RVC:
|
||||
"sampleId": "",
|
||||
"files": {"rvcModel": storeFile},
|
||||
}
|
||||
props: LoadModelParams = LoadModelParams(
|
||||
slot=targetSlot, isHalf=True, params=params
|
||||
)
|
||||
props: LoadModelParams = LoadModelParams(slot=targetSlot, isHalf=True, params=params)
|
||||
self.loadModel(props)
|
||||
self.prepareModel(targetSlot)
|
||||
self.settings.modelSlotIndex = targetSlot
|
||||
@ -471,12 +439,8 @@ class RVC:
|
||||
|
||||
def update_model_default(self):
|
||||
print("[Voice Changer] UPDATE MODEL DEFAULT!!")
|
||||
slotDir = os.path.join(
|
||||
self.params.model_dir, RVC_MODEL_DIRNAME, str(self.currentSlot)
|
||||
)
|
||||
params = json.load(
|
||||
open(os.path.join(slotDir, "params.json"), "r", encoding="utf-8")
|
||||
)
|
||||
slotDir = os.path.join(self.params.model_dir, RVC_MODEL_DIRNAME, str(self.currentSlot))
|
||||
params = json.load(open(os.path.join(slotDir, "params.json"), "r", encoding="utf-8"))
|
||||
params["defaultTune"] = self.settings.tran
|
||||
params["defaultIndexRatio"] = self.settings.indexRatio
|
||||
params["defaultProtect"] = self.settings.protect
|
||||
@ -488,14 +452,10 @@ class RVC:
|
||||
print("[Voice Changer] UPDATE MODEL INFO", newData)
|
||||
newDataDict = json.loads(newData)
|
||||
try:
|
||||
slotDir = os.path.join(
|
||||
self.params.model_dir, RVC_MODEL_DIRNAME, str(newDataDict["slot"])
|
||||
)
|
||||
slotDir = os.path.join(self.params.model_dir, RVC_MODEL_DIRNAME, str(newDataDict["slot"]))
|
||||
except Exception as e:
|
||||
print("Exception::::", e)
|
||||
params = json.load(
|
||||
open(os.path.join(slotDir, "params.json"), "r", encoding="utf-8")
|
||||
)
|
||||
params = json.load(open(os.path.join(slotDir, "params.json"), "r", encoding="utf-8"))
|
||||
params[newDataDict["key"]] = newDataDict["val"]
|
||||
json.dump(params, open(os.path.join(slotDir, "params.json"), "w"))
|
||||
self.loadSlots()
|
||||
@ -504,9 +464,7 @@ class RVC:
|
||||
print("[Voice Changer] UPLOAD ASSETS", params)
|
||||
paramsDict = json.loads(params)
|
||||
uploadPath = os.path.join(UPLOAD_DIR, paramsDict["file"])
|
||||
storeDir = os.path.join(
|
||||
self.params.model_dir, RVC_MODEL_DIRNAME, str(paramsDict["slot"])
|
||||
)
|
||||
storeDir = os.path.join(self.params.model_dir, RVC_MODEL_DIRNAME, str(paramsDict["slot"]))
|
||||
storePath = os.path.join(
|
||||
storeDir,
|
||||
paramsDict["file"],
|
||||
|
||||
@ -17,9 +17,7 @@ def checkRvcModelExist(model_dir: str):
|
||||
return True
|
||||
|
||||
|
||||
def downloadInitialSampleModels(
|
||||
sampleJsons: list[str], sampleModelIds: list[str], model_dir: str
|
||||
):
|
||||
def downloadInitialSampleModels(sampleJsons: list[str], sampleModelIds: list[str], model_dir: str):
|
||||
sampleModels = getModelSamples(sampleJsons, "RVC")
|
||||
if sampleModels is None:
|
||||
return
|
||||
@ -58,11 +56,7 @@ def downloadInitialSampleModels(
|
||||
slotInfo.modelFile = modelFilePath
|
||||
line_num += 1
|
||||
|
||||
if (
|
||||
initSampleId[1] is True
|
||||
