699 lines
25 KiB
Python
699 lines
25 KiB
Python
import os,sys
|
||
from .ft_node import SoVITS_weight_root,GPT_weight_root,\
|
||
pretrained_sovits_name,pretrained_gpt_name,work_path,\
|
||
now_dir,gsv_path,models_dir
|
||
|
||
sys.path.append(now_dir)
|
||
sys.path.append(gsv_path)
|
||
from tools.i18n.i18n import I18nAuto, scan_language_list
|
||
|
||
language=os.environ.get("language","Auto")
|
||
language=sys.argv[-1] if sys.argv[-1] in scan_language_list() else language
|
||
i18n = I18nAuto(language=language)
|
||
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:
|
||
try:
|
||
for name in os.listdir(os.path.join(work_path,path)):
|
||
if name.endswith(".pth"): SoVITS_names.append("%s/%s" % (path, name))
|
||
except:
|
||
pass
|
||
GPT_names = [i for i in pretrained_gpt_name]
|
||
for path in GPT_weight_root:
|
||
try:
|
||
for name in os.listdir(os.path.join(work_path,path)):
|
||
if name.endswith(".ckpt"): GPT_names.append("%s/%s" % (path, name))
|
||
except:
|
||
pass
|
||
|
||
return SoVITS_names, GPT_names
|
||
|
||
dict_language_v1 = {
|
||
i18n("中文"): "all_zh",#全部按中文识别
|
||
i18n("英文"): "en",#全部按英文识别#######不变
|
||
i18n("日文"): "all_ja",#全部按日文识别
|
||
i18n("中英混合"): "zh",#按中英混合识别####不变
|
||
i18n("日英混合"): "ja",#按日英混合识别####不变
|
||
i18n("多语种混合"): "auto",#多语种启动切分识别语种
|
||
}
|
||
dict_language_v2 = {
|
||
i18n("中文"): "all_zh",#全部按中文识别
|
||
i18n("英文"): "en",#全部按英文识别#######不变
|
||
i18n("日文"): "all_ja",#全部按日文识别
|
||
i18n("粤语"): "all_yue",#全部按中文识别
|
||
i18n("韩文"): "all_ko",#全部按韩文识别
|
||
i18n("中英混合"): "zh",#按中英混合识别####不变
|
||
i18n("日英混合"): "ja",#按日英混合识别####不变
|
||
i18n("粤英混合"): "yue",#按粤英混合识别####不变
|
||
i18n("韩英混合"): "ko",#按韩英混合识别####不变
|
||
i18n("多语种混合"): "auto",#多语种启动切分识别语种
|
||
i18n("多语种混合(粤语)"): "auto_yue",#多语种启动切分识别语种
|
||
}
|
||
dict_language = dict_language_v2
|
||
|
||
import re
|
||
import torch
|
||
import numpy as np
|
||
from time import time as ttime
|
||
import cuda_malloc
|
||
device = "cuda" if cuda_malloc.cuda_malloc_supported() else "cpu"
|
||
|
||
from module.models import SynthesizerTrn
|
||
splits = {"，", "。", "？", "！", ",", ".", "?", "!", "~", ":", "：", "—", "…", }
|
||
|
||
|
||
def get_first(text):
|
||
pattern = "[" + "".join(re.escape(sep) for sep in splits) + "]"
|
||
text = re.split(pattern, text)[0].strip()
|
||
return text
|
||
class DictToAttrRecursive(dict):
|
||
def __init__(self, input_dict):
|
||
super().__init__(input_dict)
|
||
for key, value in input_dict.items():
|
||
if isinstance(value, dict):
|
||
value = DictToAttrRecursive(value)
|
||
self[key] = value
|
||
setattr(self, key, value)
|
||
|
||
def __getattr__(self, item):
|
||
try:
|
||
return self[item]
|
||
except KeyError:
|
||
raise AttributeError(f"Attribute {item} not found")
|
||
|
||
def __setattr__(self, key, value):
|
||
if isinstance(value, dict):
|
||
value = DictToAttrRecursive(value)
|
||
super(DictToAttrRecursive, self).__setitem__(key, value)
|
||
super().__setattr__(key, value)
|
||
|
||
def __delattr__(self, item):
|
||
try:
|
||
del self[item]
|
||
except KeyError:
|
||
raise AttributeError(f"Attribute {item} not found")
|
||
|
||
import GPT_SoVITS.utils as utils
|
||
def change_sovits_weights(sovits_path):
|
||
global vq_model, hps, version, dict_language
|
||
comfyui_utils = sys.modules['utils']
|
||
sys.modules['utils'] = utils
|
||
dict_s2 = torch.load(sovits_path, map_location="cpu")
|
||
sys.modules['utils'] = comfyui_utils
|
||
hps = dict_s2["config"]
|
||
hps = DictToAttrRecursive(hps)
|
||
hps.model.semantic_frame_rate = "25hz"
|
||
if dict_s2['weight']['enc_p.text_embedding.weight'].shape[0] == 322:
|
||
hps.model.version = "v1"
|
||
else:
|
||
hps.model.version = "v2"
|
||
version = hps.model.version
|
||
# print("sovits版本:",hps.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
|
||
)
|
||
if ("pretrained" not in sovits_path):
|
||
del vq_model.enc_q
|
||
if is_half == True:
|
||
vq_model = vq_model.half().to(device)
|
||
else:
|
||
vq_model = vq_model.to(device)
|
||
vq_model.eval()
|
||
print(vq_model.load_state_dict(dict_s2["weight"], strict=False))
|
||
dict_language = dict_language_v1 if version =='v1' else dict_language_v2
|
||
'''
|
||
with open("./weight.json")as f:
|
||
data=f.read()
|
||
data=json.loads(data)
|
||
data["SoVITS"][version]=sovits_path
|
||
with open("./weight.json","w")as f:f.write(json.dumps(data))
|
||
'''
|
||
from AR.models.t2s_lightning_module import Text2SemanticLightningModule
|
||
|
||
def change_gpt_weights(gpt_path):
|
||
global hz, max_sec, t2s_model, config
|
||
hz = 50
|
||
dict_s1 = torch.load(gpt_path, map_location="cpu")
|
||
config = dict_s1["config"]
|
||
max_sec = config["data"]["max_sec"]
|
||
t2s_model = Text2SemanticLightningModule(config, "****", is_train=False)
|
||
t2s_model.load_state_dict(dict_s1["weight"])
|
||
if is_half == True:
|
||
t2s_model = t2s_model.half()
|
||
t2s_model = t2s_model.to(device)
|
||
t2s_model.eval()
|
||
total = sum([param.nelement() for param in t2s_model.parameters()])
|
||
print("Number of parameter: %.2fM" % (total / 1e6))
|
||
'''
|
||
with open("./weight.json")as f:
|
||
data=f.read()
|
||
data=json.loads(data)
|
||
data["GPT"][version]=gpt_path
|
||
with open("./weight.json","w")as f:f.write(json.dumps(data))
|
||
'''
|
||
def process_text(texts):
|
||
_text=[]
|
||
if all(text in [None, " ", "\n",""] for text in texts):
|
||
raise ValueError(i18n("请输入有效文本"))
|
||
for text in texts:
|
||
if text in [None, " ", ""]:
|
||
pass
|
||
else:
|
||
_text.append(text)
|
||
return _text
|
||
|
||
from text import chinese,cleaned_text_to_sequence
|
||
from text.cleaner import clean_text
|
||
import LangSegment
|
||
from module.mel_processing import spectrogram_torch
|
||
from tools.my_utils import load_audio
|
||
|
||
def get_spepc(hps, audio):
|
||
# audio = load_audio(filename, int(hps.data.sampling_rate))
|
||
# audio = torch.FloatTensor(audio)
|
||
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,
|
||
hps.data.filter_length,
|
||
hps.data.sampling_rate,
|
||
hps.data.hop_length,
|
||
hps.data.win_length,
|
||
center=False,
|
||
)
|
||
return spec
|
||
|
||
def clean_text_inf(text, language, version):
|
||
phones, word2ph, norm_text = clean_text(text, language, version)
|
||
phones = cleaned_text_to_sequence(phones, version)
|
||
return phones, word2ph, norm_text
|
||
|
||
def get_bert_feature(text, word2ph):
|
||
with torch.no_grad():
|
||
inputs = tokenizer(text, return_tensors="pt")
|
||
for i in inputs:
|
||
inputs[i] = inputs[i].to(device)
|
||
res = bert_model(**inputs, output_hidden_states=True)
|
||
res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()[1:-1]
|
||
assert len(word2ph) == len(text)
|
||
phone_level_feature = []
|
||
for i in range(len(word2ph)):
|
||
repeat_feature = res[i].repeat(word2ph[i], 1)
|
||
phone_level_feature.append(repeat_feature)
|
||
phone_level_feature = torch.cat(phone_level_feature, dim=0)
|
||
return phone_level_feature.T
|
||
|
||
def get_bert_inf(phones, word2ph, norm_text, language):
|
||
language=language.replace("all_","")
|
||
if language == "zh":
|
||
bert = get_bert_feature(norm_text, word2ph).to(device)#.to(dtype)
|
||
else:
|
||
bert = torch.zeros(
