|
|
|
|
@ -30,32 +30,14 @@ logging.getLogger("multipart.multipart").setLevel(logging.ERROR)
|
|
|
|
|
warnings.simplefilter(action="ignore", category=FutureWarning)
|
|
|
|
|
|
|
|
|
|
version = model_version = os.environ.get("version", "v2")
|
|
|
|
|
path_sovits_v3 = "GPT_SoVITS/pretrained_models/s2Gv3.pth"
|
|
|
|
|
path_sovits_v4 = "GPT_SoVITS/pretrained_models/gsv-v4-pretrained/s2Gv4.pth"
|
|
|
|
|
|
|
|
|
|
from config import name2sovits_path,name2gpt_path,change_choices,get_weights_names
|
|
|
|
|
SoVITS_names, GPT_names = get_weights_names()
|
|
|
|
|
from config import pretrained_sovits_name
|
|
|
|
|
path_sovits_v3 = pretrained_sovits_name["v3"]
|
|
|
|
|
path_sovits_v4 = pretrained_sovits_name["v4"]
|
|
|
|
|
is_exist_s2gv3 = os.path.exists(path_sovits_v3)
|
|
|
|
|
is_exist_s2gv4 = os.path.exists(path_sovits_v4)
|
|
|
|
|
pretrained_sovits_name = [
|
|
|
|
|
"GPT_SoVITS/pretrained_models/s2G488k.pth",
|
|
|
|
|
"GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2G2333k.pth",
|
|
|
|
|
"GPT_SoVITS/pretrained_models/s2Gv3.pth",
|
|
|
|
|
"GPT_SoVITS/pretrained_models/gsv-v4-pretrained/s2Gv4.pth",
|
|
|
|
|
]
|
|
|
|
|
pretrained_gpt_name = [
|
|
|
|
|
"GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt",
|
|
|
|
|
"GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt",
|
|
|
|
|
"GPT_SoVITS/pretrained_models/s1v3.ckpt",
|
|
|
|
|
"GPT_SoVITS/pretrained_models/s1v3.ckpt",
|
|
|
|
|
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
_ = [[], []]
|
|
|
|
|
for i in range(4):
|
|
|
|
|
if os.path.exists(pretrained_gpt_name[i]):
|
|
|
|
|
_[0].append(pretrained_gpt_name[i])
|
|
|
|
|
if os.path.exists(pretrained_sovits_name[i]):
|
|
|
|
|
_[-1].append(pretrained_sovits_name[i])
|
|
|
|
|
pretrained_gpt_name, pretrained_sovits_name = _
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if os.path.exists("./weight.json"):
|
|
|
|
|
pass
|
|
|
|
|
@ -66,28 +48,24 @@ else:
|
|
|
|
|
with open("./weight.json", "r", encoding="utf-8") as file:
|
|
|
|
|
weight_data = file.read()
|
|
|
|
|
weight_data = json.loads(weight_data)
|
|
|
|
|
gpt_path = os.environ.get(
|
|
|
|
|
"gpt_path", weight_data.get("GPT", {}).get(version, pretrained_gpt_name)
|
|
|
|
|
)
|
|
|
|
|
sovits_path = os.environ.get(
|
|
|
|
|
"sovits_path",
|
|
|
|
|
weight_data.get("SoVITS", {}).get(version, pretrained_sovits_name),
|
|
|
|
|
)
|
|
|
|
|
gpt_path = os.environ.get("gpt_path", weight_data.get("GPT", {}).get(version, GPT_names[-1]))
|
|
|
|
|
sovits_path = os.environ.get("sovits_path", weight_data.get("SoVITS", {}).get(version, SoVITS_names[0]))
|
|
|
|
|
if isinstance(gpt_path, list):
|
|
|
|
|
gpt_path = gpt_path[0]
|
|
|
|
|
if isinstance(sovits_path, list):
|
|
|
|
|
sovits_path = sovits_path[0]
|
|
|
|
|
|
|
|
|
|
# gpt_path = os.environ.get(
|
|
|
|
|
# "gpt_path", pretrained_gpt_name
|
|
|
|
|
# )
|
|
|
|
|
# sovits_path = os.environ.get("sovits_path", pretrained_sovits_name)
|
|
|
|
|
cnhubert_base_path = os.environ.get(
|
|
|
|
|
"cnhubert_base_path", "GPT_SoVITS/pretrained_models/chinese-hubert-base"
|
|
|
|
|
)
|
|
|
|
|
bert_path = os.environ.get(
|
|
|
|
|
"bert_path", "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large"
|
|
|
|
|
)
|
|
|
|
|
# print(2333333)
|
|
|
|
|
# print(os.environ["gpt_path"])
|
|
|
|
|
# print(gpt_path)
|
|
|
|
|
# print(GPT_names)
|
|
|
|
|
# print(weight_data)
|
|
|
|
|
# print(weight_data.get("GPT", {}))
|
|
|
|
|
# print(version)###GPT version里没有s2的v2pro
|
|
|
|
|
# print(weight_data.get("GPT", {}).get(version, GPT_names[-1]))
|
|
|
|
|
|
|
|
|
|
cnhubert_base_path = os.environ.get("cnhubert_base_path", "GPT_SoVITS/pretrained_models/chinese-hubert-base")
|
|
|
|
|
bert_path = os.environ.get("bert_path", "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large")
|
|
|
|
|
infer_ttswebui = os.environ.get("infer_ttswebui", 9872)
|
|
|
|
|
infer_ttswebui = int(infer_ttswebui)
|
|
|
|
|
