Merge branch 'dev-api' into dev

README.md

@@ -24,6 +24,10 @@ This repository is based on [Bert-VITS2](https://github.com/fishaudio/Bert-VITS2
 
 <!-- 詳しくは[こちら](docs/tutorial.md)を参照してください。 -->
 
+### 動作環境
+
+各UIとAPI Serverにおいて、Windows コマンドプロンプト・WSL2・Linux(Ubuntu Desktop)での動作を確認しています(WSLでのパス指定は相対パスなど工夫ください)。
+
 ### インストール
 
 #### GitやPythonに馴染みが無い方
@@ -90,6 +94,14 @@ model_assets
 
 注意: データセットの手動修正やノイズ除去や、より高品質なデータセットを作りたい場合は、[Aivis](https://github.com/tsukumijima/Aivis)や、そのデータセット部分のWindows対応版 [Aivis Dataset](https://github.com/litagin02/Aivis-Dataset) を使うのをおすすめします。
 
+### API Server
+
+構築した環境下で`python server_fastapi.py`するとAPIサーバーが起動します。
+API仕様は起動後に`/docs`にて確認ください。
+
+デフォルトではCORS設定を全てのドメインで許可しています。
+できる限り、`config.yml`の`server.origins`の値を変更し、信頼できるドメインに制限ください(キーを消せばCORS設定を無効にできます)。
+
 ## Bert-VITS2 v2.1との関係
 
 基本的にはBert-VITS2 v2.1のモデル構造を少し改造しただけです。[事前学習モデル](https://huggingface.co/litagin/Style-Bert-VITS2-1.0-base)も、実質Bert-VITS2 v2.1と同じものを使用しています（不要な重みを削ってsafetensorsに変換したもの）。
@@ -109,7 +121,7 @@ model_assets
 - [ ] LinuxやWSL等、Windowsの通常環境以外でのサポート？
 - [ ] 複数話者学習での音声合成対応（学習は現在でも可能）
 - [ ] 本家のver 2.1, 2.2, 2.3モデルの推論対応？（ver 2.1以外は明らかにめんどいのでたぶんやらない）
-- [ ] `server_fastapi.py`の対応、とくにAPIで使えるようになると嬉しい人が増えるのかもしれない
+- [x] `server_fastapi.py`の対応、とくにAPIで使えるようになると嬉しい人が増えるのかもしれない
 - [ ] モデルのマージで声音と感情表現を混ぜる機能の実装
 - [ ] 英語等多言語対応？
 

app.py

@@ -3,11 +3,13 @@ import datetime
 import os
 import sys
 import warnings
+import enum
 
 import gradio as gr
 import numpy as np
 import torch
 from gradio.processing_utils import convert_to_16_bit_wav
+from typing import Dict, List
 
 import utils
 from config import config
@@ -15,17 +17,33 @@ from infer import get_net_g, infer
 from tools.log import logger
 
 
+class Languages(str, enum.Enum):
+    JP = "JP"
+    EN = "EN"
+    ZH = "ZH"
+
+
+languages = [l.value for l in Languages]
+
+DEFAULT_SDP_RATIO: float = 0.2
+DEFAULT_NOISE: float = 0.6
+DEFAULT_NOISEW: float = 0.8
+DEFAULT_LENGTH: float = 1
+DEFAULT_LINE_SPLIT: bool = True
+DEFAULT_SPLIT_INTERVAL: float = 0.5
+DEFAULT_STYLE_WEIGHT: float = 0.7
+DEFAULT_EMOTION_WEIGHT: float = 1.0
+
+
 class Model:
     def __init__(self, model_path, config_path, style_vec_path, device):
         self.model_path = model_path
         self.config_path = config_path
         self.device = device
         self.style_vec_path = style_vec_path
-        self.load()
-
-    def load(self):
         self.hps = utils.get_hparams_from_file(self.config_path)
-        self.spk2id = self.hps.data.spk2id
+        self.spk2id: Dict[str, int] = self.hps.data.spk2id
+        self.id2spk: Dict[int, str] = {v: k for k, v in self.spk2id.items()}
         self.num_styles = self.hps.data.num_styles
         if hasattr(self.hps.data, "style2id"):
             self.style2id = self.hps.data.style2id
@@ -63,17 +81,17 @@ class Model:
         language="JP",
         sid=0,
         reference_audio_path=None,
-        sdp_ratio=0.2,
-        noise=0.6,
-        noisew=0.8,
-        length=1.0,
-        line_split=True,
-        split_interval=0.2,
+        sdp_ratio=DEFAULT_SDP_RATIO,
+        noise=DEFAULT_NOISE,
+        noisew=DEFAULT_NOISEW,
+        length=DEFAULT_LENGTH,
+        line_split=DEFAULT_LINE_SPLIT,
+        split_interval=DEFAULT_SPLIT_INTERVAL,
         style_text="",
-        style_weight=0.7,
+        style_weight=DEFAULT_STYLE_WEIGHT,
         use_style_text=False,
         style="0",
-        emotion_weight=1.0,
+        emotion_weight=DEFAULT_EMOTION_WEIGHT,
     ):
         if reference_audio_path == "":
             reference_audio_path = None
@@ -149,8 +167,8 @@ class ModelHolder:
         self.refresh()
 
     def refresh(self):
-        self.model_files_dict = {}
-        self.model_names = []
+        self.model_files_dict: Dict[str, List[str]] = {}
+        self.model_names: List[str] = []
         self.current_model = None
         model_dirs = [
             d
@@ -168,6 +186,7 @@ class ModelHolder:
                 logger.info(
                     f"No model files found in {self.root_dir}/{model_name}, so skip it"
                 )
+                continue
             self.model_files_dict[model_name] = model_files
             self.model_names.append(model_name)
 
@@ -373,8 +392,6 @@ if __name__ == "__main__":
 
     model_holder = ModelHolder(model_dir, device)
 