|
||
(1024, len(phones)),
|
||
dtype=torch.float16 if is_half == True else torch.float32,
|
||
).to(device)
|
||
|
||
return bert
|
||
|
||
def get_phones_and_bert(text,language,version):
|
||
if language in {"en", "all_zh", "all_ja", "all_ko", "all_yue"}:
|
||
language = language.replace("all_","")
|
||
if language == "en":
|
||
LangSegment.setfilters(["en"])
|
||
formattext = " ".join(tmp["text"] for tmp in LangSegment.getTexts(text))
|
||
else:
|
||
# 因无法区别中日韩文汉字,以用户输入为准
|
||
formattext = text
|
||
while " " in formattext:
|
||
formattext = formattext.replace(" ", " ")
|
||
if language == "zh":
|
||
if re.search(r'[A-Za-z]', formattext):
|
||
formattext = re.sub(r'[a-z]', lambda x: x.group(0).upper(), formattext)
|
||
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)
|
||
bert = get_bert_feature(norm_text, word2ph).to(device)
|
||
elif language == "yue" and re.search(r'[A-Za-z]', formattext):
|
||
formattext = re.sub(r'[a-z]', lambda x: x.group(0).upper(), formattext)
|
||
formattext = chinese.mix_text_normalize(formattext)
|
||
return get_phones_and_bert(formattext,"yue",version)
|
||
else:
|
||
phones, word2ph, norm_text = clean_text_inf(formattext, language, version)
|
||
bert = torch.zeros(
|
||
(1024, len(phones)),
|
||
dtype=torch.float16 if is_half == True else torch.float32,
|
||
).to(device)
|
||
elif language in {"zh", "ja", "ko", "yue", "auto", "auto_yue"}:
|
||
textlist=[]
|
||
langlist=[]
|
||
LangSegment.setfilters(["zh","ja","en","ko"])
|
||
if language == "auto":
|
||
for tmp in LangSegment.getTexts(text):
|
||
langlist.append(tmp["lang"])
|
||
textlist.append(tmp["text"])
|
||
elif language == "auto_yue":
|
||
for tmp in LangSegment.getTexts(text):
|
||
if tmp["lang"] == "zh":
|
||
tmp["lang"] = "yue"
|
||
langlist.append(tmp["lang"])
|
||
textlist.append(tmp["text"])
|
||
else:
|
||
for tmp in LangSegment.getTexts(text):
|
||
if tmp["lang"] == "en":
|
||
langlist.append(tmp["lang"])
|
||
else:
|
||
# 因无法区别中日韩文汉字,以用户输入为准
|
||
langlist.append(language)
|
||
textlist.append(tmp["text"])
|
||
print(textlist)
|
||
print(langlist)
|
||
phones_list = []
|
||
bert_list = []
|
||
norm_text_list = []
|
||
for i in range(len(textlist)):
|
||
lang = langlist[i]
|
||
phones, word2ph, norm_text = clean_text_inf(textlist[i], lang, version)
|
||
bert = get_bert_inf(phones, word2ph, norm_text, lang)
|
||
phones_list.append(phones)
|
||
norm_text_list.append(norm_text)
|
||
bert_list.append(bert)
|
||
bert = torch.cat(bert_list, dim=1)
|
||
phones = sum(phones_list, [])
|
||
norm_text = ''.join(norm_text_list)
|
||
dtype=torch.float16 if is_half == True else torch.float32
|
||
return phones,bert.to(dtype),norm_text
|
||
|
||
|
||
def merge_short_text_in_array(texts, threshold):
|
||
if (len(texts)) < 2:
|
||
return texts
|
||
result = []
|
||
text = ""
|
||
for ele in texts:
|
||
text += ele
|
||
if len(text) >= threshold:
|
||
result.append(text)
|
||
text = ""
|
||
if (len(text) > 0):
|
||
if len(result) == 0:
|
||
result.append(text)
|
||
else:
|
||
result[len(result) - 1] += text
|
||
return result
|
||
|
||
def get_tts_wav(ref_wav,prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切"), top_k=20, top_p=0.6, temperature=0.6, speed=1):
|
||
t= []
|
||
if len(prompt_text) == 0:
|
||
ref_free = True
|
||
t0 = ttime()
|
||
prompt_language = dict_language[prompt_language]
|
||
text_language = dict_language[text_language]
|
||
|
||
if not ref_free:
|
||
prompt_text = prompt_text.strip("\n")
|
||
if (prompt_text[-1] not in splits): prompt_text += "。" if prompt_language != "en" else "."