is_share = os.environ.get("is_share", "False")
|
|
|
|
|
@ -231,9 +209,7 @@ def resample(audio_tensor, sr0, sr1):
|
|
|
|
|
global resample_transform_dict
|
|
|
|
|
key = "%s-%s" % (sr0, sr1)
|
|
|
|
|
if key not in resample_transform_dict:
|
|
|
|
|
resample_transform_dict[key] = torchaudio.transforms.Resample(sr0, sr1).to(
|
|
|
|
|
device
|
|
|
|
|
)
|
|
|
|
|
resample_transform_dict[key] = torchaudio.transforms.Resample(sr0, sr1).to(device)
|
|
|
|
|
return resample_transform_dict[key](audio_tensor)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
@ -242,20 +218,15 @@ def resample(audio_tensor, sr0, sr1):
|
|
|
|
|
from process_ckpt import get_sovits_version_from_path_fast, load_sovits_new
|
|
|
|
|
|
|
|
|
|
v3v4set = {"v3", "v4"}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def change_sovits_weights(sovits_path, prompt_language=None, text_language=None):
|
|
|
|
|
if "！"in sovits_path:sovits_path=name2sovits_path[sovits_path]
|
|
|
|
|
global vq_model, hps, version, model_version, dict_language, if_lora_v3
|
|
|
|
|
version, model_version, if_lora_v3 = get_sovits_version_from_path_fast(sovits_path)
|
|
|
|
|
print(sovits_path, version, model_version, if_lora_v3)
|
|
|
|
|
is_exist = is_exist_s2gv3 if model_version == "v3" else is_exist_s2gv4
|
|
|
|
|
path_sovits = path_sovits_v3 if model_version == "v3" else path_sovits_v4
|
|
|
|
|
if if_lora_v3 == True and is_exist == False:
|
|
|
|
|
info = (
|
|
|
|
|
"GPT_SoVITS/pretrained_models/s2Gv3.pth"
|
|
|
|
|
+ f"SoVITS {model_version}"
|
|
|
|
|
+ " : "
|
|
|
|
|
+ i18n("底模缺失，无法加载相应 LoRA 权重")
|
|
|
|
|
)
|
|
|
|
|
info = path_sovits + i18n("SoVITS %s 底模缺失，无法加载相应 LoRA 权重" % model_version)
|
|
|
|
|
gr.Warning(info)
|
|
|
|
|
raise FileExistsError(info)
|
|
|
|
|
dict_language = dict_language_v1 if version == "v1" else dict_language_v2
|
|
|
|
|
@ -269,10 +240,7 @@ def change_sovits_weights(sovits_path, prompt_language=None, text_language=None)
|
|
|
|
|
prompt_text_update = {"__type__": "update", "value": ""}
|
|
|
|
|
prompt_language_update = {"__type__": "update", "value": i18n("中文")}
|
|
|
|
|
if text_language in list(dict_language.keys()):
|
|
|
|
|
text_update, text_language_update = {"__type__": "update"}, {
|
|
|
|
|
"__type__": "update",
|
|
|
|
|
"value": text_language,
|
|
|
|
|
}
|
|
|
|
|
text_update, text_language_update = {"__type__": "update"}, {"__type__": "update", "value": text_language}
|
|
|
|
|
else:
|
|
|
|
|
text_update = {"__type__": "update", "value": ""}
|
|
|
|
|
text_language_update = {"__type__": "update", "value": i18n("中文")}
|
|
|
|
|
@ -293,22 +261,12 @@ def change_sovits_weights(sovits_path, prompt_language=None, text_language=None)
|
|
|
|
|
"__type__": "update",
|
|
|
|
|
"visible": visible_sample_steps,
|
|
|
|
|
"value": 32 if model_version == "v3" else 8,
|
|
|
|
|
"choices": (
|
|
|
|
|
[4, 8, 16, 32, 64, 128] if model_version == "v3" else [4, 8, 16, 32]
|
|
|
|
|
),
|
|
|
|
|
"choices": [4, 8, 16, 32, 64, 128] if model_version == "v3" else [4, 8, 16, 32],
|
|
|
|
|
},
|
|
|
|
|
{"__type__": "update", "visible": visible_inp_refs},
|
|
|
|
|
{
|
|
|
|
|
"__type__": "update",
|
|
|
|
|
"value": False,
|
|
|
|
|
"interactive": True if model_version not in v3v4set else False,
|
|
|
|
|
},
|
|
|
|
|
{"__type__": "update", "value": False, "interactive": True if model_version not in v3v4set else False},
|
|
|
|
|
{"__type__": "update", "visible": True if model_version == "v3" else False},
|
|
|
|
|
{
|
|
|
|
|
"__type__": "update",
|
|
|
|
|
"value": i18n("模型加载中，请等待"),
|
|
|
|
|
"interactive": False,
|
|
|
|
|
},
|
|
|
|
|
{"__type__": "update", "value": i18n("模型加载中，请等待"), "interactive": False},
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
dict_s2 = load_sovits_new(sovits_path)
|
|
|
|
|
@ -324,13 +282,16 @@ def change_sovits_weights(sovits_path, prompt_language=None, text_language=None)
|
|
|
|
|
version = hps.model.version
|
|