-    languages = ["JP", "EN", "ZH"]
-
     model_names = model_holder.model_names
     if len(model_names) == 0:
         logger.error(f"モデルが見つかりませんでした。{model_dir}にモデルを置いてください。")
@@ -404,27 +421,43 @@ if __name__ == "__main__":
                     load_button = gr.Button("ロード", scale=1, variant="primary")
                 text_input = gr.TextArea(label="テキスト", value=initial_text)
 
-                line_split = gr.Checkbox(label="改行で分けて生成", value=True)
+                line_split = gr.Checkbox(label="改行で分けて生成", value=DEFAULT_LINE_SPLIT)
                 split_interval = gr.Slider(
                     minimum=0.0,
                     maximum=2,
-                    value=0.5,
+                    value=DEFAULT_SPLIT_INTERVAL,
                     step=0.1,
                     label="分けた場合に挟む無音の長さ（秒）",
                 )
                 language = gr.Dropdown(choices=languages, value="JP", label="Language")
                 with gr.Accordion(label="詳細設定", open=False):
                     sdp_ratio = gr.Slider(
-                        minimum=0, maximum=1, value=0.2, step=0.1, label="SDP Ratio"
+                        minimum=0,
+                        maximum=1,
+                        value=DEFAULT_SDP_RATIO,
+                        step=0.1,
+                        label="SDP Ratio",
                     )
                     noise_scale = gr.Slider(
-                        minimum=0.1, maximum=2, value=0.6, step=0.1, label="Noise"
+                        minimum=0.1,
+                        maximum=2,
+                        value=DEFAULT_NOISE,
+                        step=0.1,
+                        label="Noise",
                     )
                     noise_scale_w = gr.Slider(
-                        minimum=0.1, maximum=2, value=0.8, step=0.1, label="Noise_W"
+                        minimum=0.1,
+                        maximum=2,
+                        value=DEFAULT_NOISEW,
+                        step=0.1,
+                        label="Noise_W",
                     )
                     length_scale = gr.Slider(
-                        minimum=0.1, maximum=2, value=1.0, step=0.1, label="Length"
+                        minimum=0.1,
+                        maximum=2,
+                        value=DEFAULT_LENGTH,
+                        step=0.1,
+                        label="Length",
                     )
                     use_style_text = gr.Checkbox(label="Style textを使う", value=False)
                     style_text = gr.Textbox(
@@ -436,7 +469,7 @@ if __name__ == "__main__":
                     style_text_weight = gr.Slider(
                         minimum=0,
                         maximum=1,
-                        value=0.7,
+                        value=DEFAULT_STYLE_WEIGHT,
                         step=0.1,
                         label="Style textの強さ",
                         visible=False,
@@ -462,7 +495,7 @@ if __name__ == "__main__":
                 style_weight = gr.Slider(
                     minimum=0,
                     maximum=50,
-                    value=1,
+                    value=DEFAULT_EMOTION_WEIGHT,
                     step=0.1,
                     label="スタイルの強さ",
                 )

config.py

@@ -2,11 +2,13 @@
 @Desc: 全局配置文件读取
 """
 import argparse
-import yaml
-from typing import Dict, List
 import os
 import shutil
-import sys
+from typing import Dict, List
+
+import yaml
+
+from tools.log import logger
 
 
 class Resample_config:
@@ -172,11 +174,18 @@ class Webui_config:
 
 class Server_config:
     def __init__(
-        self, models: List[Dict[str, any]], port: int = 5000, device: str = "cuda"
+        self,
+        port: int = 5000,
+        device: str = "cuda",
+        limit: int = 100,
+        language: str = "JP",
+        origins: List[str] = None,
     ):
-        self.models: List[Dict[str, any]] = models  # 需要加载的所有模型的配置
-        self.port: int = port  # 端口号
-        self.device: str = device  # 模型默认使用设备
+        self.port: int = port
+        self.device: str = device
+        self.language: str = language
+        self.limit: int = limit
+        self.origins: List[str] = origins
 
     @classmethod
     def from_dict(cls, data: Dict[str, any]):
@@ -251,4 +260,10 @@ parser = argparse.ArgumentParser()
 # 为避免与以前的config.json起冲突，将其更名如下
 parser.add_argument("-y", "--yml_config", type=str, default="config.yml")
 args, _ = parser.parse_known_args()
+
+try:
     config = Config(args.yml_config)
+except TypeError:
+    logger.warning("Old config.yml found. Replace it with default_config.yml.")
+    shutil.copy(src="default_config.yml", dst="config.yml")
+    config = Config("config.yml")

default_config.yml

@@ -65,17 +65,10 @@ webui:
   language_identification_library: "langid"
 
 # server_fastapi's config
-# TODO: `server_fastapi.py` is not implemented yet for this version
 server:
   port: 5000
   device: "cuda"
-  models:
-    - model: ""
-      config: ""
-      device: "cuda"
-      language: "ZH"
-    - model: ""
-      config: ""
-      device: "cpu"
   language: "JP"
-      speakers: []
+  limit: 100
+  origins:
+    - "*"

