mirror of
https://github.com/RVC-Boss/GPT-SoVITS.git
synced 2026-06-03 20:40:30 +08:00
Merge 0235857b895183cae7d296389db16f3c783f189a into c767f0b83b998e996a4d230d86da575a03f54a3f
This commit is contained in:
hsoftxl
GitHub
commit
ac432bb24b
@ -1,6 +1,3 @@
|
||||
GPT_SoVITS/pretrained_models/*
|
||||
tools/asr/models/*
|
||||
tools/uvr5/uvr5_weights/*
|
||||
|
||||
.git
|
||||
.DS_Store
|
||||
@ -11,10 +8,7 @@ runtime
|
||||
.idea
|
||||
output
|
||||
logs
|
||||
SoVITS_weights*/
|
||||
GPT_weights*/
|
||||
TEMP
|
||||
weight.json
|
||||
ffmpeg*
|
||||
ffprobe*
|
||||
cfg.json
|
||||
|
||||
@ -18,7 +18,7 @@ ln -s /workspace/models/pretrained_models /workspace/GPT-SoVITS/GPT_SoVITS/pretr
|
||||
|
||||
ln -s /workspace/models/G2PWModel /workspace/GPT-SoVITS/GPT_SoVITS/text/G2PWModel
|
||||
|
||||
bash install.sh --device "CU${CUDA_VERSION//./}" --source HF
|
||||
bash install.sh --device "MPS" --source HF
|
||||
|
||||
pip cache purge
|
||||
|
||||
|
||||
70
Dockerfile
70
Dockerfile
@ -1,62 +1,20 @@
|
||||
ARG CUDA_VERSION=12.6
|
||||
ARG TORCH_BASE=full
|
||||
FROM python:3.10.18-bullseye
|
||||
|
||||
FROM xxxxrt666/torch-base:cu${CUDA_VERSION}-${TORCH_BASE}
|
||||
|
||||
LABEL maintainer="XXXXRT"
|
||||
LABEL version="V4"
|
||||
LABEL version="V2pro"
|
||||
LABEL description="Docker image for GPT-SoVITS"
|
||||
|
||||
ARG CUDA_VERSION=12.6
|
||||
WORKDIR /GPT-SoVITS
|
||||
COPY requirements.txt /GPT-SoVITS
|
||||
RUN pip install -r requirements.txt
|
||||
|
||||
ENV CUDA_VERSION=${CUDA_VERSION}
|
||||
COPY GPT_SoVITS /GPT-SoVITS/GPT_SoVITS
|
||||
COPY tools /GPT-SoVITS/tools
|
||||
COPY api.py /GPT-SoVITS
|
||||
COPY api_v2.py /GPT-SoVITS
|
||||
COPY config.py /GPT-SoVITS
|
||||
COPY webui.py /GPT-SoVITS
|
||||
COPY ref_audio /GPT-SoVITS/ref_audio
|
||||
|
||||
SHELL ["/bin/bash", "-c"]
|
||||
EXPOSE 9871 9872 9873 9874 9880 8001 8002
|
||||
|
||||
WORKDIR /workspace/GPT-SoVITS
|
||||
|
||||
COPY Docker /workspace/GPT-SoVITS/Docker/
|
||||
|
||||
ARG LITE=false
|
||||
ENV LITE=${LITE}
|
||||
|
||||
ARG WORKFLOW=false
|
||||
ENV WORKFLOW=${WORKFLOW}
|
||||
|
||||
ARG TARGETPLATFORM
|
||||
ENV TARGETPLATFORM=${TARGETPLATFORM}
|
||||
|
||||
RUN bash Docker/miniconda_install.sh
|
||||
|
||||
COPY extra-req.txt /workspace/GPT-SoVITS/
|
||||
|
||||
COPY requirements.txt /workspace/GPT-SoVITS/
|
||||
|
||||
COPY install.sh /workspace/GPT-SoVITS/
|
||||
|
||||
RUN bash Docker/install_wrapper.sh
|
||||
|
||||
EXPOSE 9871 9872 9873 9874 9880
|
||||
|
||||
ENV PYTHONPATH="/workspace/GPT-SoVITS"
|
||||
|
||||
RUN conda init bash && echo "conda activate base" >> ~/.bashrc
|
||||
|
||||
WORKDIR /workspace
|
||||
|
||||
RUN rm -rf /workspace/GPT-SoVITS
|
||||
|
||||
WORKDIR /workspace/GPT-SoVITS
|
||||
|
||||
COPY . /workspace/GPT-SoVITS
|
||||
|
||||
CMD ["/bin/bash", "-c", "\
|
||||
rm -rf /workspace/GPT-SoVITS/GPT_SoVITS/pretrained_models && \
|
||||
rm -rf /workspace/GPT-SoVITS/GPT_SoVITS/text/G2PWModel && \
|
||||
rm -rf /workspace/GPT-SoVITS/tools/asr/models && \
|
||||
rm -rf /workspace/GPT-SoVITS/tools/uvr5/uvr5_weights && \
|
||||
ln -s /workspace/models/pretrained_models /workspace/GPT-SoVITS/GPT_SoVITS/pretrained_models && \
|
||||
ln -s /workspace/models/G2PWModel /workspace/GPT-SoVITS/GPT_SoVITS/text/G2PWModel && \
|
||||
ln -s /workspace/models/asr_models /workspace/GPT-SoVITS/tools/asr/models && \
|
||||
ln -s /workspace/models/uvr5_weights /workspace/GPT-SoVITS/tools/uvr5/uvr5_weights && \
|
||||
exec bash"]
|
||||
CMD ["/bin/bash", "-c", "python GPT_SoVITS/inference_webui_api.py"]
|
||||
@ -1,11 +1,11 @@
|
||||