|
|
|
# print("sovits版本:",hps.model.version)
|
|
|
|
|
if model_version not in v3v4set:
|
|
|
|
|
if "Pro"not in model_version:
|
|
|
|
|
model_version = version
|
|
|
|
|
else:
|
|
|
|
|
hps.model.version = model_version
|
|
|
|
|
vq_model = SynthesizerTrn(
|
|
|
|
|
hps.data.filter_length // 2 + 1,
|
|
|
|
|
hps.train.segment_size // hps.data.hop_length,
|
|
|
|
|
n_speakers=hps.data.n_speakers,
|
|
|
|
|
**hps.model,
|
|
|
|
|
)
|
|
|
|
|
model_version = version
|
|
|
|
|
else:
|
|
|
|
|
hps.model.version = model_version
|
|
|
|
|
vq_model = SynthesizerTrnV3(
|
|
|
|
|
@ -350,17 +311,12 @@ def change_sovits_weights(sovits_path, prompt_language=None, text_language=None)
|
|
|
|
|
vq_model = vq_model.to(device)
|
|
|
|
|
vq_model.eval()
|
|
|
|
|
if if_lora_v3 == False:
|
|
|
|
|
print(
|
|
|
|
|
"loading sovits_%s" % model_version,
|
|
|
|
|
vq_model.load_state_dict(dict_s2["weight"], strict=False),
|
|
|
|
|
)
|
|
|
|
|
print("loading sovits_%s" % model_version, vq_model.load_state_dict(dict_s2["weight"], strict=False))
|
|
|
|
|
else:
|
|
|
|
|
path_sovits = path_sovits_v3 if model_version == "v3" else path_sovits_v4
|
|
|
|
|
print(
|
|
|
|
|
"loading sovits_%spretrained_G" % model_version,
|
|
|
|
|
vq_model.load_state_dict(
|
|
|
|
|
load_sovits_new(path_sovits)["weight"], strict=False
|
|
|
|
|
),
|
|
|
|
|
vq_model.load_state_dict(load_sovits_new(path_sovits)["weight"], strict=False),
|
|
|
|
|
)
|
|
|
|
|
lora_rank = dict_s2["lora_rank"]
|
|
|
|
|
lora_config = LoraConfig(
|
|
|
|
|
@ -387,16 +343,10 @@ def change_sovits_weights(sovits_path, prompt_language=None, text_language=None)
|
|
|
|
|
"__type__": "update",
|
|
|
|
|
"visible": visible_sample_steps,
|
|
|
|
|
"value": 32 if model_version == "v3" else 8,
|
|
|
|
|
"choices": (
|
|
|
|
|
[4, 8, 16, 32, 64, 128] if model_version == "v3" else [4, 8, 16, 32]
|
|
|
|
|
),
|
|
|
|
|
"choices": [4, 8, 16, 32, 64, 128] if model_version == "v3" else [4, 8, 16, 32],
|
|
|
|
|
},
|
|
|
|
|
{"__type__": "update", "visible": visible_inp_refs},
|
|
|
|
|
{
|
|
|
|
|
"__type__": "update",
|
|
|
|
|
"value": False,
|
|
|
|
|
"interactive": True if model_version not in v3v4set else False,
|
|
|
|
|
},
|
|
|
|
|
{"__type__": "update", "value": False, "interactive": True if model_version not in v3v4set else False},
|
|
|
|
|
{"__type__": "update", "visible": True if model_version == "v3" else False},
|
|
|
|
|
{"__type__": "update", "value": i18n("合成语音"), "interactive": True},
|
|
|
|
|
)
|
|
|
|
|
@ -415,9 +365,10 @@ except:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def change_gpt_weights(gpt_path):
|
|
|
|
|
if "！"in gpt_path:gpt_path=name2gpt_path[gpt_path]
|
|
|
|
|
global hz, max_sec, t2s_model, config
|
|
|
|
|
hz = 50
|
|
|
|
|
dict_s1 = torch.load(gpt_path, map_location="cpu")
|
|
|
|
|
dict_s1 = torch.load(gpt_path, map_location="cpu", weights_only=False)
|
|
|
|
|
config = dict_s1["config"]
|
|
|
|
|
max_sec = config["data"]["max_sec"]
|
|
|
|
|
t2s_model = Text2SemanticLightningModule(config, "****", is_train=False)
|
|
|
|
|
@ -442,34 +393,54 @@ import torch
|
|
|
|
|
|
|
|
|
|
now_dir = os.getcwd()
|
|
|
|
|
|
|
|
|
|
def clean_hifigan_model():
|
|
|
|
|
global hifigan_model
|
|
|
|
|
if hifigan_model:
|
|
|
|
|
hifigan_model = hifigan_model.cpu()
|
|
|
|
|
hifigan_model = None
|
|
|
|
|
try:
|
|
|
|
|
torch.cuda.empty_cache()
|
|
|
|
|
except:
|
|
|
|
|
pass
|
|
|
|
|
def clean_bigvgan_model():
|
|
|
|
|
global bigvgan_model
|
|
|
|
|
if bigvgan_model:
|
|
|
|
|
bigvgan_model = bigvgan_model.cpu()
|
|
|
|
|
bigvgan_model = None
|
|
|
|
|
try:
|
|
|
|
|
torch.cuda.empty_cache()
|
|
|
|
|
except:
|
|
|
|
|
pass
|
|
|
|
|
def clean_sv_cn_model():
|
|
|
|
|
global sv_cn_model
|
|
|
|
|
if sv_cn_model:
|
|
|
|
|
sv_cn_model.embedding_model = sv_cn_model.embedding_model.cpu()
|
|
|
|
|
sv_cn_model = None
|
|
|
|
|
try:
|
|
|
|
|
torch.cuda.empty_cache()
|
|
|
|
|