server_fastapi.py

@@ -1,535 +1,210 @@
 """
-TODO: This file is not supported in this fork.
 api服务 多版本多模型 fastapi实现
 """
-import logging
-import gc
-import random
-
-import librosa
-import gradio
-import numpy as np
-import utils
-from fastapi import FastAPI, Query, Request, File, UploadFile, Form
+import argparse
+from fastapi import FastAPI, Query, Request, status, HTTPException
 from fastapi.responses import Response, FileResponse
-from fastapi.staticfiles import StaticFiles
+from fastapi.middleware.cors import CORSMiddleware
 from io import BytesIO
 from scipy.io import wavfile
 import uvicorn
 import torch
-import webbrowser
 import psutil
 import GPUtil
-from typing import Dict, Optional, List, Set, Union
-import os
+from typing import Dict, Optional, List, Union
+import os, sys
 from tools.log import logger
 from urllib.parse import unquote
-
-from infer import infer, get_net_g, latest_version
-import tools.translate as trans
-from re_matching import cut_sent
-
-
 from config import config
-
-os.environ["TOKENIZERS_PARALLELISM"] = "false"
-
-
-class Model:
-    """模型封装类"""
-
-    def __init__(self, config_path: str, model_path: str, device: str, language: str):
-        self.config_path: str = os.path.normpath(config_path)
-        self.model_path: str = os.path.normpath(model_path)
-        self.device: str = device
-        self.language: str = language
-        self.hps = utils.get_hparams_from_file(config_path)
-        self.spk2id: Dict[str, int] = self.hps.data.spk2id  # spk - id 映射字典
-        self.id2spk: Dict[int, str] = dict()  # id - spk 映射字典
-        for speaker, speaker_id in self.hps.data.spk2id.items():
-            self.id2spk[speaker_id] = speaker
-        self.version: str = (
-            self.hps.version if hasattr(self.hps, "version") else latest_version
+from app import (
+    Model,
+    ModelHolder,
+    Languages,
+    DEFAULT_SDP_RATIO,
+    DEFAULT_NOISE,
+    DEFAULT_NOISEW,
+    DEFAULT_LENGTH,
+    DEFAULT_LINE_SPLIT,
+    DEFAULT_SPLIT_INTERVAL,
+    DEFAULT_STYLE_WEIGHT,
+    DEFAULT_EMOTION_WEIGHT,
 )
-        self.net_g = get_net_g(
-            model_path=model_path,
-            version=self.version,
-            device=device,
-            hps=self.hps,
+from webui_style_vectors import DEFAULT_EMOTION
+
+ln = config.server_config.language
+
+
+def raise_validation_error(msg: str, param: str):
+    raise HTTPException(
+        status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+        detail=[dict(type="invalid_params", msg=msg, loc=["query", param])],
     )
 
-    def to_dict(self) -> Dict[str, any]:
-        return {
-            "config_path": self.config_path,
-            "model_path": self.model_path,
-            "device": self.device,
-            "language": self.language,
-            "spk2id": self.spk2id,
-            "id2spk": self.id2spk,
-            "version": self.version,
-        }
+
+class AudioResponse(Response):
+    media_type = "audio/wav"
 
 
-class Models:
-    def __init__(self):
-        self.models: Dict[int, Model] = dict()
-        self.num = 0
-        # spkInfo[角色名][模型id] = 角色id
-        self.spk_info: Dict[str, Dict[int, int]] = dict()
-        self.path2ids: Dict[str, Set[int]] = dict()  # 路径指向的model的id
-
-    def init_model(
-        self, config_path: str, model_path: str, device: str, language: str
-    ) -> int:
-        """
-        初始化并添加一个模型
-
-        :param config_path: 模型config.json路径
-        :param model_path: 模型路径
-        :param device: 模型推理使用设备
-        :param language: 模型推理默认语言
-        """
-        # 若文件不存在则不进行加载
-        if not os.path.isfile(model_path):
-            if model_path != "":
-                logger.warning(f"模型文件{model_path} 不存在，不进行初始化")
-            return self.num
-        if not os.path.isfile(config_path):
-            if config_path != "":
-                logger.warning(f"配置文件{config_path} 不存在，不进行初始化")
-            return self.num
-
-        # 若路径中的模型已存在，则不添加模型，若不存在，则进行初始化。
-        model_path = os.path.realpath(model_path)
-        if model_path not in self.path2ids.keys():
-            self.path2ids[model_path] = {self.num}
-            self.models[self.num] = Model(
-                config_path=config_path,
-                model_path=model_path,
-                device=device,
-                language=language,
+def load_models(model_holder: ModelHolder):
+    model_holder.models = []
+    for model_name, model_paths in model_holder.model_files_dict.items():
+        model = Model(
+            model_path=model_paths[0],
+            config_path=os.path.join(model_holder.root_dir, model_name, "config.json"),
+            style_vec_path=os.path.join(
+                model_holder.root_dir, model_name, "style_vectors.npy"
+            ),
+            device=model_holder.device,
         )
-            logger.success(f"添加模型{model_path}，使用配置文件{os.path.realpath(config_path)}")
-        else:
-            # 获取一个指向id
-            m_id = next(iter(self.path2ids[model_path]))
-            self.models[self.num] = self.models[m_id]
-            self.path2ids[model_path].add(self.num)
-            logger.success("模型已存在，添加模型引用。")
-        # 添加角色信息
-        for speaker, speaker_id in self.models[self.num].spk2id.items():
-            if speaker not in self.spk_info.keys():
-                self.spk_info[speaker] = {self.num: speaker_id}
-            else:
-                self.spk_info[speaker][self.num] = speaker_id
-        # 修改计数
-        self.num += 1
-        return self.num - 1
-
-    def del_model(self, index: int) -> Optional[int]:
-        """删除对应序号的模型，若不存在则返回None"""
-        if index not in self.models.keys():
-            return None
-        # 删除角色信息
-        for speaker, speaker_id in self.models[index].spk2id.items():
-            self.spk_info[speaker].pop(index)
-            if len(self.spk_info[speaker]) == 0:
-                # 若对应角色的所有模型都被删除，则清除该角色信息
-                self.spk_info.pop(speaker)
-        # 删除路径信息
-        model_path = os.path.realpath(self.models[index].model_path)
-        self.path2ids[model_path].remove(index)
-        if len(self.path2ids[model_path]) == 0:
-            self.path2ids.pop(model_path)
-            logger.success(f"删除模型{model_path}, id = {index}")
-        else:
-            logger.success(f"删除模型引用{model_path}, id = {index}")
-        # 删除模型
-        self.models.pop(index)
-        gc.collect()
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-        return index
-
-    def get_models(self):
-        """获取所有模型"""
-        return self.models
+        model.load_net_g()
+        model_holder.models.append(model)
 