custom:
|
||||
bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
|
||||
cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
|
||||
device: cuda
|
||||
is_half: true
|
||||
t2s_weights_path: GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt
|
||||
device: cpu
|
||||
is_half: false
|
||||
t2s_weights_path: GPT_SoVITS/pretrained_models/meiv2pp-e15.ckpt
|
||||
version: v2
|
||||
vits_weights_path: GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2G2333k.pth
|
||||
vits_weights_path: GPT_SoVITS/pretrained_models/meiv2pp_e8_s232.pth
|
||||
v1:
|
||||
bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
|
||||
cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
|
||||
|
||||
685
GPT_SoVITS/inference_webui_api.py
Normal file
685
GPT_SoVITS/inference_webui_api.py
Normal file
@ -0,0 +1,685 @@
|
||||
"""
|
||||
按中英混合识别
|
||||
按日英混合识别
|
||||
多语种启动切分识别语种
|
||||
全部按中文识别
|
||||
全部按英文识别
|
||||
全部按日文识别
|
||||
"""
|
||||
|
||||
import json
|
||||
import logging
|
||||
import os
|
||||
import random
|
||||
import re
|
||||
import sys
|
||||
import time
|
||||
import io
|
||||
import traceback
|
||||
import wave
|
||||
import torch
|
||||
import numpy as np
|
||||
from fastapi.responses import StreamingResponse
|
||||
|
||||
now_dir = os.getcwd()
|
||||
sys.path.append(now_dir)
|
||||
sys.path.append("%s/GPT_SoVITS" % (now_dir))
|
||||
|
||||
logging.getLogger("markdown_it").setLevel(logging.ERROR)
|
||||
logging.getLogger("urllib3").setLevel(logging.ERROR)
|
||||
logging.getLogger("httpcore").setLevel(logging.ERROR)
|
||||
logging.getLogger("httpx").setLevel(logging.ERROR)
|
||||
logging.getLogger("asyncio").setLevel(logging.ERROR)
|
||||
logging.getLogger("charset_normalizer").setLevel(logging.ERROR)
|
||||
logging.getLogger("torchaudio._extension").setLevel(logging.ERROR)
|
||||
|
||||
|
||||
infer_ttswebui = os.environ.get("infer_ttswebui", 9872)
|
||||
infer_ttswebui = int(infer_ttswebui)
|
||||
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"]
|
||||
|
||||
is_half = eval(os.environ.get("is_half", "True")) and torch.cuda.is_available()
|
||||
gpt_path = os.environ.get("gpt_path", None)
|
||||
sovits_path = os.environ.get("sovits_path", None)
|
||||
cnhubert_base_path = os.environ.get("cnhubert_base_path", None)
|
||||
bert_path = os.environ.get("bert_path", None)
|
||||
version = model_version = os.environ.get("version", "v2")
|
||||
|
||||
import gradio as gr
|
||||
from TTS_infer_pack.text_segmentation_method import get_method
|
||||
from TTS_infer_pack.TTS import NO_PROMPT_ERROR, TTS, TTS_Config
|
||||
|
||||
from tools.assets import css, js, top_html
|
||||
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)
|
||||
|
||||
|
||||
# os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 确保直接启动推理UI时也能够设置。
|
||||
|
||||
if torch.cuda.is_available():
|
||||
device = "cuda"
|
||||
# elif torch.backends.mps.is_available():
|
||||
# device = "mps"
|
||||
else:
|
||||
device = "cpu"
|
||||
|
||||
# is_half = False
|
||||
# device = "cpu"
|
||||
|
||||
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_v1 if version == "v1" else dict_language_v2
|
||||
|
||||
cut_method = {
|
||||
i18n("不切"): "cut0",
|
||||
i18n("凑四句一切"): "cut1",
|
||||
i18n("凑50字一切"): "cut2",
|
||||
i18n("按中文句号。切"): "cut3",
|
||||
i18n("按英文句号.切"): "cut4",
|
||||
i18n("按标点符号切"): "cut5",
|
||||
}
|
||||
|
||||
from config import change_choices, get_weights_names, name2gpt_path, name2sovits_path
|
||||
|
||||
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)
|
||||
|
||||
tts_config = TTS_Config("GPT_SoVITS/configs/tts_infer.yaml")
