mirror of
https://github.com/RVC-Boss/GPT-SoVITS.git
synced 2025-10-07 23:48:48 +08:00
476 lines
16 KiB
Python
476 lines
16 KiB
Python
import os, re, logging
|
||
import LangSegment
|
||
import pdb
|
||
import torch
|
||
import gradio as gr
|
||
from transformers import AutoModelForMaskedLM, AutoTokenizer
|
||
import numpy as np
|
||
import librosa
|
||
from feature_extractor import cnhubert
|
||
from module.models import SynthesizerTrn
|
||
from AR.models.t2s_lightning_module import Text2SemanticLightningModule
|
||
from text import cleaned_text_to_sequence
|
||
from text.cleaner import clean_text
|
||
from time import time as ttime
|
||
from module.mel_processing import spectrogram_torch
|
||
from tools.my_utils import load_audio
|
||
from tools.i18n.i18n import I18nAuto
|
||
import scipy.io.wavfile as wavfile
|
||
|
||
device = "cuda" if torch.cuda.is_available() else "cpu"
|
||
|
||
i18n = I18nAuto()
|
||
|
||
dict_language = {
|
||
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
|
||
}
|
||
|
||
is_share = os.environ.get("is_share", "False")
|
||
is_share = eval(is_share)
|
||
if "_CUDA_VISIBLE_DEVICES" in os.environ:
|
||
os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["_CUDA_VISIBLE_DEVICES"]
|
||
half_precision = True
|
||
is_half = half_precision and torch.cuda.is_available()
|
||
splits = {"，", "。", "？", "！", ",", ".", "?", "!", "~", ":", "：", "—", "…", }
|
||
punctuation = set(['!', '?', '…', ',', '.', '-', " "])
|
||
|
||
|
||
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")
|
||
|
||
|
||
def replace_consecutive_punctuation(text):
|
||
punctuations = ''.join(re.escape(p) for p in punctuation)
|
||
pattern = f'([{punctuations}])([{punctuations}])+'
|
||
result = re.sub(pattern, r'\1', text)
|
||
return result
|
||
|
||
|
||
def get_first(text):
|
||
pattern = "[" + "".join(re.escape(sep) for sep in splits) + "]"
|
||
text = re.split(pattern, text)[0].strip()
|
||
return text
|
||
|
||
|
||
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
|
||
|
||
|
||
# 四句一切
|
||
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)
|
||
|
||
|
||
# 句号切
|
||
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 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
|
||
|
||
|
||
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 clean_text_inf(text, language):
|
||
phones, word2ph, norm_text = clean_text(text, language)
|
||
phones = cleaned_text_to_sequence(phones)
|
||
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
|
||
|
||
|
||
dtype = torch.float16 if is_half else torch.float32
|
||
|
||
|
||
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 else torch.float32,
|
||
).to(device)
|
||
|
||
return bert
|
||
|
||
|
||
def get_spepc(hps, filename):
|
||
audio = load_audio(filename, int(hps.data.sampling_rate))
|
||
audio = torch.FloatTensor(audio)
|
||
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 get_phones_and_bert(text, language):
|
||
if language in {"en", "all_zh", "all_ja"}:
|
||
language = language.replace("all_", "")
|
||
if language == "en":
|
||
LangSegment.setfilters(["en"])
|
||
formattext = " ".join(tmp["text"] for tmp in LangSegment.getTexts(text))
|
||
else:
|
||
# 因无法区别中日文汉字,以用户输入为准
|
||
formattext = text
|
||
while " " in formattext:
|
||
formattext = formattext.replace(" ", " ")
|
||
phones, word2ph, norm_text = clean_text_inf(formattext, language)
|
||
if language == "zh":
|
||
bert = get_bert_feature(norm_text, word2ph).to(device)
|
||
else:
|
||
bert = torch.zeros(
|
||
(1024, len(phones)),
|
||
dtype=torch.float16 if is_half == True else torch.float32,
|
||
).to(device)
|
||
elif language in {"zh", "ja", "auto"}:
|
||
textlist = []
|
||
langlist = []
|
||
LangSegment.setfilters(["zh", "ja", "en", "ko"])
|
||
if language == "auto":
|
||
for tmp in LangSegment.getTexts(text):
|
||
if tmp["lang"] == "ko":
|
||
langlist.append("zh")
|
||
textlist.append(tmp["text"])
|
||
else:
|
||
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)
|
||
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)
|
||
|
||
return phones, bert.to(dtype), norm_text
|
||
|
||
|
||
def set_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:
|
||
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("./gweight.txt", "w", encoding="utf-8") as f: f.write(gpt_path)
|
||
|
||
|
||
def set_sovits_weights(sovits_path):