and hasattr(sample, "indexUrl")
|
||||
and sample.indexUrl != ""
|
||||
):
|
||||
if initSampleId[1] is True and hasattr(sample, "indexUrl") and sample.indexUrl != "":
|
||||
indexPath = os.path.join(
|
||||
slotDir,
|
||||
os.path.basename(sample.indexUrl),
|
||||
@ -97,7 +91,7 @@ def downloadInitialSampleModels(
|
||||
slotInfo.name = sample.name
|
||||
slotInfo.termsOfUseUrl = sample.termsOfUseUrl
|
||||
slotInfo.defaultTune = 0
|
||||
slotInfo.defaultIndexRatio = 1
|
||||
slotInfo.defaultIndexRatio = 0.0
|
||||
slotInfo.defaultProtect = 0.5
|
||||
slotInfo.isONNX = slotInfo.modelFile.endswith(".onnx")
|
||||
|
||||
@ -141,11 +135,7 @@ def downloadModelFiles(sampleInfo: RVCModelSample, useIndex: bool = True):
|
||||
)
|
||||
|
||||
indexPath = None
|
||||
if (
|
||||
useIndex is True
|
||||
and hasattr(sampleInfo, "indexUrl")
|
||||
and sampleInfo.indexUrl != ""
|
||||
):
|
||||
if useIndex is True and hasattr(sampleInfo, "indexUrl") and sampleInfo.indexUrl != "":
|
||||
print("[Voice Changer] Download sample with index.")
|
||||
indexPath = os.path.join(TMP_DIR, os.path.basename(sampleInfo.indexUrl))
|
||||
downloadParams.append(
|
||||
|
||||
@ -47,9 +47,7 @@ class Pipeline(object):
|
||||
print("GENERATE PITCH EXTRACTOR", self.pitchExtractor)
|
||||
|
||||
self.index = index
|
||||
self.big_npy = (
|
||||
index.reconstruct_n(0, index.ntotal) if index is not None else None
|
||||
)
|
||||
self.big_npy = index.reconstruct_n(0, index.ntotal) if index is not None else None
|
||||
# self.feature = feature
|
||||
|
||||
self.targetSR = targetSR
|
||||
@ -63,11 +61,7 @@ class Pipeline(object):
|
||||
inferencerInfo = self.inferencer.getInferencerInfo() if self.inferencer else {}
|
||||
embedderInfo = self.embedder.getEmbedderInfo()
|
||||
pitchExtractorInfo = self.pitchExtractor.getPitchExtractorInfo()
|
||||
return {
|
||||
"inferencer": inferencerInfo,
|
||||
"embedder": embedderInfo,
|
||||
"pitchExtractor": pitchExtractorInfo,
|
||||
}
|
||||
return {"inferencer": inferencerInfo, "embedder": embedderInfo, "pitchExtractor": pitchExtractorInfo, "isHalf": self.isHalf}
|
||||
|
||||
def setPitchExtractor(self, pitchExtractor: PitchExtractor):
|
||||
self.pitchExtractor = pitchExtractor
|
||||
@ -87,17 +81,24 @@ class Pipeline(object):
|
||||
):
|
||||
# 16000のサンプリングレートで入ってきている。以降この世界は16000で処理。
|
||||
|
||||
search_index = (
|
||||
self.index is not None and self.big_npy is not None and index_rate != 0
|
||||
)
|
||||
self.t_pad = self.sr * repeat
|
||||
self.t_pad_tgt = self.targetSR * repeat
|
||||
audio_pad = F.pad(
|
||||
audio.unsqueeze(0), (self.t_pad, self.t_pad), mode="reflect"
|
||||
).squeeze(0)
|
||||
search_index = self.index is not None and self.big_npy is not None and index_rate != 0
|
||||
# self.t_pad = self.sr * repeat # 1秒
|
||||
# self.t_pad_tgt = self.targetSR * repeat # 1秒 出力時のトリミング(モデルのサンプリングで出力される)
|
||||
audio = audio.unsqueeze(0)
|
||||
|
||||
quality_padding_sec = (repeat * (audio.shape[1] - 1)) / self.sr # padding(reflect)のサイズは元のサイズより小さい必要がある。