except:
|
|
|
|
|
pass
|
|
|
|
|
|
|
|
|
|
def init_bigvgan():
|
|
|
|
|
global bigvgan_model, hifigan_model
|
|
|
|
|
global bigvgan_model, hifigan_model,sv_cn_model
|
|
|
|
|
from BigVGAN import bigvgan
|
|
|
|
|
|
|
|
|
|
bigvgan_model = bigvgan.BigVGAN.from_pretrained(
|
|
|
|
|
"%s/GPT_SoVITS/pretrained_models/models--nvidia--bigvgan_v2_24khz_100band_256x"
|
|
|
|
|
% (now_dir,),
|
|
|
|
|
"%s/GPT_SoVITS/pretrained_models/models--nvidia--bigvgan_v2_24khz_100band_256x" % (now_dir,),
|
|
|
|
|
use_cuda_kernel=False,
|
|
|
|
|
) # if True, RuntimeError: Ninja is required to load C++ extensions
|
|
|
|
|
# remove weight norm in the model and set to eval mode
|
|
|
|
|
bigvgan_model.remove_weight_norm()
|
|
|
|
|
bigvgan_model = bigvgan_model.eval()
|
|
|
|
|
if hifigan_model:
|
|
|
|
|
hifigan_model = hifigan_model.cpu()
|
|
|
|
|
hifigan_model = None
|
|
|
|
|
try:
|
|
|
|
|
torch.cuda.empty_cache()
|
|
|
|
|
except:
|
|
|
|
|
pass
|
|
|
|
|
clean_hifigan_model()
|
|
|
|
|
clean_sv_cn_model()
|
|
|
|
|
if is_half == True:
|
|
|
|
|
bigvgan_model = bigvgan_model.half().to(device)
|
|
|
|
|
else:
|
|
|
|
|
bigvgan_model = bigvgan_model.to(device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def init_hifigan():
|
|
|
|
|
global hifigan_model, bigvgan_model
|
|
|
|
|
global hifigan_model, bigvgan_model,sv_cn_model
|
|
|
|
|
hifigan_model = Generator(
|
|
|
|
|
initial_channel=100,
|
|
|
|
|
resblock="1",
|
|
|
|
|
@ -484,48 +455,73 @@ def init_hifigan():
|
|
|
|
|
hifigan_model.eval()
|
|
|
|
|
hifigan_model.remove_weight_norm()
|
|
|
|
|
state_dict_g = torch.load(
|
|
|
|
|
"%s/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth" % (now_dir,),
|
|
|
|
|
map_location="cpu",
|
|
|
|
|
"%s/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth" % (now_dir,), map_location="cpu", weights_only=False
|
|
|
|
|
)
|
|
|
|
|
print("loading vocoder", hifigan_model.load_state_dict(state_dict_g))
|
|
|
|
|
if bigvgan_model:
|
|
|
|
|
bigvgan_model = bigvgan_model.cpu()
|
|
|
|
|
bigvgan_model = None
|
|
|
|
|
try:
|
|
|
|
|
torch.cuda.empty_cache()
|
|
|
|
|
except:
|
|
|
|
|
pass
|
|
|
|
|
clean_bigvgan_model()
|
|
|
|
|
clean_sv_cn_model()
|
|
|
|
|
if is_half == True:
|
|
|
|
|
hifigan_model = hifigan_model.half().to(device)
|
|
|
|
|
else:
|
|
|
|
|
hifigan_model = hifigan_model.to(device)
|
|
|
|
|
|
|
|
|
|
from sv import SV
|
|
|
|
|
def init_sv_cn():
|
|
|
|
|
global hifigan_model, bigvgan_model,sv_cn_model
|
|
|
|
|
sv_cn_model = SV(device, is_half)
|
|
|
|
|
clean_bigvgan_model()
|
|
|
|
|
clean_hifigan_model()
|
|
|
|
|
|
|
|
|
|
bigvgan_model = hifigan_model = None
|
|
|
|
|
|
|
|
|
|
bigvgan_model = hifigan_model = sv_cn_model = None
|
|
|
|
|
if model_version == "v3":
|
|
|
|
|
init_bigvgan()
|
|
|
|
|
if model_version == "v4":
|
|
|
|
|
init_hifigan()
|
|
|
|
|
if model_version in {"v2Pro","v2ProPlus"}:
|
|
|
|
|
init_sv_cn()
|
|
|
|
|
|
|
|
|
|
resample_transform_dict={}
|
|
|
|
|
def resample(audio_tensor, sr0,sr1,device):
|
|
|
|
|
global resample_transform_dict
|
|
|
|
|
key="%s-%s-%s"%(sr0,sr1,str(device))
|
|
|
|
|
if key not in resample_transform_dict:
|
|
|
|
|
resample_transform_dict[key] = torchaudio.transforms.Resample(
|
|
|
|
|
sr0, sr1
|
|
|
|
|
).to(device)
|
|
|
|
|
return resample_transform_dict[key](audio_tensor)
|
|
|
|
|
|
|
|
|
|
def get_spepc(hps, filename):
|
|
|
|
|
def get_spepc(hps, filename,dtype,device,is_v2pro=False):
|
|
|
|
|
# audio = load_audio(filename, int(hps.data.sampling_rate))
|
|
|
|
|
audio, sampling_rate = librosa.load(filename, sr=int(hps.data.sampling_rate))
|
|
|
|
|
audio = torch.FloatTensor(audio)
|
|
|
|
|
|
|
|
|
|
# audio, sampling_rate = librosa.load(filename, sr=int(hps.data.sampling_rate))
|
|
|
|
|
# audio = torch.FloatTensor(audio)
|
|
|
|
|
|
|
|
|
|
sr1=int(hps.data.sampling_rate)
|
|
|
|
|
audio, sr0=torchaudio.load(filename)
|
|
|
|
|
if sr0!=sr1:
|
|
|
|
|