|
||||
tts_config.device = device
|
||||
tts_config.is_half = is_half
|
||||
tts_config.version = version
|
||||
if gpt_path is not None:
|
||||
if "！" in gpt_path or "!" in gpt_path:
|
||||
gpt_path = name2gpt_path[gpt_path]
|
||||
tts_config.t2s_weights_path = gpt_path
|
||||
if sovits_path is not None:
|
||||
if "！" in sovits_path or "!" in sovits_path:
|
||||
sovits_path = name2sovits_path[sovits_path]
|
||||
tts_config.vits_weights_path = sovits_path
|
||||
if cnhubert_base_path is not None:
|
||||
tts_config.cnhuhbert_base_path = cnhubert_base_path
|
||||
if bert_path is not None:
|
||||
tts_config.bert_base_path = bert_path
|
||||
|
||||
print(tts_config)
|
||||
tts_pipeline = TTS(tts_config)
|
||||
gpt_path = tts_config.t2s_weights_path
|
||||
sovits_path = tts_config.vits_weights_path
|
||||
version = tts_config.version
|
||||
|
||||
|
||||
def inference(
|
||||
text,
|
||||
text_lang,
|
||||
ref_audio_path,
|
||||
aux_ref_audio_paths,
|
||||
prompt_text,
|
||||
prompt_lang,
|
||||
top_k,
|
||||
top_p,
|
||||
temperature,
|
||||
text_split_method,
|
||||
batch_size,
|
||||
speed_factor,
|
||||
ref_text_free,
|
||||
split_bucket,
|
||||
fragment_interval,
|
||||
seed,
|
||||
keep_random,
|
||||
parallel_infer,
|
||||
repetition_penalty,
|
||||
sample_steps,
|
||||
super_sampling,
|
||||
):
|
||||
|
||||
seed = -1 if keep_random else seed
|
||||
actual_seed = seed if seed not in [-1, "", None] else random.randint(0, 2**32 - 1)
|
||||
inputs = {
|
||||
"text": text,
|
||||
"text_lang": dict_language[text_lang],
|
||||
"ref_audio_path": ref_audio_path,
|
||||
"aux_ref_audio_paths": [item.name for item in aux_ref_audio_paths] if aux_ref_audio_paths is not None else [],
|
||||
"prompt_text": prompt_text if not ref_text_free else "",
|
||||
"prompt_lang": dict_language[prompt_lang],
|
||||
"top_k": top_k,
|
||||
"top_p": top_p,
|
||||
"temperature": temperature,
|
||||
"text_split_method": cut_method[text_split_method],
|
||||
"batch_size": int(batch_size),
|
||||
"speed_factor": float(speed_factor),
|
||||
"split_bucket": split_bucket,
|
||||
"return_fragment": False,
|
||||
"fragment_interval": fragment_interval,
|
||||
"seed": actual_seed,
|
||||
"parallel_infer": parallel_infer,
|
||||
"repetition_penalty": repetition_penalty,
|
||||
"sample_steps": int(sample_steps),
|
||||
"super_sampling": super_sampling,
|
||||
}
|
||||
|
||||
|
||||
logging.info(
|
||||
f"inference_button请求耗时: {inputs}"
|
||||
)
|
||||
try:
|
||||
|
||||
start_time = time.time()
|
||||
|
||||
for item in tts_pipeline.run(inputs):
|
||||
yield item, actual_seed
|
||||
|
||||
logging.info(
|
||||
f"TTS请求耗时: {time.time() - start_time:.3f}s | 文本: {text}"
|
||||
)
|
||||
except NO_PROMPT_ERROR:
|
||||
gr.Warning(i18n("V3不支持无参考文本模式，请填写参考文本！"))
|
||||
|
||||
|
||||
def custom_sort_key(s):
|
||||
# 使用正则表达式提取字符串中的数字部分和非数字部分
|
||||
parts = re.split("(\d+)", s)
|
||||
# 将数字部分转换为整数，非数字部分保持不变
|
||||
parts = [int(part) if part.isdigit() else part for part in parts]
|
||||
return parts
|
||||
|
||||
|
||||
if os.path.exists("./weight.json"):
|
||||
pass
|
||||
else:
|
||||
with open("./weight.json", "w", encoding="utf-8") as file:
|
||||
json.dump({"GPT": {}, "SoVITS": {}}, file)
|
||||
|
||||
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, 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]
|
||||
|
||||
from process_ckpt import get_sovits_version_from_path_fast
|
||||
|
||||
v3v4set = {"v3", "v4"}
|
||||
|
||||
|
||||
def change_sovits_weights(sovits_path, prompt_language=None, text_language=None):
|
||||
if "！" in sovits_path or "!" in sovits_path:
|
||||
sovits_path = name2sovits_path[sovits_path]
|
||||
global 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 = path_sovits + "SoVITS %s" % model_version + i18n("底模缺失，无法加载相应 LoRA 权重")
|
||||
gr.Warning(info)
|
||||
raise FileExistsError(info)
|
||||
dict_language = dict_language_v1 if version == "v1" else dict_language_v2
|
||||
if prompt_language is not None and text_language is not None:
|
||||
if prompt_language in list(dict_language.keys()):
|
||||
prompt_text_update, prompt_language_update = (
|
||||
{"__type__": "update"},
|
||||
{"__type__": "update", "value": prompt_language},
|
||||
)
|
||||
else:
|
||||
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}
|
||||
else:
|
||||
text_update = {"__type__": "update", "value": ""}
|
||||
text_language_update = {"__type__": "update", "value": i18n("中文")}
|
||||
if model_version in v3v4set:
|
||||
visible_sample_steps = True
|
||||
visible_inp_refs = False
|
||||
else:
|
||||
visible_sample_steps = False
|
||||
visible_inp_refs = True
|
||||
yield (
|
||||
{"__type__": "update", "choices": list(dict_language.keys())},
|
||||
{"__type__": "update", "choices": list(dict_language.keys())},
|
||||
prompt_text_update,
|
||||
prompt_language_update,
|
||||
text_update,
|
||||
text_language_update,
|
||||
{"__type__": "update", "interactive": visible_sample_steps, "value": 32},
|
||||
{"__type__": "update", "visible": visible_inp_refs},
|
||||
{"__type__": "update", "interactive": True if model_version not in v3v4set else False},
|
||||
{"__type__": "update", "value": i18n("模型加载中，请等待"), "interactive": False},
|
||||
)
|
||||
|
||||
tts_pipeline.init_vits_weights(sovits_path)
|
||||
yield (
|
||||
{"__type__": "update", "choices": list(dict_language.keys())},
|
||||
{"__type__": "update", "choices": list(dict_language.keys())},
|
||||
prompt_text_update,
|
||||
prompt_language_update,
|
||||
text_update,
|
||||
text_language_update,
|
||||
{"__type__": "update", "interactive": visible_sample_steps, "value": 32},
|
||||
{"__type__": "update", "visible": visible_inp_refs},
|
||||
{"__type__": "update", "interactive": True if model_version not in v3v4set else False},
|
||||
{"__type__": "update", "value": i18n("合成语音"), "interactive": True},
|
||||
)
|
||||
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))
|
||||
|
||||
|
||||
def change_gpt_weights(gpt_path):
|
||||
if "！" in gpt_path or "!" in gpt_path:
|
||||
gpt_path = name2gpt_path[gpt_path]
|
||||
tts_pipeline.init_t2s_weights(gpt_path)
|
||||
|
||||
|
||||
with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False, js=js, css=css) as app:
|
||||
gr.HTML(
|
||||
top_html.format(
|
||||
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.")
|
||||
+ 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.")
|
||||
),
|
||||
elem_classes="markdown",
|
||||
)
|
||||
|
||||
with gr.Column():
|
||||
# with gr.Group():
|
||||
gr.Markdown(value=i18n("模型切换"))
|
||||
with gr.Row():
|
||||
GPT_dropdown = gr.Dropdown(
|
||||
label=i18n("GPT模型列表"),
|
||||
choices=sorted(GPT_names, key=custom_sort_key),
|
||||
value=gpt_path,
|
||||
interactive=True,
|
||||
)
|
||||
SoVITS_dropdown = gr.Dropdown(
|
||||
label=i18n("SoVITS模型列表"),
|
||||
choices=sorted(SoVITS_names, key=custom_sort_key),
|
||||
value=sovits_path,
|
||||
interactive=True,
|
||||
)
|
||||
refresh_button = gr.Button(i18n("刷新模型路径"), variant="primary")
|
||||
refresh_button.click(fn=change_choices, inputs=[], outputs=[SoVITS_dropdown, GPT_dropdown])
|
||||
|
||||
with gr.Row():
|
||||
with gr.Column():
|
||||
gr.Markdown(value=i18n("*请上传并填写参考信息"))