 
 if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--cpu", action="store_true", help="Use CPU instead of GPU")
+    parser.add_argument(
+        "--dir", "-d", type=str, help="Model directory", default=config.out_dir
+    )
+    args = parser.parse_args()
+
+    if args.cpu:
+        device = "cpu"
+    else:
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+
+    model_dir = args.dir
+    model_holder = ModelHolder(model_dir, device)
+    if len(model_holder.model_names) == 0:
+        logger.error(f"Models not found in {model_dir}.")
+        sys.exit(1)
+
+    logger.info("Loading models...")
+    load_models(model_holder)
+    limit = config.server_config.limit
     app = FastAPI()
-    app.logger = logger
-    # 挂载静态文件
-    logger.info("开始挂载网页页面")
-    StaticDir: str = "./Web"
-    if not os.path.isdir(StaticDir):
+    allow_origins = config.server_config.origins
+    if allow_origins:
         logger.warning(
-            "缺少网页资源，无法开启网页页面，如有需要请在 https://github.com/jiangyuxiaoxiao/Bert-VITS2-UI 或者Bert-VITS对应版本的release页面下载"
+            f"CORS allow_origins={config.server_config.origins}. If you don't want, modify config.yml"
         )
-    else:
-        dirs = [fir.name for fir in os.scandir(StaticDir) if fir.is_dir()]
-        files = [fir.name for fir in os.scandir(StaticDir) if fir.is_dir()]
-        for dirName in dirs:
-            app.mount(
-                f"/{dirName}",
-                StaticFiles(directory=f"./{StaticDir}/{dirName}"),
-                name=dirName,
-            )
-    loaded_models = Models()
-    # 加载模型
-    logger.info("开始加载模型")
-    models_info = config.server_config.models
-    for model_info in models_info:
-        loaded_models.init_model(
-            config_path=model_info["config"],
-            model_path=model_info["model"],
-            device=model_info["device"],
-            language=model_info["language"],
+        app.add_middleware(
+            CORSMiddleware,
+            allow_origins=config.server_config.origins,
+            allow_credentials=True,
+            allow_methods=["*"],
+            allow_headers=["*"],
         )
+    app.logger = logger
 