audio=audio.to(device)
|
|
|
|
|
if(audio.shape[0]==2):audio=audio.mean(0).unsqueeze(0)
|
|
|
|
|
audio=resample(audio,sr0,sr1,device)
|
|
|
|
|
else:
|
|
|
|
|
audio=audio.to(device)
|
|
|
|
|
if(audio.shape[0]==2):audio=audio.mean(0).unsqueeze(0)
|
|
|
|
|
|
|
|
|
|
maxx = audio.abs().max()
|
|
|
|
|
if maxx > 1:
|
|
|
|
|
audio /= min(2, maxx)
|
|
|
|
|
audio_norm = audio
|
|
|
|
|
audio_norm = audio_norm.unsqueeze(0)
|
|
|
|
|
spec = spectrogram_torch(
|
|
|
|
|
audio_norm,
|
|
|
|
|
audio,
|
|
|
|
|
hps.data.filter_length,
|
|
|
|
|
hps.data.sampling_rate,
|
|
|
|
|
hps.data.hop_length,
|
|
|
|
|
hps.data.win_length,
|
|
|
|
|
center=False,
|
|
|
|
|
)
|
|
|
|
|
return spec
|
|
|
|
|
spec=spec.to(dtype)
|
|
|
|
|
if is_v2pro==True:
|
|
|
|
|
audio=resample(audio,sr1,16000,device).to(dtype)
|
|
|
|
|
return spec,audio
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def clean_text_inf(text, language, version):
|
|
|
|
|
@ -588,9 +584,7 @@ def get_phones_and_bert(text, language, version, final=False):
|
|
|
|
|
formattext = chinese.mix_text_normalize(formattext)
|
|
|
|
|
return get_phones_and_bert(formattext, "zh", version)
|
|
|
|
|
else:
|
|
|
|
|
phones, word2ph, norm_text = clean_text_inf(
|
|
|
|
|
formattext, language, version
|
|
|
|
|
)
|
|
|
|
|
phones, word2ph, norm_text = clean_text_inf(formattext, language, version)
|
|
|
|
|
bert = get_bert_feature(norm_text, word2ph).to(device)
|
|
|
|
|
elif language == "all_yue" and re.search(r"[A-Za-z]", formattext):
|
|
|
|
|
formattext = re.sub(r"[a-z]", lambda x: x.group(0).upper(), formattext)
|
|
|
|
|
@ -716,11 +710,7 @@ def audio_sr(audio, sr):
|
|
|
|
|
try:
|
|
|
|
|
sr_model = AP_BWE(device, DictToAttrRecursive)
|
|
|
|
|
except FileNotFoundError:
|
|
|
|
|
gr.Warning(
|
|
|
|
|
i18n(
|
|
|
|
|
"你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好"
|
|
|
|
|
)
|
|
|
|
|
)
|
|
|
|
|
gr.Warning(i18n("你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好"))
|
|
|
|
|
return audio.cpu().detach().numpy(), sr
|
|
|
|
|
return sr_model(audio, sr)
|
|
|
|
|
|
|
|
|
|
@ -764,6 +754,10 @@ def get_tts_wav(
|
|
|
|
|
ref_free = False # s2v3暂不支持ref_free
|
|
|
|
|
else:
|
|
|
|
|
if_sr = False
|
|
|
|
|
if model_version not in {"v3","v4","v2Pro","v2ProPlus"}:
|
|
|
|
|
clean_bigvgan_model()
|
|
|
|
|
clean_hifigan_model()
|
|
|
|
|
clean_sv_cn_model()
|
|
|
|
|
t0 = ttime()
|
|
|
|
|
prompt_language = dict_language[prompt_language]
|
|
|
|
|
text_language = dict_language[text_language]
|
|
|
|
|
@ -798,11 +792,7 @@ def get_tts_wav(
|
|
|
|
|
else:
|
|
|
|
|
wav16k = wav16k.to(device)
|
|
|
|
|
wav16k = torch.cat([wav16k, zero_wav_torch])
|
|
|
|
|
ssl_content = ssl_model.model(wav16k.unsqueeze(0))[
|
|
|
|
|
"last_hidden_state"
|
|
|
|
|
].transpose(
|
|
|
|
|
1, 2
|
|
|
|
|
) # .float()
|
|
|
|
|
ssl_content = ssl_model.model(wav16k.unsqueeze(0))["last_hidden_state"].transpose(1, 2) # .float()
|
|
|
|
|
codes = vq_model.extract_latent(ssl_content)
|
|
|
|
|
prompt_semantic = codes[0, 0]
|
|
|
|
|
prompt = prompt_semantic.unsqueeze(0).to(device)
|
|
|
|
|
@ -829,9 +819,7 @@ def get_tts_wav(
|
|
|
|
|
audio_opt = []
|
|
|
|
|
###s2v3暂不支持ref_free
|
|
|
|
|
if not ref_free:
|
|
|
|
|
phones1, bert1, norm_text1 = get_phones_and_bert(
|
|
|
|
|
prompt_text, prompt_language, version
|
|
|
|
|
)
|
|
|
|
|
phones1, bert1, norm_text1 = get_phones_and_bert(prompt_text, prompt_language, version)
|
|
|
|
|
|
|
|
|
|
for i_text, text in enumerate(texts):
|
|
|
|
|
# 解决输入目标文本的空行导致报错的问题
|
|
|
|
|
@ -844,9 +832,7 @@ def get_tts_wav(
|
|
|
|
|
print(i18n("前端处理后的文本(每句):"), norm_text2)
|
|
|
|
|
if not ref_free:
|
|
|
|
|
bert = torch.cat([bert1, bert2], 1)
|
|
|
|
|
all_phoneme_ids = (
|
|
|
|
|
torch.LongTensor(phones1 + phones2).to(device).unsqueeze(0)
|
|
|
|
|
)
|
|
|
|
|
all_phoneme_ids = torch.LongTensor(phones1 + phones2).to(device).unsqueeze(0)
|
|
|
|
|
else:
|
|
|
|
|
bert = bert2
|