|
||
global vq_model, hps
|
||
dict_s2 = torch.load(sovits_path, map_location="cpu")
|
||
hps = dict_s2["config"]
|
||
hps = DictToAttrRecursive(hps)
|
||
hps.model.semantic_frame_rate = "25hz"
|
||
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:
|
||
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))
|
||
with open("./sweight.txt", "w", encoding="utf-8") as f:
|
||
f.write(sovits_path)
|
||
|
||
|
||
def gen_audio(ref_wav_path, prompt_text, text_to_speak, output_file, top_k=20, top_p=0.6, temperature=0.6, ref_free=False):
|
||
if prompt_text is None or len(prompt_text) == 0:
|
||
ref_free = True
|
||
t0 = ttime()
|
||
prompt_language = "zh"
|
||
text_language = "zh"
|
||
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_to_speak = text_to_speak.strip("\n")
|
||
text_to_speak = replace_consecutive_punctuation(text_to_speak)
|
||
if text_to_speak[0] not in splits and len(get_first(text_to_speak)) < 4:
|
||
text_to_speak = "。" + text_to_speak if text_language != "en" else "." + text_to_speak
|
||
|
||
print(i18n("实际输入的目标文本:"), text_to_speak)
|
||
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:
|
||
raise OSError(i18n("参考音频在3~10秒范围外，请更换！"))
|
||
wav16k = torch.from_numpy(wav16k)
|
||
zero_wav_torch = torch.from_numpy(zero_wav)
|
||
if is_half:
|
||
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()
|
||
# text_to_speak = cut1(text_to_speak)
|
||
text_to_speak = cut3(text_to_speak)
|
||
while "\n\n" in text_to_speak:
|
||
text_to_speak = text_to_speak.replace("\n\n", "\n")
|
||
print(i18n("实际输入的目标文本(切句后):"), text_to_speak)
|
||
texts = text_to_speak.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)
|
||
|
||
for text_to_speak in texts:
|
||
# 解决输入目标文本的空行导致报错的问题
|
||
if len(text_to_speak.strip()) == 0:
|
||
continue
|
||
if text_to_speak[-1] not in splits:
|
||
text_to_speak += "。" if text_language != "en" else "."
|
||
print(i18n("实际输入的目标文本(每句):"), text_to_speak)
|
||
phones2, bert2, norm_text2 = get_phones_and_bert(text_to_speak, text_language)
|
||
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()
|
||
with torch.no_grad():
|
||
# pred_semantic = t2s_model.model.infer(
|
||
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,
|
||
)
|
||
t3 = ttime()
|
||
# print(pred_semantic.shape,idx)
|
||
pred_semantic = pred_semantic[:, -idx:].unsqueeze(
|
||
0
|
||
) # .unsqueeze(0)#mq要多unsqueeze一次
|
||
refer = get_spepc(hps, ref_wav_path) # .to(device)
|
||
if is_half:
|
||
refer = refer.half().to(device)
|
||
else:
|
||
refer = refer.to(device)
|
||
# audio = vq_model.decode(pred_semantic, all_phoneme_ids, refer).detach().cpu().numpy()[0, 0]
|
||
audio = (
|
||
vq_model.decode(
|
||
pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer
|
||
)
|
||
.detach()
|
||
.cpu()
|
||
.numpy()[0, 0]
|
||
) # 试试重建不带上prompt部分
|
||
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()
|
||
|
||
# 将音频数据合并
|
||
audio_data = np.concatenate(audio_opt, 0) * 32768
|
||
audio_data = audio_data.astype(np.int16)
|
||
wavfile.write(output_file, hps.data.sampling_rate, audio_data)
|
||
|
||
print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3))
|
||
|
||
|
||
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"
|
||
)
|
||
|
||
cnhubert.cnhubert_base_path = cnhubert_base_path
|
||
ssl_model = cnhubert.get_model()
|
||
if is_half:
|
||
ssl_model = ssl_model.half().to(device)
|
||
else:
|
||
ssl_model = ssl_model.to(device)
|
||
|
||
tokenizer = AutoTokenizer.from_pretrained(bert_path)
|
||
bert_model = AutoModelForMaskedLM.from_pretrained(bert_path)
|
||
if is_half:
|
||
bert_model = bert_model.half().to(device)
|
||
else:
|
||
bert_model = bert_model.to(device)
|
||
|
||
|
||
def speak(text_to_speak):
|
||
sovits_path = "SoVITS_weights/阿贝多_e12_s2748.pth"
|
||
set_sovits_weights(sovits_path)
|
||
gpt_path = "GPT_weights/阿贝多-e10.ckpt"
|
||
set_gpt_weights(gpt_path)
|
||
ref_wav_path = "audio/首先，先看看这不明来源的元素力，究竟是如何对外流动的.wav"
|
||
prompt_text = "首先，先看看这不明来源的元素力，究竟是如何对外流动的。"
|
||
# text_to_speak = "我...我...我不知道你在说什么，我们之间没有秘密呀。可能你弄错了，我们平时关系很好的，请不要误会。"
|
||
# 创建一个时间戳的文件名
|
||
output_file = "outputs/" + str(int(ttime())) + ".wav"
|
||
gen_audio(ref_wav_path, prompt_text, text_to_speak, output_file)
|
||
return output_file
|
||
|
||
|
||
def main():
|
||
speak("放学了，我该回家了，你叫我留下来干什么？")
|
||
|
||
|
||
if __name__ == '__main__':
|
||
main()
|