-    @app.get("/")
-    async def index():
-        return FileResponse("./Web/index.html")
-
-    async def _voice(
-        text: str,
-        model_id: int,
-        speaker_name: str,
-        speaker_id: int,
-        sdp_ratio: float,
-        noise: float,
-        noisew: float,
-        length: float,
-        language: str,
-        auto_translate: bool,
-        auto_split: bool,
-        emotion: Optional[Union[int, str]] = None,
-        reference_audio=None,
-        style_text: Optional[str] = None,
-        style_weight: float = 0.7,
-    ) -> Union[Response, Dict[str, any]]:
-        """TTS实现函数"""
-        # 检查模型是否存在
-        if model_id not in loaded_models.models.keys():
-            logger.error(f"/voice 请求错误：模型model_id={model_id}未加载")
-            return {"status": 10, "detail": f"模型model_id={model_id}未加载"}
-        # 检查是否提供speaker
-        if speaker_name is None and speaker_id is None:
-            logger.error("/voice 请求错误：推理请求未提供speaker_name或speaker_id")
-            return {"status": 11, "detail": "请提供speaker_name或speaker_id"}
-        elif speaker_name is None:
-            # 检查speaker_id是否存在
-            if speaker_id not in loaded_models.models[model_id].id2spk.keys():
-                logger.error(f"/voice 请求错误：角色speaker_id={speaker_id}不存在")
-                return {"status": 12, "detail": f"角色speaker_id={speaker_id}不存在"}
-            speaker_name = loaded_models.models[model_id].id2spk[speaker_id]
-        # 检查speaker_name是否存在
-        if speaker_name not in loaded_models.models[model_id].spk2id.keys():
-            logger.error(f"/voice 请求错误：角色speaker_name={speaker_name}不存在")
-            return {"status": 13, "detail": f"角色speaker_name={speaker_name}不存在"}
-        # 未传入则使用默认语言
-        if language is None:
-            language = loaded_models.models[model_id].language
-        # 翻译会破坏mix结构，auto也会变得无意义。不要在这两个模式下使用
-        if auto_translate:
-            if language == "auto" or language == "mix":
-                logger.error(
-                    f"/voice 请求错误：请勿同时使用language = {language}与auto_translate模式"
-                )
-                return {
-                    "status": 20,
-                    "detail": f"请勿同时使用language = {language}与auto_translate模式",
-                }
-            text = trans.translate(Sentence=text, to_Language=language.lower())
-        if reference_audio is not None:
-            ref_audio = BytesIO(await reference_audio.read())
-            # 2.2 适配
-            if loaded_models.models[model_id].version == "2.2":
-                ref_audio, _ = librosa.load(ref_audio, 48000)
-
-        else:
-            ref_audio = reference_audio
-        if not auto_split:
-            with torch.no_grad():
-                audio = infer(
-                    text=text,
-                    sdp_ratio=sdp_ratio,
-                    noise_scale=noise,
-                    noise_scale_w=noisew,
-                    length_scale=length,
-                    sid=speaker_name,
-                    language=language,
-                    hps=loaded_models.models[model_id].hps,
-                    net_g=loaded_models.models[model_id].net_g,
-                    device=loaded_models.models[model_id].device,
-                    emotion=emotion,
-                    reference_audio=ref_audio,
-                    style_text=style_text,
-                    style_weight=style_weight,
-                )
-                audio = gradio.processing_utils.convert_to_16_bit_wav(audio)
-        else:
-            texts = cut_sent(text)
-            audios = []
-            with torch.no_grad():
-                for t in texts:
-                    audios.append(
-                        infer(
-                            text=t,
-                            sdp_ratio=sdp_ratio,
-                            noise_scale=noise,
-                            noise_scale_w=noisew,
-                            length_scale=length,
-                            sid=speaker_name,
-                            language=language,
-                            hps=loaded_models.models[model_id].hps,
-                            net_g=loaded_models.models[model_id].net_g,
-                            device=loaded_models.models[model_id].device,
-                            emotion=emotion,
-                            reference_audio=ref_audio,
-                            style_text=style_text,
-                            style_weight=style_weight,
-                        )
-                    )
-                    audios.append(np.zeros(int(44100 * 0.2)))
-                audio = np.concatenate(audios)
-                audio = gradio.processing_utils.convert_to_16_bit_wav(audio)
-        with BytesIO() as wavContent:
-            wavfile.write(
-                wavContent, loaded_models.models[model_id].hps.data.sampling_rate, audio
-            )
-            response = Response(content=wavContent.getvalue(), media_type="audio/wav")
-            return response
-
-    @app.post("/voice")
+    @app.get("/voice", response_class=AudioResponse)
     async def voice(
-        request: Request,  # fastapi自动注入
-        text: str = Form(...),
-        model_id: int = Query(..., description="模型ID"),  # 模型序号
+        request: Request,
+        text: str = Query(..., min_length=1, max_length=limit, description=f"セリフ"),
+        encoding: str = Query(None, description="textをURLデコードする(ex, `utf-8`)"),
+        model_id: int = Query(0, description="モデルID。`GET /models/info`のkeyの値を指定ください"),
         speaker_name: str = Query(
-            None, description="说话人名"
-        ),  # speaker_name与 speaker_id二者选其一
-        speaker_id: int = Query(None, description="说话人id，与speaker_name二选一"),
-        sdp_ratio: float = Query(0.2, description="SDP/DP混合比"),
-        noise: float = Query(0.2, description="感情"),
-        noisew: float = Query(0.9, description="音素长度"),
-        length: float = Query(1, description="语速"),
-        language: str = Query(None, description="语言"),  # 若不指定使用语言则使用默认值
-        auto_translate: bool = Query(False, description="自动翻译"),
-        auto_split: bool = Query(False, description="自动切分"),
-        emotion: Optional[Union[int, str]] = Query(None, description="emo"),
-        reference_audio: UploadFile = File(None),
-        style_text: Optional[str] = Form(None, description="风格文本"),
-        style_weight: float = Query(0.7, description="风格权重"),
+            None, description="話者名(speaker_idより優先)。esd.listの2列目の文字列を指定"
+        ),
+        speaker_id: int = Query(
+            0, description="話者ID。model_assets>[model]>config.json内のspk2idを確認"
+        ),
+        sdp_ratio: float = Query(
+            DEFAULT_SDP_RATIO,
+            description="SDP(Stochastic Duration Predictor)/DP混合比。比率が高くなるほどトーンのばらつきが大きくなる",
+        ),
+        noise: float = Query(DEFAULT_NOISE, description="サンプルノイズの割合。大きくするほどランダム性が高まる"),
+        noisew: float = Query(
+            DEFAULT_NOISEW, description="SDPノイズ。大きくするほど発音の間隔にばらつきが出やすくなる"
+        ),
+        length: float = Query(
+            DEFAULT_LENGTH, description="話速。基準は1で大きくするほど音声は長くなり読み上げが遅まる"
+        ),
+        language: Languages = Query(ln, description=f"textの言語"),
+        auto_split: bool = Query(DEFAULT_LINE_SPLIT, description="改行で分けて生成"),
+        split_interval: float = Query(
+            DEFAULT_SPLIT_INTERVAL, description="分けた場合に挟む無音の長さ（秒）"
+        ),
+        style_text: Optional[str] = Query(
+            None, description="このテキストの読み上げと似た声音・感情になりやすくなる。ただし抑揚やテンポ等が犠牲になる傾向がある"
+        ),
+        style_weight: float = Query(DEFAULT_STYLE_WEIGHT, description="style_textの強さ"),
+        emotion: Optional[Union[int, str]] = Query(DEFAULT_EMOTION, description="スタイル"),
+        emotion_weight: float = Query(DEFAULT_EMOTION_WEIGHT, description="emotionの強さ"),
+        reference_audio_path: Optional[str] = Query(
+            None, description="emotionを音声ファイルで行う"
+        ),
     ):
-        """语音接口，若需要上传参考音频请仅使用post请求"""
-        logger.info(
-            f"{request.client.host}:{request.client.port}/voice  { unquote(str(request.query_params) )} text={text}"
-        )
-        return await _voice(
-            text=text,
-            model_id=model_id,
-            speaker_name=speaker_name,
-            speaker_id=speaker_id,
-            sdp_ratio=sdp_ratio,
-            noise=noise,
-            noisew=noisew,
-            length=length,
-            language=language,
-            auto_translate=auto_translate,
-            auto_split=auto_split,
-            emotion=emotion,
-            reference_audio=reference_audio,
-            style_text=style_text,
-            style_weight=style_weight,
-        )
-
-    @app.get("/voice")
-    async def voice(
-        request: Request,  # fastapi自动注入
-        text: str = Query(..., description="输入文字"),
-        model_id: int = Query(..., description="模型ID"),  # 模型序号
-        speaker_name: str = Query(
-            None, description="说话人名"
-        ),  # speaker_name与 speaker_id二者选其一
-        speaker_id: int = Query(None, description="说话人id，与speaker_name二选一"),
-        sdp_ratio: float = Query(0.2, description="SDP/DP混合比"),
-        noise: float = Query(0.2, description="感情"),
-        noisew: float = Query(0.9, description="音素长度"),
-        length: float = Query(1, description="语速"),
-        language: str = Query(None, description="语言"),  # 若不指定使用语言则使用默认值
-        auto_translate: bool = Query(False, description="自动翻译"),
-        auto_split: bool = Query(False, description="自动切分"),
-        emotion: Optional[Union[int, str]] = Query(None, description="emo"),
-        style_text: Optional[str] = Query(None, description="风格文本"),
-        style_weight: float = Query(0.7, description="风格权重"),
-    ):
-        """语音接口"""
+        """Infer text to speech(テキストから感情付き音声を生成する)"""
         logger.info(
             f"{request.client.host}:{request.client.port}/voice  { unquote(str(request.query_params) )}"
         )
-        return await _voice(
+        if model_id >= len(model_holder.models):  # /models/refresh があるためQuery(le)で表現不可
+            raise_validation_error(f"model_id={model_id} not found", "model_id")
+
+        model = model_holder.models[model_id]
+        if speaker_name is None:
+            if speaker_id not in model.id2spk.keys():
+                raise_validation_error(
+                    f"speaker_id={speaker_id} not found", "speaker_id"
+                )
+        else:
+            if speaker_name not in model.spk2id.keys():
+                raise_validation_error(
+                    f"speaker_name={speaker_name} not found", "speaker_name"
+                )
+            speaker_id = model.spk2id[speaker_name]
+        if emotion not in model.style2id.keys():
+            raise_validation_error(f"emotion={emotion} not found", "emotion")
+        if encoding is not None:
+            text = unquote(text, encoding=encoding)
+        sr, audio = model.infer(
             text=text,
-            model_id=model_id,
-            speaker_name=speaker_name,
-            speaker_id=speaker_id,
+            language=language,
+            sid=speaker_id,
+            reference_audio_path=reference_audio_path,
             sdp_ratio=sdp_ratio,
             noise=noise,
             noisew=noisew,
             length=length,
-            language=language,
-            auto_translate=auto_translate,
-            auto_split=auto_split,
-            emotion=emotion,
+            line_split=auto_split,
+            split_interval=split_interval,
             style_text=style_text,
             style_weight=style_weight,
+            use_style_text=bool(style_text),
+            style=emotion,
+            emotion_weight=emotion_weight,
         )
+        with BytesIO() as wavContent:
+            wavfile.write(wavContent, sr, audio)
+            return Response(content=wavContent.getvalue(), media_type="audio/wav")
 