|
|
|
|
all_phoneme_ids = torch.LongTensor(phones2).to(device).unsqueeze(0)
|
|
|
|
|
@ -875,31 +861,37 @@ def get_tts_wav(
|
|
|
|
|
pred_semantic = pred_semantic[:, -idx:].unsqueeze(0)
|
|
|
|
|
cache[i_text] = pred_semantic
|
|
|
|
|
t3 = ttime()
|
|
|
|
|
is_v2pro=model_version in {"v2Pro","v2ProPlus"}
|
|
|
|
|
# print(23333,is_v2pro,model_version)
|
|
|
|
|
###v3不存在以下逻辑和inp_refs
|
|
|
|
|
if model_version not in v3v4set:
|
|
|
|
|
refers = []
|
|
|
|
|
if is_v2pro:
|
|
|
|
|
sv_emb=[]
|
|
|
|
|
if sv_cn_model == None:
|
|
|
|
|
init_sv_cn()
|
|
|
|
|
if inp_refs:
|
|
|
|
|
for path in inp_refs:
|
|
|
|
|
try:
|
|
|
|
|
refer = get_spepc(hps, path.name).to(dtype).to(device)
|
|
|
|
|
try:#####这里加上提取sv的逻辑，要么一堆sv一堆refer，要么单个sv单个refer
|
|
|
|
|
refer,audio_tensor = get_spepc(hps, path.name,dtype,device,is_v2pro)
|
|
|
|
|
refers.append(refer)
|
|
|
|
|
if is_v2pro:
|
|
|
|
|
sv_emb.append(sv_cn_model.compute_embedding3(audio_tensor))
|
|
|
|
|
except:
|
|
|
|
|
traceback.print_exc()
|
|
|
|
|
if len(refers) == 0:
|
|
|
|
|
refers = [get_spepc(hps, ref_wav_path).to(dtype).to(device)]
|
|
|
|
|
audio = vq_model.decode(
|
|
|
|
|
pred_semantic,
|
|
|
|
|
torch.LongTensor(phones2).to(device).unsqueeze(0),
|
|
|
|
|
refers,
|
|
|
|
|
speed=speed,
|
|
|
|
|
)[0][
|
|
|
|
|
0
|
|
|
|
|
] # .cpu().detach().numpy()
|
|
|
|
|
refers,audio_tensor = get_spepc(hps, ref_wav_path,dtype,device,is_v2pro)
|
|
|
|
|
refers=[refers]
|
|
|
|
|
if is_v2pro:
|
|
|
|
|
sv_emb=[sv_cn_model.compute_embedding3(audio_tensor)]
|
|
|
|
|
if is_v2pro:
|
|
|
|
|
audio = vq_model.decode(pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refers, speed=speed,sv_emb=sv_emb)[0][0]
|
|
|
|
|
else:
|
|
|
|
|
audio = vq_model.decode(pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refers, speed=speed)[0][0]
|
|
|
|
|
else:
|
|
|
|
|
refer = get_spepc(hps, ref_wav_path).to(device).to(dtype)
|
|
|
|
|
refer,audio_tensor = get_spepc(hps, ref_wav_path,dtype,device)
|
|
|
|
|
phoneme_ids0 = torch.LongTensor(phones1).to(device).unsqueeze(0)
|
|
|
|
|
phoneme_ids1 = torch.LongTensor(phones2).to(device).unsqueeze(0)
|
|
|
|
|
# print(11111111, phoneme_ids0, phoneme_ids1)
|
|
|
|
|
fea_ref, ge = vq_model.decode_encp(prompt.unsqueeze(0), phoneme_ids0, refer)
|
|
|
|
|
ref_audio, sr = torchaudio.load(ref_wav_path)
|
|
|
|
|
ref_audio = ref_audio.to(device).float()
|
|
|
|
|
@ -922,9 +914,7 @@ def get_tts_wav(
|
|
|
|
|
T_min = Tref
|
|
|
|
|
chunk_len = Tchunk - T_min
|
|
|
|
|
mel2 = mel2.to(dtype)
|
|
|
|
|
fea_todo, ge = vq_model.decode_encp(
|
|
|
|
|
pred_semantic, phoneme_ids1, refer, ge, speed
|
|
|
|
|
)
|
|
|
|
|
fea_todo, ge = vq_model.decode_encp(pred_semantic, phoneme_ids1, refer, ge, speed)
|
|
|
|
|
cfm_resss = []
|
|
|
|
|
idx = 0
|
|
|
|
|
while 1:
|
|
|
|
|
@ -934,11 +924,7 @@ def get_tts_wav(
|
|
|
|
|
idx += chunk_len
|
|
|
|
|
fea = torch.cat([fea_ref, fea_todo_chunk], 2).transpose(2, 1)
|
|
|
|
|
cfm_res = vq_model.cfm.inference(
|
|
|
|
|
fea,
|
|
|
|
|
torch.LongTensor([fea.size(1)]).to(fea.device),
|
|
|
|
|
mel2,
|
|
|
|
|
sample_steps,
|
|
|
|
|
inference_cfg_rate=0,
|
|
|
|
|
fea, torch.LongTensor([fea.size(1)]).to(fea.device), mel2, sample_steps, inference_cfg_rate=0
|
|
|
|
|
)
|
|
|
|
|
cfm_res = cfm_res[:, :, mel2.shape[2] :]
|
|
|
|
|
mel2 = cfm_res[:, :, -T_min:]
|
|
|
|
|
@ -966,7 +952,7 @@ def get_tts_wav(
|
|
|
|
|
t1 = ttime()
|
|
|
|
|
print("%.3f\t%.3f\t%.3f\t%.3f" % (t[0], sum(t[1::3]), sum(t[2::3]), sum(t[3::3])))
|
|
|
|
|
audio_opt = torch.cat(audio_opt, 0) # np.concatenate
|
|
|
|
|
if model_version in {"v1", "v2"}:
|
|
|
|
|
if model_version in {"v1", "v2", "v2Pro", "v2ProPlus"}:
|
|
|
|
|
opt_sr = 32000
|
|
|
|
|
elif model_version == "v3":
|
|
|
|
|
opt_sr = 24000
|
|
|
|
|
@ -1065,13 +1051,7 @@ def cut5(inp):
|
|
|
|
|
|
|
|
|
|
for i, char in enumerate(inp):
|
|
|
|
|
if char in punds:
|
|
|
|
|
if (
|
|
|
|
|
char == "."