|
||||
with gr.Row():
|
||||
inp_ref = gr.Audio(label=i18n("主参考音频(请上传3~10秒内参考音频，超过会报错！)"), type="filepath")
|
||||
inp_refs = gr.File(
|
||||
label=i18n("辅参考音频(可选多个，或不选)"),
|
||||
file_count="multiple",
|
||||
visible=True if model_version != "v3" else False,
|
||||
)
|
||||
prompt_text = gr.Textbox(label=i18n("主参考音频的文本"), value="", lines=2)
|
||||
with gr.Row():
|
||||
prompt_language = gr.Dropdown(
|
||||
label=i18n("主参考音频的语种"), choices=list(dict_language.keys()), value=i18n("中文")
|
||||
)
|
||||
with gr.Column():
|
||||
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。"),
|
||||
value=False,
|
||||
interactive=True if model_version != "v3" else False,
|
||||
show_label=True,
|
||||
)
|
||||
gr.Markdown(
|
||||
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。")
|
||||
)
|
||||
|
||||
with gr.Column():
|
||||
gr.Markdown(value=i18n("*请填写需要合成的目标文本和语种模式"))
|
||||
text = gr.Textbox(label=i18n("需要合成的文本"), value="", lines=20, max_lines=20)
|
||||
text_language = gr.Dropdown(
|
||||
label=i18n("需要合成的文本的语种"), choices=list(dict_language.keys()), value=i18n("中文")
|
||||
)
|
||||
|
||||
with gr.Group():
|
||||
gr.Markdown(value=i18n("推理设置"))
|
||||
with gr.Row():
|
||||
with gr.Column():
|
||||
with gr.Row():
|
||||
batch_size = gr.Slider(
|
||||
minimum=1, maximum=200, step=1, label=i18n("batch_size"), value=20, interactive=True
|
||||
)
|
||||
sample_steps = gr.Radio(
|
||||
label=i18n("采样步数(仅对V3/4生效)"), value=32, choices=[4, 8, 16, 32, 64, 128], visible=True
|
||||
)
|
||||
with gr.Row():
|
||||
fragment_interval = gr.Slider(
|
||||
minimum=0.01, maximum=1, step=0.01, label=i18n("分段间隔(秒)"), value=0.3, interactive=True
|
||||
)
|
||||
speed_factor = gr.Slider(
|
||||
minimum=0.6, maximum=1.65, step=0.05, label="语速", value=1.0, interactive=True
|
||||
)
|
||||
with gr.Row():
|
||||
top_k = gr.Slider(minimum=1, maximum=100, step=1, label=i18n("top_k"), value=5, interactive=True)
|
||||
top_p = gr.Slider(minimum=0, maximum=1, step=0.05, label=i18n("top_p"), value=1, interactive=True)
|
||||
with gr.Row():
|
||||
temperature = gr.Slider(
|
||||
minimum=0, maximum=1, step=0.05, label=i18n("temperature"), value=1, interactive=True
|
||||
)
|
||||
repetition_penalty = gr.Slider(
|
||||
minimum=0, maximum=2, step=0.05, label=i18n("重复惩罚"), value=1.35, interactive=True
|
||||
)
|
||||
|
||||
with gr.Column():
|
||||
with gr.Row():
|
||||
how_to_cut = gr.Dropdown(
|
||||
label=i18n("怎么切"),
|
||||
choices=[
|
||||
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"),
|
||||
],
|
||||
value=i18n("凑四句一切"),
|
||||
interactive=True,
|
||||
scale=1,
|
||||
)
|
||||
super_sampling = gr.Checkbox(
|
||||
label=i18n("音频超采样(仅对V3生效))"), value=False, interactive=True, show_label=True
|
||||
)
|
||||
|
||||
with gr.Row():
|
||||
parallel_infer = gr.Checkbox(label=i18n("并行推理"), value=True, interactive=True, show_label=True)
|
||||
split_bucket = gr.Checkbox(
|
||||
label=i18n("数据分桶(并行推理时会降低一点计算量)"),
|
||||
value=True,
|
||||
interactive=True,
|
||||
show_label=True,
|
||||
)
|
||||
|
||||
with gr.Row():
|
||||
seed = gr.Number(label=i18n("随机种子"), value=-1)
|
||||
keep_random = gr.Checkbox(label=i18n("保持随机"), value=True, interactive=True, show_label=True)
|
||||
|
||||
output = gr.Audio(label=i18n("输出的语音"))
|
||||
with gr.Row():
|
||||
inference_button = gr.Button(i18n("合成语音"), variant="primary")
|
||||
stop_infer = gr.Button(i18n("终止合成"), variant="primary")
|
||||
|
||||
|
||||
inference_button.click(
|
||||
inference,
|
||||
[
|
||||
text,
|
||||
text_language,
|
||||
inp_ref,
|
||||
inp_refs,
|
||||
prompt_text,
|
||||
prompt_language,