     @app.get("/models/info")
-    def get_loaded_models_info(request: Request):
-        """获取已加载模型信息"""
+    def get_loaded_models_info():
+        """ロードされたモデル情報の取得"""
 
         result: Dict[str, Dict] = dict()
-        for key, model in loaded_models.models.items():
-            result[str(key)] = model.to_dict()
+        for model_id, model in enumerate(model_holder.models):
+            result[str(model_id)] = {
+                "config_path": model.config_path,
+                "model_path": model.model_path,
+                "device": model.device,
+                "spk2id": model.spk2id,
+                "id2spk": model.id2spk,
+                "style2id": model.style2id,
+            }
         return result
 
-    @app.get("/models/delete")
-    def delete_model(
-        request: Request, model_id: int = Query(..., description="删除模型id")
-    ):
-        """删除指定模型"""
-        logger.info(
-            f"{request.client.host}:{request.client.port}/models/delete  { unquote(str(request.query_params) )}"
-        )
-        result = loaded_models.del_model(model_id)
-        if result is None:
-            logger.error(f"/models/delete 模型删除错误：模型{model_id}不存在，删除失败")
-            return {"status": 14, "detail": f"模型{model_id}不存在，删除失败"}
-
-        return {"status": 0, "detail": "删除成功"}
-
-    @app.get("/models/add")
-    def add_model(
-        request: Request,
-        model_path: str = Query(..., description="添加模型路径"),
-        config_path: str = Query(
-            None, description="添加模型配置文件路径，不填则使用./config.json或../config.json"
-        ),
-        device: str = Query("cuda", description="推理使用设备"),
-        language: str = Query("ZH", description="模型默认语言"),
-    ):
-        """添加指定模型：允许重复添加相同路径模型，且不重复占用内存"""
-        logger.info(
-            f"{request.client.host}:{request.client.port}/models/add  { unquote(str(request.query_params) )}"
-        )
-        if config_path is None:
-            model_dir = os.path.dirname(model_path)
-            if os.path.isfile(os.path.join(model_dir, "config.json")):
-                config_path = os.path.join(model_dir, "config.json")
-            elif os.path.isfile(os.path.join(model_dir, "../config.json")):
-                config_path = os.path.join(model_dir, "../config.json")
-            else:
-                logger.error("/models/add 模型添加失败：未在模型所在目录以及上级目录找到config.json文件")
-                return {
-                    "status": 15,
-                    "detail": "查询未传入配置文件路径，同时默认路径./与../中不存在配置文件config.json。",
-                }
-        try:
-            model_id = loaded_models.init_model(
-                config_path=config_path,
-                model_path=model_path,
-                device=device,
-                language=language,
-            )
-        except Exception:
-            logging.exception("模型加载出错")
-            return {
-                "status": 16,
-                "detail": "模型加载出错，详细查看日志",
-            }
-        return {
-            "status": 0,
-            "detail": "模型添加成功",
-            "Data": {
-                "model_id": model_id,
-                "model_info": loaded_models.models[model_id].to_dict(),
-            },
-        }
-
-    def _get_all_models(root_dir: str = "Data", only_unloaded: bool = False):
-        """从root_dir搜索获取所有可用模型"""
-        result: Dict[str, List[str]] = dict()
-        files = os.listdir(root_dir) + ["."]
-        for file in files:
-            if os.path.isdir(os.path.join(root_dir, file)):
-                sub_dir = os.path.join(root_dir, file)
-                # 搜索 "sub_dir" 、 "sub_dir/models" 两个路径
-                result[file] = list()
-                sub_files = os.listdir(sub_dir)
-                model_files = []
-                for sub_file in sub_files:
-                    relpath = os.path.realpath(os.path.join(sub_dir, sub_file))
-                    if only_unloaded and relpath in loaded_models.path2ids.keys():
-                        continue
-                    if sub_file.endswith(".pth") and sub_file.startswith("G_"):
-                        if os.path.isfile(relpath):
-                            model_files.append(sub_file)
-                # 对模型文件按步数排序
-                model_files = sorted(
-                    model_files,
-                    key=lambda pth: int(pth.lstrip("G_").rstrip(".pth"))
-                    if pth.lstrip("G_").rstrip(".pth").isdigit()
-                    else 10**10,
-                )
-                result[file] = model_files
-                models_dir = os.path.join(sub_dir, "models")
-                model_files = []
-                if os.path.isdir(models_dir):
-                    sub_files = os.listdir(models_dir)
-                    for sub_file in sub_files:
-                        relpath = os.path.realpath(os.path.join(models_dir, sub_file))
-                        if only_unloaded and relpath in loaded_models.path2ids.keys():
-                            continue
-                        if sub_file.endswith(".pth") and sub_file.startswith("G_"):
-                            if os.path.isfile(os.path.join(models_dir, sub_file)):
-                                model_files.append(f"models/{sub_file}")
-                    # 对模型文件按步数排序
-                    model_files = sorted(
-                        model_files,
-                        key=lambda pth: int(pth.lstrip("models/G_").rstrip(".pth"))
-                        if pth.lstrip("models/G_").rstrip(".pth").isdigit()
-                        else 10**10,
-                    )
-                    result[file] += model_files
-                if len(result[file]) == 0:
-                    result.pop(file)
-
-        return result
-
-    @app.get("/models/get_unloaded")
-    def get_unloaded_models_info(
-        request: Request, root_dir: str = Query("Data", description="搜索根目录")
-    ):
-        """获取未加载模型"""
-        logger.info(
-            f"{request.client.host}:{request.client.port}/models/get_unloaded  { unquote(str(request.query_params) )}"
-        )
-        return _get_all_models(root_dir, only_unloaded=True)
-
-    @app.get("/models/get_local")
-    def get_local_models_info(
-        request: Request, root_dir: str = Query("Data", description="搜索根目录")
-    ):
-        """获取全部本地模型"""
-        logger.info(
-            f"{request.client.host}:{request.client.port}/models/get_local  { unquote(str(request.query_params) )}"
-        )
-        return _get_all_models(root_dir, only_unloaded=False)
+    @app.post("/models/refresh")
+    def refresh():
+        """モデルをパスに追加/削除した際などに読み込ませる"""
+        model_holder.refresh()
+        load_models(model_holder)
+        return get_loaded_models_info()
 