|
|
|
|
|
and i > 0
|
|
|
|
|
and i < len(inp) - 1
|
|
|
|
|
and inp[i - 1].isdigit()
|
|
|
|
|
and inp[i + 1].isdigit()
|
|
|
|
|
):
|
|
|
|
|
if char == "." and i > 0 and i < len(inp) - 1 and inp[i - 1].isdigit() and inp[i + 1].isdigit():
|
|
|
|
|
items.append(char)
|
|
|
|
|
else:
|
|
|
|
|
items.append(char)
|
|
|
|
|
@ -1106,46 +1086,6 @@ def process_text(texts):
|
|
|
|
|
_text.append(text)
|
|
|
|
|
return _text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def change_choices():
|
|
|
|
|
SoVITS_names, GPT_names = get_weights_names(GPT_weight_root, SoVITS_weight_root)
|
|
|
|
|
return {
|
|
|
|
|
"choices": sorted(SoVITS_names, key=custom_sort_key),
|
|
|
|
|
"__type__": "update",
|
|
|
|
|
}, {
|
|
|
|
|
"choices": sorted(GPT_names, key=custom_sort_key),
|
|
|
|
|
"__type__": "update",
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
SoVITS_weight_root = [
|
|
|
|
|
"SoVITS_weights",
|
|
|
|
|
"SoVITS_weights_v2",
|
|
|
|
|
"SoVITS_weights_v3",
|
|
|
|
|
"SoVITS_weights_v4",
|
|
|
|
|
]
|
|
|
|
|
GPT_weight_root = ["GPT_weights", "GPT_weights_v2", "GPT_weights_v3", "GPT_weights_v4"]
|
|
|
|
|
for path in SoVITS_weight_root + GPT_weight_root:
|
|
|
|
|
os.makedirs(path, exist_ok=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def get_weights_names(GPT_weight_root, SoVITS_weight_root):
|
|
|
|
|
SoVITS_names = [i for i in pretrained_sovits_name]
|
|
|
|
|
for path in SoVITS_weight_root:
|
|
|
|
|
for name in os.listdir(path):
|
|
|
|
|
if name.endswith(".pth"):
|
|
|
|
|
SoVITS_names.append("%s/%s" % (path, name))
|
|
|
|
|
GPT_names = [i for i in pretrained_gpt_name]
|
|
|
|
|
for path in GPT_weight_root:
|
|
|
|
|
for name in os.listdir(path):
|
|
|
|
|
if name.endswith(".ckpt"):
|
|
|
|
|
GPT_names.append("%s/%s" % (path, name))
|
|
|
|
|
return SoVITS_names, GPT_names
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
SoVITS_names, GPT_names = get_weights_names(GPT_weight_root, SoVITS_weight_root)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def html_center(text, label="p"):
|
|
|
|
|
return f"""<div style="text-align: center; margin: 100; padding: 50;">
|
|
|
|
|
<{label} style="margin: 0; padding: 0;">{text}</{label}>
|
|
|
|
|
@ -1160,13 +1100,9 @@ def html_left(text, label="p"):
|
|
|
|
|
|
|
|
|
|
with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False) as app:
|
|
|
|
|
gr.Markdown(
|
|
|
|
|
value=i18n(
|
|
|
|
|
"本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责."
|
|
|
|
|
)
|
|
|
|
|
value=i18n("本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责.")
|
|
|
|
|
+ "<br>"
|
|
|
|
|
+ i18n(
|
|
|
|
|
"如不认可该条款, 则不能使用或引用软件包内任何代码和文件. 详见根目录LICENSE."
|
|
|
|
|
)
|
|
|
|
|
+ i18n("如不认可该条款, 则不能使用或引用软件包内任何代码和文件. 详见根目录LICENSE.")
|
|
|
|
|
)
|
|
|
|
|
with gr.Group():
|
|
|
|
|
gr.Markdown(html_center(i18n("模型切换"), "h3"))
|
|
|
|
|
@ -1185,19 +1121,11 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False) as app:
|
|
|
|
|
interactive=True,
|
|
|
|
|
scale=14,
|
|
|
|
|
)
|
|
|
|
|
refresh_button = gr.Button(
|
|
|
|
|
i18n("刷新模型路径"), variant="primary", scale=14
|
|
|
|
|
)
|
|
|
|
|
refresh_button.click(
|
|
|
|
|
fn=change_choices, inputs=[], outputs=[SoVITS_dropdown, GPT_dropdown]
|
|
|
|
|
)
|
|
|
|
|
refresh_button = gr.Button(i18n("刷新模型路径"), variant="primary", scale=14)
|
|
|
|
|
refresh_button.click(fn=change_choices, inputs=[], outputs=[SoVITS_dropdown, GPT_dropdown])
|
|
|
|
|
gr.Markdown(html_center(i18n("*请上传并填写参考信息"), "h3"))
|
|
|
|
|
with gr.Row():
|
|
|
|
|
inp_ref = gr.Audio(
|
|
|
|
|
label=i18n("请上传3~10秒内参考音频，超过会报错！"),
|
|
|
|
|
type="filepath",
|
|
|
|
|
scale=13,
|
|
|
|
|
)
|
|
|
|
|
inp_ref = gr.Audio(label=i18n("请上传3~10秒内参考音频，超过会报错！"), type="filepath", scale=13)
|
|
|
|
|
with gr.Column(scale=13):
|
|
|
|
|
ref_text_free = gr.Checkbox(
|
|
|
|
|
label=i18n("开启无参考文本模式。不填参考文本亦相当于开启。")
|
|
|
|
|
@ -1211,18 +1139,10 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False) as app:
|
|
|
|
|
html_left(
|
|
|
|
|
i18n("使用无参考文本模式时建议使用微调的GPT")
|
|
|
|
|
+ "<br>"
|
|
|
|
|
+ i18n(
|
|
|
|
|
"听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。"
|
|
|
|
|
)
|
|
|
|
|
+ i18n("听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。")
|
|
|
|
|
)
|
|
|
|
|
)
|
|
|
|
|