|
||
print(i18n("实际输入的参考文本:"), prompt_text)
|
||
text = text.strip("\n")
|
||
if (text[0] not in splits and len(get_first(text)) < 4): text = "。" + text if text_language != "en" else "." + text
|
||
|
||
print(i18n("实际输入的目标文本:"), text)
|
||
zero_wav = np.zeros(
|
||
int(hps.data.sampling_rate * 0.3),
|
||
dtype=np.float16 if is_half == True else np.float32,
|
||
)
|
||
if not ref_free:
|
||
with torch.no_grad():
|
||
'''
|
||
wav16k, sr = librosa.load(ref_wav_path, sr=16000)
|
||
if (wav16k.shape[0] > 160000 or wav16k.shape[0] < 48000):
|
||
gr.Warning(i18n("参考音频在3~10秒范围外，请更换！"))
|
||
raise OSError(i18n("参考音频在3~10秒范围外，请更换！"))
|
||
wav16k = torch.from_numpy(wav16k)
|
||
'''
|
||
wav16k = ref_wav
|
||
zero_wav_torch = torch.from_numpy(zero_wav)
|
||
if is_half == True:
|
||
wav16k = wav16k.half().to(device)
|
||
zero_wav_torch = zero_wav_torch.half().to(device)
|
||
else:
|
||
wav16k = wav16k.to(device)
|
||
zero_wav_torch = zero_wav_torch.to(device)
|
||
wav16k = torch.cat([wav16k, zero_wav_torch])
|
||
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)
|
||
|
||
t1 = ttime()
|
||
t.append(t1-t0)
|
||
|
||
if (how_to_cut == i18n("凑四句一切")):
|
||
text = cut1(text)
|
||
elif (how_to_cut == i18n("凑50字一切")):
|
||
text = cut2(text)
|
||
elif (how_to_cut == i18n("按中文句号。切")):
|
||
text = cut3(text)
|
||
elif (how_to_cut == i18n("按英文句号.切")):
|
||
text = cut4(text)
|
||
elif (how_to_cut == i18n("按标点符号切")):
|
||
text = cut5(text)
|
||
while "\n\n" in text:
|
||
text = text.replace("\n\n", "\n")
|
||
print(i18n("实际输入的目标文本(切句后):"), text)
|
||
texts = text.split("\n")
|
||
texts = process_text(texts)
|
||
texts = merge_short_text_in_array(texts, 5)
|
||
audio_opt = []
|
||
if not ref_free:
|
||
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
|
||
if (len(text.strip()) == 0):
|
||
continue
|
||
if (text[-1] not in splits): text += "。" if text_language != "en" else "."