|
||||
|
||||
self.t_pad = round(self.sr * quality_padding_sec) # 前後に音声を追加
|
||||
self.t_pad_tgt = round(self.targetSR * quality_padding_sec) # 前後に音声を追加 出力時のトリミング(モデルのサンプリングで出力される)
|
||||
print("audio shape", self.t_pad, self.t_pad_tgt, audio.shape)
|
||||
audio_pad = F.pad(audio, (self.t_pad, self.t_pad), mode="reflect").squeeze(0)
|
||||
p_len = audio_pad.shape[0] // self.window
|
||||
sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
|
||||
|
||||
# RVC QualityがOnのときにはsilence_frontをオフに。
|
||||
silence_front = silence_front if repeat == 0 else 0
|
||||
print("silence_front", silence_front)
|
||||
|
||||
# ピッチ検出
|
||||
pitch, pitchf = None, None
|
||||
try:
|
||||
@ -112,9 +113,7 @@ class Pipeline(object):
|
||||
pitch = pitch[:p_len]
|
||||
pitchf = pitchf[:p_len]
|
||||
pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
|
||||
pitchf = torch.tensor(
|
||||
pitchf, device=self.device, dtype=torch.float
|
||||
).unsqueeze(0)
|
||||
pitchf = torch.tensor(pitchf, device=self.device, dtype=torch.float).unsqueeze(0)
|
||||
except IndexError:
|
||||
# print(e)
|
||||
raise NotEnoughDataExtimateF0()
|
||||
@ -169,21 +168,14 @@ class Pipeline(object):
|
||||
npy = np.sum(self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
|
||||
|
||||
# recover silient font
|
||||
npy = np.concatenate(
|
||||
[np.zeros([npyOffset, npy.shape[1]]).astype("float32"), npy]
|
||||
)
|
||||
npy = np.concatenate([np.zeros([npyOffset, npy.shape[1]]).astype("float32"), npy])
|
||||
if self.isHalf is True:
|
||||
npy = npy.astype("float16")
|
||||
|
||||
feats = (
|
||||
torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate
|
||||
+ (1 - index_rate) * feats
|
||||
)
|
||||
feats = torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate + (1 - index_rate) * feats
|
||||
feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
|
||||
if protect < 0.5 and search_index:
|
||||
feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(
|
||||
0, 2, 1
|
||||
)
|
||||
feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
|
||||
|
||||
# ピッチサイズ調整
|
||||
p_len = audio_pad.shape[0] // self.window
|
||||
@ -219,14 +211,11 @@ class Pipeline(object):
|
||||
with torch.no_grad():
|
||||
audio1 = (
|
||||
torch.clip(
|
||||
self.inferencer.infer(feats, p_len, pitch, pitchf, sid)[0][
|
||||
0, 0
|
||||
].to(dtype=torch.float32),
|
||||
self.inferencer.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0].to(dtype=torch.float32),
|
||||
-1.0,
|
||||
1.0,
|
||||
)
|
||||
* 32767.5
|
||||
- 0.5
|
||||
).data.to(dtype=torch.int16)
|
||||
except RuntimeError as e:
|
||||
if "HALF" in e.__str__().upper():
|
||||
@ -238,6 +227,8 @@ class Pipeline(object):
|
||||
del feats, p_len, padding_mask
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
# inferで出力されるサンプリングレートはモデルのサンプリングレートになる。
|
||||
# pipelineに（入力されるときはhubertように16k）
|
||||
if self.t_pad_tgt != 0:
|
||||
offset = self.t_pad_tgt
|
||||
end = -1 * self.t_pad_tgt
|
||||
|
||||
Loading…
Reference in New Issue
Block a user