|
||||
top_k,
|
||||
top_p,
|
||||
temperature,
|
||||
how_to_cut,
|
||||
batch_size,
|
||||
speed_factor,
|
||||
ref_text_free,
|
||||
split_bucket,
|
||||
fragment_interval,
|
||||
seed,
|
||||
keep_random,
|
||||
parallel_infer,
|
||||
repetition_penalty,
|
||||
sample_steps,
|
||||
super_sampling,
|
||||
],
|
||||
[output, seed],
|
||||
)
|
||||
stop_infer.click(tts_pipeline.stop, [], [])
|
||||
SoVITS_dropdown.change(
|
||||
change_sovits_weights,
|
||||
[SoVITS_dropdown, prompt_language, text_language],
|
||||
[
|
||||
prompt_language,
|
||||
text_language,
|
||||
prompt_text,
|
||||
prompt_language,
|
||||
text,
|
||||
text_language,
|
||||
sample_steps,
|
||||
inp_refs,
|
||||
ref_text_free,
|
||||
inference_button,
|
||||
],
|
||||
) #
|
||||
GPT_dropdown.change(change_gpt_weights, [GPT_dropdown], [])
|
||||
|
||||
with gr.Group():
|
||||
gr.Markdown(
|
||||
value=i18n(
|
||||
"文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。"
|
||||
)
|
||||
)
|
||||
with gr.Row():
|
||||
text_inp = gr.Textbox(label=i18n("需要合成的切分前文本"), value="", lines=4)
|
||||
with gr.Column():
|
||||
_how_to_cut = gr.Radio(
|
||||
label=i18n("怎么切"),
|
||||
choices=[
|
||||
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"),
|
||||
],
|
||||
value=i18n("凑四句一切"),
|
||||
interactive=True,
|
||||
)
|
||||
cut_text = gr.Button(i18n("切分"), variant="primary")
|
||||
|
||||
def to_cut(text_inp, how_to_cut):
|
||||
if len(text_inp.strip()) == 0 or text_inp == []:
|
||||
return ""
|
||||
method = get_method(cut_method[how_to_cut])
|
||||
return method(text_inp)
|
||||
|
||||
text_opt = gr.Textbox(label=i18n("切分后文本"), value="", lines=4)
|
||||
cut_text.click(to_cut, [text_inp, _how_to_cut], [text_opt])
|
||||
gr.Markdown(value=i18n("后续将支持转音素、手工修改音素、语音合成分步执行。"))
|
||||
|
||||
|
||||
from fastapi import FastAPI, UploadFile, File, Form
|
||||
from fastapi.responses import FileResponse
|
||||
import tempfile
|
||||
import shutil
|
||||
import os
|
||||
from pydantic import BaseModel
|
||||
import soundfile as sf
|
||||
|
||||
app = FastAPI()
|
||||
|
||||
|
||||
class InferenceRequest(BaseModel):
|
||||
text: str
|
||||
text_lang: str = i18n("中文")
|
||||
ref_audio: str # 这里是base64编码的音频文件内容
|
||||
prompt_text: str
|
||||
prompt_lang: str = i18n("中文")
|
||||
top_k: int = 6
|
||||
top_p: float = 0.9
|
||||
temperature: float = 0.95
|
||||
text_split_method: str = i18n("按标点符号切")
|
||||
batch_size: int = 20
|
||||
speed_factor: float = 1.1
|
||||
ref_text_free: bool = False
|
||||
split_bucket: bool = True
|
||||
fragment_interval: float = 0.3
|
||||
seed: int = -1
|
||||
keep_random: bool = True
|
||||
parallel_infer: bool = True
|
||||
repetition_penalty: float = 1.45
|
||||
sample_steps: int = 32
|
||||
super_sampling: bool = False
|
||||
|
||||
@app.post("/tts")
|
||||
async def api_inference(req: InferenceRequest):
|
||||
|
||||
try:
|
||||
start_time = time.time()
|
||||
result = inference(
|
||||
text=req.text,
|
||||
text_lang=req.text_lang,
|
||||
ref_audio_path=req.ref_audio,
|
||||
aux_ref_audio_paths=None,
|
||||
prompt_text=req.prompt_text,
|
||||
prompt_lang=req.prompt_lang,
|
||||
top_k=req.top_k,
|
||||
top_p=req.top_p,
|
||||
temperature=req.temperature,
|
||||
text_split_method=req.text_split_method,
|
||||
batch_size=req.batch_size,
|
||||
speed_factor=req.speed_factor,
|
||||
ref_text_free=req.ref_text_free,
|
||||
split_bucket=req.split_bucket,
|
||||
fragment_interval=req.fragment_interval,
|
||||
seed=req.seed,
|
||||
keep_random=req.keep_random,
|
||||
parallel_infer=req.parallel_infer,
|
||||
repetition_penalty=req.repetition_penalty,