|
||
print(i18n("实际输入的目标文本(每句):"), text)
|
||
phones2,bert2,norm_text2=get_phones_and_bert(text, text_language, version)
|
||
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)
|
||
else:
|
||
bert = bert2
|
||
all_phoneme_ids = torch.LongTensor(phones2).to(device).unsqueeze(0)
|
||
|
||
bert = bert.to(device).unsqueeze(0)
|
||
all_phoneme_len = torch.tensor([all_phoneme_ids.shape[-1]]).to(device)
|
||
|
||
t2 = ttime()
|
||
# cache_key="%s-%s-%s-%s-%s-%s-%s-%s"%(ref_wav_path,prompt_text,prompt_language,text,text_language,top_k,top_p,temperature)
|
||
# print(cache.keys(),if_freeze)
|
||
'''
|
||
if(i_text in cache and if_freeze==True):pred_semantic=cache[i_text]
|
||
else:
|
||
with torch.no_grad():
|
||
pred_semantic, idx = t2s_model.model.infer_panel(
|
||
all_phoneme_ids,
|
||
all_phoneme_len,
|
||
None if ref_free else prompt,
|
||
bert,
|
||
# prompt_phone_len=ph_offset,
|
||
top_k=top_k,
|
||
top_p=top_p,
|
||
temperature=temperature,
|
||
early_stop_num=hz * max_sec,
|
||
)
|
||
pred_semantic = pred_semantic[:, -idx:].unsqueeze(0)
|
||
cache[i_text]=pred_semantic
|
||
'''
|
||
with torch.no_grad():
|
||
pred_semantic, idx = t2s_model.model.infer_panel(
|
||
all_phoneme_ids,
|
||
all_phoneme_len,
|
||
None if ref_free else prompt,
|
||
bert,
|
||
# prompt_phone_len=ph_offset,
|
||
top_k=top_k,
|
||
top_p=top_p,
|
||
temperature=temperature,
|
||
early_stop_num=hz * max_sec,
|
||
)
|
||
pred_semantic = pred_semantic[:, -idx:].unsqueeze(0)
|
||
t3 = ttime()
|
||
refers=[]
|
||
'''
|
||
if(inp_refs):
|
||
for path in inp_refs:
|
||
try:
|
||
refer = get_spepc(hps, path.name).to(dtype).to(device)
|
||
refers.append(refer)
|
||
except:
|
||
traceback.print_exc()
|
||
'''
|
||
dtype=torch.float16 if is_half == True else torch.float32
|
||
if(len(refers)==0):refers = [get_spepc(hps, ref_wav).to(dtype).to(device)]
|
||
audio = (vq_model.decode(pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refers,speed=speed).detach().cpu().numpy()[0, 0])
|
||
max_audio=np.abs(audio).max()#简单防止16bit爆音
|
||
if max_audio>1:audio/=max_audio
|
||
audio_opt.append(audio)
|
||
audio_opt.append(zero_wav)
|
||
t4 = ttime()
|
||
t.extend([t2 - t1,t3 - t2, t4 - t3])
|
||
t1 = ttime()
|
||
print("%.3f\t%.3f\t%.3f\t%.3f" %
|
||
(t[0], sum(t[1::3]), sum(t[2::3]), sum(t[3::3]))
|
||
)
|
||
return torch.Tensor(np.concatenate(audio_opt, 0)).unsqueeze(0)
|
||
'''
|
||
yield hps.data.sampling_rate, (np.concatenate(audio_opt, 0) * 32768).astype(
|
||
np.int16
|
||
)
|
||
'''
|
||
|
||
punctuation = set(['!', '?', '…', ',', '.', '-'," "])
|
||
|
||
def split(todo_text):
|
||
todo_text = todo_text.replace("……", "。").replace("——", "，")
|
||
if todo_text[-1] not in splits:
|
||
todo_text += "。"
|
||
i_split_head = i_split_tail = 0
|
||
len_text = len(todo_text)
|
||
todo_texts = []
|
||
while 1:
|
||
if i_split_head >= len_text:
|
||
break # 结尾一定有标点，所以直接跳出即可，最后一段在上次已加入
|
||
if todo_text[i_split_head] in splits:
|
||
i_split_head += 1
|
||
todo_texts.append(todo_text[i_split_tail:i_split_head])
|
||
i_split_tail = i_split_head
|
||
else:
|
||
i_split_head += 1
|
||