     @app.get("/status")
     def get_status():
-        """获取电脑运行状态"""
+        """実行環境のステータスを取得"""
         cpu_percent = psutil.cpu_percent(interval=1)
         memory_info = psutil.virtual_memory()
         memory_total = memory_info.total
@@ -563,119 +238,22 @@ if __name__ == "__main__":
             "gpu": gpuInfo,
         }
 
-    @app.get("/tools/translate")
-    def translate(
-        request: Request,
-        texts: str = Query(..., description="待翻译文本"),
-        to_language: str = Query(..., description="翻译目标语言"),
+    @app.get("/tools/get_audio", response_class=AudioResponse)
+    def get_audio(
+        request: Request, path: str = Query(..., description="local wav path")
     ):
-        """翻译"""
-        logger.info(
-            f"{request.client.host}:{request.client.port}/tools/translate  { unquote(str(request.query_params) )}"
-        )
-        return {"texts": trans.translate(Sentence=texts, to_Language=to_language)}
-
-    all_examples: Dict[str, Dict[str, List]] = dict()  # 存放示例
-
-    @app.get("/tools/random_example")
-    def random_example(
-        request: Request,
-        language: str = Query(None, description="指定语言，未指定则随机返回"),
-        root_dir: str = Query("Data", description="搜索根目录"),
-    ):
-        """
-        获取一个随机音频+文本，用于对比，音频会从本地目录随机选择。
-        """
-        logger.info(
-            f"{request.client.host}:{request.client.port}/tools/random_example  { unquote(str(request.query_params) )}"
-        )
-        global all_examples
-        # 数据初始化
-        if root_dir not in all_examples.keys():
-            all_examples[root_dir] = {"ZH": [], "JP": [], "EN": []}
-
-            examples = all_examples[root_dir]
-
-            # 从项目Data目录中搜索train/val.list
-            for root, directories, _files in os.walk(root_dir):
-                for file in _files:
-                    if file in ["train.list", "val.list"]:
-                        with open(
-                            os.path.join(root, file), mode="r", encoding="utf-8"
-                        ) as f:
-                            lines = f.readlines()
-                            for line in lines:
-                                data = line.split("|")
-                                if len(data) != 7:
-                                    continue
-                                # 音频存在 且语言为ZH/EN/JP
-                                if os.path.isfile(data[0]) and data[2] in [
-                                    "ZH",
-                                    "JP",
-                                    "EN",
-                                ]:
-                                    examples[data[2]].append(
-                                        {
-                                            "text": data[3],
-                                            "audio": data[0],
-                                            "speaker": data[1],
-                                        }
-                                    )
-
-        examples = all_examples[root_dir]
-        if language is None:
-            if len(examples["ZH"]) + len(examples["JP"]) + len(examples["EN"]) == 0:
-                return {"status": 17, "detail": "没有加载任何示例数据"}
-            else:
-                # 随机选一个
-                rand_num = random.randint(
-                    0,
-                    len(examples["ZH"]) + len(examples["JP"]) + len(examples["EN"]) - 1,
-                )
-                # ZH
-                if rand_num < len(examples["ZH"]):
-                    return {"status": 0, "Data": examples["ZH"][rand_num]}
-                # JP
-                if rand_num < len(examples["ZH"]) + len(examples["JP"]):
-                    return {
-                        "status": 0,
-                        "Data": examples["JP"][rand_num - len(examples["ZH"])],
-                    }
-                # EN
-                return {
-                    "status": 0,
-                    "Data": examples["EN"][
-                        rand_num - len(examples["ZH"]) - len(examples["JP"])
-                    ],
-                }
-
-        else:
-            if len(examples[language]) == 0:
-                return {"status": 17, "detail": f"没有加载任何{language}数据"}
-            return {
-                "status": 0,
-                "Data": examples[language][
-                    random.randint(0, len(examples[language]) - 1)
-                ],
-            }
-
-    @app.get("/tools/get_audio")
-    def get_audio(request: Request, path: str = Query(..., description="本地音频路径")):
+        """wavデータを取得する"""
         logger.info(
             f"{request.client.host}:{request.client.port}/tools/get_audio  { unquote(str(request.query_params) )}"
         )
         if not os.path.isfile(path):
-            logger.error(f"/tools/get_audio 获取音频错误：指定音频{path}不存在")
-            return {"status": 18, "detail": "指定音频不存在"}
+            raise_validation_error(f"path={path} not found", "path")
         if not path.lower().endswith(".wav"):
-            logger.error(f"/tools/get_audio 获取音频错误：音频{path}非wav文件")
-            return {"status": 19, "detail": "非wav格式文件"}
-        return FileResponse(path=path)
+            raise_validation_error(f"wav file not found in {path}", "path")
+        return FileResponse(path=path, media_type="audio/wav")
 