|
||||
sample_steps=req.sample_steps,
|
||||
super_sampling=req.super_sampling,
|
||||
)
|
||||
|
||||
logging.info(
|
||||
f"TTS请求infer ence耗时: {time.time() - start_time:.3f}s | 文本: {req.text}"
|
||||
)
|
||||
for wav_data, _ in result:
|
||||
sr, audio = wav_data
|
||||
# 确保音频数据为16位整数格式
|
||||
if not isinstance(audio, np.ndarray):
|
||||
audio = np.array(audio)
|
||||
if audio.dtype != np.int16:
|
||||
audio = (audio * 32768).astype(np.int16)
|
||||
|
||||
# 创建临时WAV文件
|
||||
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_wav:
|
||||
temp_path = temp_wav.name
|
||||
# 写入WAV格式
|
||||
import wave
|
||||
import struct
|
||||
with wave.open(temp_path, "wb") as wav_file:
|
||||
wav_file.setnchannels(1) # 单声道
|
||||
wav_file.setsampwidth(2) # 16位
|
||||
wav_file.setframerate(sr)
|
||||
wav_file.writeframes(audio.tobytes())
|
||||
logging.info(
|
||||
f"TTS请求耗时: {time.time() - start_time:.3f}s | 文本: {req.text}"
|
||||
)
|
||||
# 返回WAV文件
|
||||
return FileResponse(
|
||||
temp_path,
|
||||
media_type="audio/wav",
|
||||
headers={
|
||||
"Content-Disposition": "attachment;filename=output.wav"
|
||||
}
|
||||
)
|
||||
|
||||
except Exception as e:
|
||||
|
||||
traceback.print_exc()
|
||||
logging.error(f"Error during inference: {e}")
|
||||
# 返回错误信息
|
||||
return {"error": "未能生成音频"}
|
||||
|
||||
|
||||
def wav_chunk_streamer(infer_gen):
|
||||
def encode_wav_chunk(sr, audio):
|
||||
buffer = io.BytesIO()
|
||||
with wave.open(buffer, 'wb') as wav_file:
|
||||
wav_file.setnchannels(1)
|
||||
wav_file.setsampwidth(2)
|
||||
wav_file.setframerate(sr)
|
||||
wav_file.writeframes(audio.tobytes())
|
||||
return buffer.getvalue()
|
||||
|
||||
for audio, _ in infer_gen:
|
||||
audio_data = audio[0] if isinstance(audio[0], np.ndarray) else audio[1]
|
||||
yield encode_wav_chunk(32000, audio_data) # 每段 WAV 数据
|
||||
|
||||
|
||||
@app.post("/tts_stream")
|
||||
async def api_inference(req: InferenceRequest):
|
||||
try:
|
||||
infer_gen = inference(
|
||||
text=req.text,
|
||||
text_lang=i18n(req.text_lang),
|
||||
ref_audio_path=req.ref_audio,
|
||||
aux_ref_audio_paths=[],
|
||||
prompt_text=req.prompt_text,
|
||||
prompt_lang=i18n(req.prompt_lang),
|
||||
top_k=req.top_k,
|
||||
top_p=req.top_p,
|
||||
temperature=req.temperature,
|
||||
text_split_method=req.text_split_method,
|
||||
batch_size=req.batch_size,
|
||||
speed_factor=req.speed_factor,
|
||||
ref_text_free=req.ref_text_free,
|
||||
split_bucket=req.split_bucket,
|
||||
fragment_interval=req.fragment_interval,
|
||||
seed=req.seed,
|
||||
keep_random=req.keep_random,
|
||||
parallel_infer=req.parallel_infer,
|
||||
repetition_penalty=req.repetition_penalty,
|
||||
sample_steps=req.sample_steps,
|
||||
super_sampling=req.super_sampling,
|
||||
)
|
||||
|
||||
return StreamingResponse(
|
||||
wav_chunk_streamer(infer_gen),
|
||||
media_type="audio/wav",
|
||||
headers={
|
||||
"Content-Disposition": "inline; filename=output.wav"
|
||||
}
|
||||
)
|
||||
|
||||
except Exception as e:
|
||||
import traceback
|
||||
traceback.print_exc()
|
||||
return {"error": f"生成失败: {str(e)}"}
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
import uvicorn
|
||||
port = int(os.environ.get("PORT", 8001)) # 默认端口8001
|
||||
uvicorn.run(app, host="0.0.0.0", port=port)
|
||||
@ -29,6 +29,14 @@ import sys
|
||||
|
||||
import torch
|
||||
|
||||
import logging
|
||||
import time
|
||||
import numpy
|
||||
|
||||
# 在文件开头添加输出目录配置
|
||||
output_dir = os.environ.get("output_dir", "outputs")
|
||||
os.makedirs(output_dir, exist_ok=True)
|
||||
|
||||
now_dir = os.getcwd()