prompt_text = gr.Textbox(
|
|
|
|
|
label=i18n("参考音频的文本"),
|
|
|
|
|
value="",
|
|
|
|
|
lines=5,
|
|
|
|
|
max_lines=5,
|
|
|
|
|
scale=1,
|
|
|
|
|
)
|
|
|
|
|
prompt_text = gr.Textbox(label=i18n("参考音频的文本"), value="", lines=5, max_lines=5, scale=1)
|
|
|
|
|
with gr.Column(scale=14):
|
|
|
|
|
prompt_language = gr.Dropdown(
|
|
|
|
|
label=i18n("参考音频的语种"),
|
|
|
|
|
@ -1249,21 +1169,13 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False) as app:
|
|
|
|
|
gr.Radio(
|
|
|
|
|
label=i18n("采样步数,如果觉得电,提高试试,如果觉得慢,降低试试"),
|
|
|
|
|
value=32 if model_version == "v3" else 8,
|
|
|
|
|
choices=(
|
|
|
|
|
[4, 8, 16, 32, 64, 128]
|
|
|
|
|
if model_version == "v3"
|
|
|
|
|
else [4, 8, 16, 32]
|
|
|
|
|
),
|
|
|
|
|
choices=[4, 8, 16, 32, 64, 128] if model_version == "v3" else [4, 8, 16, 32],
|
|
|
|
|
visible=True,
|
|
|
|
|
)
|
|
|
|
|
if model_version in v3v4set
|
|
|
|
|
else gr.Radio(
|
|
|
|
|
label=i18n("采样步数,如果觉得电,提高试试,如果觉得慢,降低试试"),
|
|
|
|
|
choices=(
|
|
|
|
|
[4, 8, 16, 32, 64, 128]
|
|
|
|
|
if model_version == "v3"
|
|
|
|
|
else [4, 8, 16, 32]
|
|
|
|
|
),
|
|
|
|
|
choices=[4, 8, 16, 32, 64, 128] if model_version == "v3" else [4, 8, 16, 32],
|
|
|
|
|
visible=False,
|
|
|
|
|
value=32 if model_version == "v3" else 8,
|
|
|
|
|
)
|
|
|
|
|
@ -1278,9 +1190,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False) as app:
|
|
|
|
|
gr.Markdown(html_center(i18n("*请填写需要合成的目标文本和语种模式"), "h3"))
|
|
|
|
|
with gr.Row():
|
|
|
|
|
with gr.Column(scale=13):
|
|
|
|
|
text = gr.Textbox(
|
|
|
|
|
label=i18n("需要合成的文本"), value="", lines=26, max_lines=26
|
|
|
|
|
)
|
|
|
|
|
text = gr.Textbox(label=i18n("需要合成的文本"), value="", lines=26, max_lines=26)
|
|
|
|
|
with gr.Column(scale=7):
|
|
|
|
|
text_language = gr.Dropdown(
|
|
|
|
|
label=i18n("需要合成的语种") + i18n(".限制范围越小判别效果越好。"),
|
|
|
|
|
@ -1312,13 +1222,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False) as app:
|
|
|
|
|
)
|
|
|
|
|
with gr.Row():
|
|
|
|
|
speed = gr.Slider(
|
|
|
|
|
minimum=0.6,
|
|
|
|
|
maximum=1.65,
|
|
|
|
|
step=0.05,
|
|
|
|
|
label=i18n("语速"),
|
|
|
|
|
value=1,
|
|
|
|
|
interactive=True,
|
|
|
|
|
scale=1,
|
|
|
|
|
minimum=0.6, maximum=1.65, step=0.05, label=i18n("语速"), value=1, interactive=True, scale=1
|
|
|
|
|
)
|
|
|
|
|
pause_second_slider = gr.Slider(
|
|
|
|
|
minimum=0.1,
|
|
|
|
|
@ -1329,46 +1233,22 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False) as app:
|
|
|
|
|
interactive=True,
|
|
|
|
|
scale=1,
|
|
|
|
|
)
|
|
|
|
|
gr.Markdown(
|
|
|
|
|
html_center(
|
|
|
|
|
i18n("GPT采样参数(无参考文本时不要太低。不懂就用默认)：")
|
|
|
|
|
)
|
|
|
|
|
)
|
|
|
|
|
gr.Markdown(html_center(i18n("GPT采样参数(无参考文本时不要太低。不懂就用默认)：")))
|
|
|
|
|
top_k = gr.Slider(
|
|
|
|
|
minimum=1,
|
|
|
|
|
maximum=100,
|
|
|
|
|
step=1,
|
|
|
|
|
label=i18n("top_k"),
|
|
|
|
|
value=15,
|
|
|
|
|
interactive=True,
|
|
|
|
|
scale=1,
|
|
|
|
|
minimum=1, maximum=100, step=1, label=i18n("top_k"), value=15, interactive=True, scale=1
|
|
|
|
|
)
|
|
|
|
|
top_p = gr.Slider(
|
|
|
|
|
minimum=0,
|
|
|
|
|
maximum=1,
|
|
|
|
|
step=0.05,
|
|
|
|
|
label=i18n("top_p"),
|
|
|
|
|
value=1,
|
|
|
|
|
interactive=True,
|
|
|
|
|
scale=1,
|
|
|
|
|
minimum=0, maximum=1, step=0.05, label=i18n("top_p"), value=1, interactive=True, scale=1
|
|
|
|
|
)
|
|
|
|
|
temperature = gr.Slider(
|
|
|
|
|
minimum=0,
|
|
|
|
|
maximum=1,
|
|
|
|
|
step=0.05,
|
|
|
|
|
label=i18n("temperature"),
|
|
|
|
|
value=1,
|
|
|
|
|
interactive=True,
|
|
|
|
|
scale=1,
|
|
|
|
|
minimum=0, maximum=1, step=0.05, label=i18n("temperature"), value=1, interactive=True, scale=1
|
|
|
|
|
)
|
|
|
|
|
# with gr.Column():
|
|
|
|
|
# gr.Markdown(value=i18n("手工调整音素。当音素框不为空时使用手工音素输入推理，无视目标文本框。"))
|
|
|
|
|
# phoneme=gr.Textbox(label=i18n("音素框"), value="")
|
|
|
|
|
# get_phoneme_button = gr.Button(i18n("目标文本转音素"), variant="primary")
|
|
|
|
|
with gr.Row():
|
|
|
|
|
inference_button = gr.Button(
|
|
|
|
|
value=i18n("合成语音"), variant="primary", size="lg", scale=25
|
|
|
|
|
)
|
|
|
|
|
inference_button = gr.Button(value=i18n("合成语音"), variant="primary", size="lg", scale=25)
|
|
|
|
|
output = gr.Audio(label=i18n("输出的语音"), scale=14)
|
|
|
|
|
|
|
|
|
|
inference_button.click(
|
|
|
|
|
|
RVC-Boss
2 months ago
committed by
GitHub