return todo_texts
|
||
|
||
|
||
def cut1(inp):
|
||
inp = inp.strip("\n")
|
||
inps = split(inp)
|
||
split_idx = list(range(0, len(inps), 4))
|
||
split_idx[-1] = None
|
||
if len(split_idx) > 1:
|
||
opts = []
|
||
for idx in range(len(split_idx) - 1):
|
||
opts.append("".join(inps[split_idx[idx]: split_idx[idx + 1]]))
|
||
else:
|
||
opts = [inp]
|
||
opts = [item for item in opts if not set(item).issubset(punctuation)]
|
||
return "\n".join(opts)
|
||
|
||
|
||
def cut2(inp):
|
||
inp = inp.strip("\n")
|
||
inps = split(inp)
|
||
if len(inps) < 2:
|
||
return inp
|
||
opts = []
|
||
summ = 0
|
||
tmp_str = ""
|
||
for i in range(len(inps)):
|
||
summ += len(inps[i])
|
||
tmp_str += inps[i]
|
||
if summ > 50:
|
||
summ = 0
|
||
opts.append(tmp_str)
|
||
tmp_str = ""
|
||
if tmp_str != "":
|
||
opts.append(tmp_str)
|
||
# print(opts)
|
||
if len(opts) > 1 and len(opts[-1]) < 50: ##如果最后一个太短了，和前一个合一起
|
||
opts[-2] = opts[-2] + opts[-1]
|
||
opts = opts[:-1]
|
||
opts = [item for item in opts if not set(item).issubset(punctuation)]
|
||
return "\n".join(opts)
|
||
|
||
|
||
def cut3(inp):
|
||
inp = inp.strip("\n")
|
||
opts = ["%s" % item for item in inp.strip("。").split("。")]
|
||
opts = [item for item in opts if not set(item).issubset(punctuation)]
|
||
return "\n".join(opts)
|
||
|
||
def cut4(inp):
|
||
inp = inp.strip("\n")
|
||
opts = ["%s" % item for item in inp.strip(".").split(".")]
|
||
opts = [item for item in opts if not set(item).issubset(punctuation)]
|
||
return "\n".join(opts)
|
||
|
||
|
||
# contributed by https://github.com/AI-Hobbyist/GPT-SoVITS/blob/main/GPT_SoVITS/inference_webui.py
|
||
def cut5(inp):
|
||
inp = inp.strip("\n")
|
||
punds = {',', '.', ';', '?', '!', '、', '，', '。', '？', '！', ';', '：', '…'}
|
||
mergeitems = []
|
||
items = []
|
||
|
||
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():
|
||
items.append(char)
|
||
else:
|
||
items.append(char)
|
||
mergeitems.append("".join(items))
|
||
items = []
|
||
else:
|
||
items.append(char)
|
||
|
||
if items:
|
||
mergeitems.append("".join(items))
|
||
|
||
opt = [item for item in mergeitems if not set(item).issubset(punds)]
|
||
return "\n".join(opt)
|
||
|
||
class TextDictNode:
|
||
@classmethod
|
||
def INPUT_TYPES(s):
|
||
return {
|
||
"required": {
|
||
"text": ("STRING", {"multiline": True, "dynamicPrompts": True}),
|
||
"language": (list(dict_language.keys()),)
|
||
}
|
||
}
|
||
RETURN_TYPES = ("TEXTDICT",)
|
||
FUNCTION = "encode"
|
||
|
||
CATEGORY = "AIFSH_GPT-SoVITS"
|
||
|
||
def encode(self,text,language):
|
||
res = {
|
||
"text": text,
|
||
"language":language
|
||
}
|
||
return (res, )
|
||
|
||
|
||
prompt_sr = 32000
|
||
import torchaudio
|
||
from feature_extractor import cnhubert
|
||
from transformers import AutoModelForMaskedLM, AutoTokenizer
|
||
|
||
class GSVTTSNode:
|
||
|
||
def __init__(self):
|
||
self.GPT_weight = None
|
||
self.SoVITS_weight = None
|
||
|
||
@classmethod
|
||
def INPUT_TYPES(s):
|
||
SoVITS_names, GPT_names = get_weights_names(GPT_weight_root, SoVITS_weight_root)
|
||
return {
|
||
"required":{
|
||