|
||||
sys.path.append(now_dir)
|
||||
sys.path.append("%s/GPT_SoVITS" % (now_dir))
|
||||
@ -170,6 +178,7 @@ def inference(
|
||||
sample_steps,
|
||||
super_sampling,
|
||||
):
|
||||
|
||||
seed = -1 if keep_random else seed
|
||||
actual_seed = seed if seed not in [-1, "", None] else random.randint(0, 2**32 - 1)
|
||||
inputs = {
|
||||
@ -194,9 +203,32 @@ def inference(
|
||||
"sample_steps": int(sample_steps),
|
||||
"super_sampling": super_sampling,
|
||||
}
|
||||
|
||||
|
||||
logging.info(
|
||||
f"inference_button请求耗时: {inputs}"
|
||||
)
|
||||
try:
|
||||
for item in tts_pipeline.run(inputs):
|
||||
yield item, actual_seed
|
||||
|
||||
start_time = time.time()
|
||||
|
||||
for audio in tts_pipeline.run(inputs):
|
||||
if isinstance(audio, tuple):
|
||||
# 保存到本地
|
||||
output_filename = f"tts_{int(time.time())}.wav"
|
||||
output_path = os.path.join(output_dir, output_filename)
|
||||
audio_data = audio[0] if isinstance(audio[0], numpy.ndarray) else audio[1]
|
||||
import soundfile as sf
|
||||
sf.write(output_path, audio_data, 32000)
|
||||
logging.info(f"音频已保存至: {output_path}")
|
||||
# 返回原始音频数据给 Gradio
|
||||
yield audio, actual_seed
|
||||
else:
|
||||
yield audio, actual_seed
|
||||
|
||||
logging.info(
|
||||
f"TTS请求耗时: {time.time() - start_time:.3f}s | 文本: {text}"
|
||||
)
|
||||
except NO_PROMPT_ERROR:
|
||||
gr.Warning(i18n("V3不支持无参考文本模式，请填写参考文本！"))
|
||||
|
||||
@ -433,6 +465,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False, js=js, css=css
|
||||
inference_button = gr.Button(i18n("合成语音"), variant="primary")
|
||||
stop_infer = gr.Button(i18n("终止合成"), variant="primary")
|
||||
|
||||
|
||||
inference_button.click(
|
||||
inference,
|
||||
[
|
||||
|
||||
BIN
GPT_SoVITS/text/ja_userdic/user.dict
Normal file
BIN
GPT_SoVITS/text/ja_userdic/user.dict
Normal file
Binary file not shown.
1
GPT_SoVITS/text/ja_userdic/userdict.md5
Normal file
1
GPT_SoVITS/text/ja_userdic/userdict.md5
Normal file
@ -0,0 +1 @@
|
||||
d36bd5ffba62f195d22bf4f1a41cd08f
|
||||
33
compress.sh
Normal file
33
compress.sh
Normal file
@ -0,0 +1,33 @@
|
||||
#!/bin/bash
|
||||
|
||||
# 定义压缩文件名（包含时间戳）
|
||||
ARCHIVE_NAME="gpt-sovits_$(date +%Y%m%d_%H%M%S).tar.gz"
|
||||
|
||||
# 创建临时目录
|
||||
TEMP_DIR=$(mktemp -d)
|
||||
echo "临时目录: $TEMP_DIR"
|
||||
DEST_DIR="$TEMP_DIR/GPT-SoVITS"
|
||||
echo "临时DEST_DIR目录: $DEST_DIR"
|
||||
mkdir -p "$DEST_DIR"
|
||||
|
||||
# 复制文件和目录到临时目录
|
||||
echo "复制文件开始..."
|
||||
cp -r GPT_SoVITS "$DEST_DIR/"
|
||||
cp -r tools "$DEST_DIR/"
|
||||
cp api.py "$DEST_DIR/"
|
||||
cp api_v2.py "$DEST_DIR/"
|
||||
cp config.py "$DEST_DIR/"
|
||||
cp webui.py "$DEST_DIR/"
|
||||
cp -r ref_audio "$DEST_DIR/"
|
||||
cp requirements.txt "$DEST_DIR/"
|
||||
cp install.sh "$DEST_DIR/"
|
||||
cp extra-req.txt "$DEST_DIR/"
|
||||
|
||||
echo "复制文件结束..."
|
||||
# 创建压缩包
|
||||
tar -czf "$ARCHIVE_NAME" -C "$TEMP_DIR" .
|
||||
|
||||
# 清理临时目录
|
||||
rm -rf "$TEMP_DIR"
|
||||
|
||||
echo "已创建压缩包: $ARCHIVE_NAME"
|
||||
0
ref_audio/1. 「苍城」？我怎么没听过，六座仙舟里有叫这个名字的吗？
Normal file
0
ref_audio/1. 「苍城」？我怎么没听过，六座仙舟里有叫这个名字的吗？
Normal file
BIN
ref_audio/1.wav
Normal file
BIN
ref_audio/1.wav
Normal file
Binary file not shown.
0
ref_audio/2. 把泰科铵大球馆砸出个洞的事，你自己说还是我来？
Normal file
0
ref_audio/2. 把泰科铵大球馆砸出个洞的事，你自己说还是我来？
Normal file
BIN
ref_audio/2.wav
Normal file
BIN
ref_audio/2.wav
Normal file
Binary file not shown.
0
ref_audio/3. 。而且参与的人多了，就会有麻烦
Normal file
0
ref_audio/3. 。而且参与的人多了，就会有麻烦
Normal file
BIN
ref_audio/3.wav
Normal file
BIN
ref_audio/3.wav
Normal file
Binary file not shown.
Loading…
x
Reference in New Issue
Block a user