-    logger.warning("本地服务，请勿将服务端口暴露于外网")
-    logger.info(f"api文档地址 http://127.0.0.1:{config.server_config.port}/docs")
-    if os.path.isdir(StaticDir):
-        webbrowser.open(f"http://127.0.0.1:{config.server_config.port}")
+    logger.info(f"server listen: http://127.0.0.1:{config.server_config.port}")
+    logger.info(f"API docs: http://127.0.0.1:{config.server_config.port}/docs")
     uvicorn.run(
         app, port=config.server_config.port, host="0.0.0.0", log_level="warning"
     )

webui_style_vectors.py

@@ -9,6 +9,7 @@ from sklearn.manifold import TSNE
 from config import config
 
 MAX_CLUSTER_NUM = 10
+DEFAULT_EMOTION: str = "Neutral"
 
 tsne = TSNE(n_components=2, random_state=42, metric="cosine")
 
@@ -124,7 +125,7 @@ def save_style_vectors(model_name, style_names: str):
     config_path = os.path.join(result_dir, "config.json")
     if not os.path.exists(config_path):
         return f"{config_path}が存在しません。"
-    style_name_list = ["Neutral"]
+    style_name_list = [DEFAULT_EMOTION]
     style_name_list = style_name_list + style_names.split(",")
     if len(style_name_list) != len(centroids) + 1:
         return f"スタイルの数が合いません。`,`で正しく{len(centroids)}個に区切られているか確認してください: {style_names}"
@@ -173,7 +174,7 @@ def save_style_vectors_from_files(model_name, audio_files_text, style_names_text
     config_path = os.path.join(result_dir, "config.json")
     if not os.path.exists(config_path):
         return f"{config_path}が存在しません。"
-    style_name_list = ["Neutral"]
+    style_name_list = [DEFAULT_EMOTION]
     style_name_list = style_name_list + style_names
     assert len(style_name_list) == len(style_vectors)
 
@@ -188,7 +189,7 @@ def save_style_vectors_from_files(model_name, audio_files_text, style_names_text
     return f"成功!\n{style_vector_path}に保存し{config_path}を更新しました。"
 
 
-initial_md = """
+initial_md = f"""
 # Style Bert-VITS2 スタイルベクトルの作成
 
 Style-Bert-VITS2でこまかくスタイルを指定して音声合成するには、モデルごとにスタイルベクトルのファイル`style_vectors.npy`を手動で作成する必要があります。
@@ -216,7 +217,7 @@ method1 = """
 
 詳細: スタイルベクトル(256次元)たちを適当なアルゴリズムでクラスタリングして、各クラスタの中心のベクトル（と全体の平均ベクトル）を保存します。
 
-平均スタイル（Neutral）は自動的に保存されます。
+平均スタイル（{DEFAULT_EMOTION}）は自動的に保存されます。
 """
 
 with gr.Blocks(theme="NoCrypt/miku") as app:
@@ -274,7 +275,7 @@ with gr.Blocks(theme="NoCrypt/miku") as app:
         style_names = gr.Textbox(
             "Angry, Sad, Happy",
             label="スタイルの名前",
-            info="スタイルの名前を`,`で区切って入力してください（日本語可）。例: `Angry, Sad, Happy`や`怒り, 悲しみ, 喜び`など。平均音声はNeutralとして自動的に保存されます。",
+            info=f"スタイルの名前を`,`で区切って入力してください（日本語可）。例: `Angry, Sad, Happy`や`怒り, 悲しみ, 喜び`など。平均音声は{DEFAULT_EMOTION}として自動的に保存されます。",
         )
         with gr.Row():
             save_button = gr.Button("スタイルベクトルを保存", variant="primary")
@@ -286,7 +287,9 @@ with gr.Blocks(theme="NoCrypt/miku") as app:
     with gr.Tab("方法2: 手動でスタイルを選ぶ"):
         gr.Markdown("下のテキスト欄に、各スタイルの代表音声のファイル名を`,`区切りで、その横に対応するスタイル名を`,`区切りで入力してください。")
         gr.Markdown("例: `angry.wav, sad.wav, happy.wav`と`Angry, Sad, Happy`")
-        gr.Markdown("注意: Neutralスタイルは自動的に保存されます、手動ではNeutralという名前のスタイルは指定しないでください。")
+        gr.Markdown(
+            f"注意: {DEFAULT_EMOTION}スタイルは自動的に保存されます、手動では{DEFAULT_EMOTION}という名前のスタイルは指定しないでください。"
+        )
         with gr.Row():
             audio_files_text = gr.Textbox(
                 label="音声ファイル名", placeholder="angry.wav, sad.wav, happy.wav"
@@ -307,5 +310,4 @@ with gr.Blocks(theme="NoCrypt/miku") as app:
             "`clustering.ipynb`にjvnvコーパスの場合の作り方とかクラスタ分けのいろいろを書いています。これを参考に自分で頑張って作ってください。"
         )
 
-
     app.launch(inbrowser=True)