"text_dict": ("TEXTDICT",),
|
||
"prompt_text_dict":("TEXTDICT",),
|
||
"prompt_audio":("AUDIO",),
|
||
"config":("CONFIG",),
|
||
"GPT_weight":(GPT_names,),
|
||
"SoVITS_weight":(SoVITS_names,),
|
||
"how_to_cut":([i18n("不切"), i18n("凑四句一切"), i18n("凑50字一切"), i18n("按中文句号。切"), i18n("按英文句号.切"), i18n("按标点符号切"), ],{
|
||
"default": i18n("凑四句一切")
|
||
}),
|
||
"speed":("FLOAT",{
|
||
"min": 0.6,
|
||
"max":1.65,
|
||
"step":0.05,
|
||
"rond": 0.001,
|
||
"display":"slider",
|
||
"default": 1.0
|
||
}),
|
||
"top_k":("INT",{
|
||
"min": 1,
|
||
"max":100,
|
||
"step":1,
|
||
"display":"slider",
|
||
"default": 15
|
||
}),
|
||
"top_p":("FLOAT",{
|
||
"min": 0.,
|
||
"max":1.,
|
||
"step":0.05,
|
||
"rond": 0.001,
|
||
"display":"slider",
|
||
"default": 1.0
|
||
}),
|
||
"temperature":("FLOAT",{
|
||
"min": 0.,
|
||
"max":1.,
|
||
"step":0.05,
|
||
"rond": 0.001,
|
||
"display":"slider",
|
||
"default": 1.0
|
||
}),
|
||
}
|
||
}
|
||
|
||
RETURN_TYPES = ("AUDIO",)
|
||
FUNCTION = "tts"
|
||
|
||
CATEGORY = "AIFSH_GPT-SoVITS"
|
||
|
||
def tts(self,text_dict,prompt_text_dict,prompt_audio,
|
||
config,GPT_weight,SoVITS_weight,how_to_cut,
|
||
speed,top_k,top_p,temperature):
|
||
global ssl_model,is_half,tokenizer,bert_model
|
||
|
||
is_half = config['is_half']
|
||
if self.GPT_weight is None:
|
||
cnhubert.cnhubert_base_path = os.path.join(models_dir,"chinese-hubert-base")
|
||
ssl_model = cnhubert.get_model()
|
||
if is_half == True:
|
||
ssl_model = ssl_model.half().to(device)
|
||
else:
|
||
ssl_model = ssl_model.to(device)
|
||
|
||
|
||
bert_path = os.path.join(models_dir,"chinese-roberta-wwm-ext-large")
|
||
|
||
tokenizer = AutoTokenizer.from_pretrained(bert_path)
|
||
bert_model = AutoModelForMaskedLM.from_pretrained(bert_path)
|
||
if is_half == True:
|
||
bert_model = bert_model.half().to(device)
|
||
else:
|
||
bert_model = bert_model.to(device)
|
||
|
||
text = text_dict ['text']
|
||
text_language = text_dict['language']
|
||
prompt_text = prompt_text_dict['text']
|
||
prompt_language = prompt_text_dict['language']
|
||
|
||
waveform = prompt_audio['waveform'].squeeze(0)
|
||
source_sr = prompt_audio['sample_rate']
|
||
speech = waveform.mean(dim=0,keepdim=True)
|
||
if source_sr != prompt_sr:
|
||
speech = torchaudio.transforms.Resample(orig_freq=source_sr, new_freq=prompt_sr)(speech)
|
||
if self.SoVITS_weight != SoVITS_weight:
|
||
self.SoVITS_weight = SoVITS_weight
|
||
SoVITS_weight_path = os.path.join(models_dir,SoVITS_weight) if "s2G" in SoVITS_weight else os.path.join(work_path,SoVITS_weight)
|
||
change_sovits_weights(SoVITS_weight_path)
|
||
|
||
if self.GPT_weight != GPT_weight:
|
||
self.GPT_weight = GPT_weight
|
||
GPT_weight_path = os.path.join(models_dir,GPT_weight) if "epoch=" in GPT_weight else os.path.join(work_path,GPT_weight)
|
||
change_gpt_weights(GPT_weight_path)
|
||
|
||
res_audio = get_tts_wav(speech.squeeze(0),prompt_text,prompt_language,text,
|
||
text_language,how_to_cut,top_k,top_p,
|
||
temperature,speed)
|
||
res = {
|
||
"waveform": res_audio.unsqueeze(0),
|
||
"sample_rate": 32000,
|
||
}
|
||
return (res,)
|