无法加载bert vits2模型？

[drawgamer]

2025-10-05 05:05:20 [INFO] initializing identifier [in langid.load_model:162]
R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\py310\lib\site-packages\onnxru
ntime\capi\onnxruntime_validation.py:26: UserWarning: Unsupported Windows versio
n (7). ONNX Runtime supports Windows 10 and above, only.
warnings.warn(
2025-10-05 05:05:28 [INFO] Loading failed. 'NoneType' object has no attribute 'l
ower' [in ModelManager.load_model:357]
2025-10-05 05:05:28 [ERROR] Traceback (most recent call last):
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\manager\ModelManager.p
y", line 338, in load_model
model_data = self._load_model_from_path(tts_model)
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\manager\ModelManager.p
y", line 253, in load_model_from_path
model = model_class(**model_args)
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\bert_vits2\bert_vits2.
py", line 39, in init
self.version = process_legacy_versions(self.hps_ms).lower().replace("-", ""
)
AttributeError: 'NoneType' object has no attribute 'lower'
[in ModelManager.load_model:358]

我把pth文件和config.json文件和pytorch_model.bin放在一个文件夹里，然后文件夹放在data\models\下报错？但是把这个文件夹放在data\bert\bert-base-japanese-v3也没有加载，到底应该放哪里？
说明文档里没有举例bert vits2模型具体怎么放，只是说放到data\bert\对应位置，我搞不明白这三个文件到底放哪里？但是vits模型我倒是加载成功了。

[drawgamer]

@Artrajz 现在我在配置文件里加了"version": "1.0.1",就成功加载了，但明明是日语模型却加载成中文，我再在配置文件里设置为 "lang": ["ja"]却提示不能获取音频？

[drawgamer]

现在设置 "lang": ["zh","ja"],发现中文可以生成但是鸟语，而日语提示不能获取音频数据并且加载的时候就提示：
2025-10-05 06:43:24 [INFO] Loading BERT model: R:\MyDownload\vits-simple-api-win
dows-cpu-v0.6.16\data\bert/chinese-roberta-wwm-ext-large [in model_handler.load_
bert:169]
2025-10-05 06:43:31 [INFO] Success loading: R:\MyDownload\vits-simple-api-window
s-cpu-v0.6.16\data\bert/chinese-roberta-wwm-ext-large [in model_handler.load_ber
t:180]
2025-10-05 06:43:40 [ERROR] enc_p.emb.weight is not in the checkpoint [in utils.
load_checkpoint:59]
2025-10-05 06:43:40 [INFO] Loaded checkpoint 'R:\MyDownload\vits-simple-api-wind
ows-cpu-v0.6.16\data\models\kafka\kafka.pth' (iteration 286) [in utils.load_chec
kpoint:67]
2025-10-05 06:43:40 [INFO] model_type:BERT-VITS2 model_id:0 n_speakers:1 vits_pa
th:R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\data\models\kafka\kafka.pth
[in ModelManager._load_model_from_path:331]

[drawgamer]

生成日语时提示：
2025-10-05 07:02:24 [INFO] [BERT-VITS2] text_prompt:Happy [in views.voice_bert_v
its2_api:511]
2025-10-05 07:02:24 [ERROR] Exception on /voice/bert-vits2 [POST] [in app.log_ex
ception:1744]
Traceback (most recent call last):
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\py310\lib\site-package
s\flask\app.py", line 2528, in wsgi_app
response = self.full_dispatch_request()
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\py310\lib\site-package
s\flask\app.py", line 1825, in full_dispatch_request
rv = self.handle_user_exception(e)
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\py310\lib\site-package
s\flask_cors\extension.py", line 194, in wrapped_function
return cors_after_request(app.make_response(f(*args, **kwargs)))
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\py310\lib\site-package
s\flask\app.py", line 1823, in full_dispatch_request
rv = self.dispatch_request()
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\py310\lib\site-package
s\flask\app.py", line 1799, in dispatch_request
return self.ensure_sync(self.view_functions[rule.endpoint])(**view_args)
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\tts_app\voice_api\auth
.py", line 10, in check_api_key
return func(*args, **kwargs)
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\tts_app\voice_api\view
s.py", line 585, in voice_bert_vits2_api
audio = tts_manager.bert_vits2_infer(state)
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\manager\TTSManager.py"
, line 420, in bert_vits2_infer
audio = infer_func(**state)
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\bert_vits2\bert_vits2.
py", line 343, in infer
zh_bert, ja_bert, en_bert, phones, tones, lang_ids = self.get_text(text, lan
g, self.hps_ms, style_text,
File "R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\bert_vits2\bert_vits2.
py", line 232, in get_text
tokenizer, _ = self.model_handler.get_bert_model(self.bert_model_names[langu
age_str])
KeyError: 'ja'
2025-10-05 07:02:24 [INFO] 192.168.123.100 - - [05/Oct/2025 07:02:24] "POST /voi
ce/bert-vits2 HTTP/1.1" 500 - [in _internal._log:187]

[Artrajz]

data\models下放包含模型的文件夹

data\bert\bert-base-japanese-v3放的是bert-base-japanese-v3模型，正常检测到需要了会自动下载

"version": "1.0.1" 是只支持中文的bv2版本，你需要搞清楚你模型训练的版本，不然版本号错误也会有各种问题无法使用

[drawgamer]

data\models下放包含模型的文件夹

data\bert\bert-base-japanese-v3放的是bert-base-japanese-v3模型，正常检测到需要了会自动下载

"version": "1.0.1" 是只支持中文的bv2版本，你需要搞清楚你模型训练的版本，不然版本号错误也会有各种问题无法使用

原来如此，这个模型是 https://huggingface.co/spaces/soiz1/bert_vits2/tree/main 里的模型，但不知道要怎么查是什么版本？或者一共有哪些版本号？我一个一个尝试？

[drawgamer]

title: Bert Vits2 JP
emoji: 📊
colorFrom: red
colorTo: green
sdk: gradio
sdk_version: 3.47.1
app_file: app.py
pinned: false

是不是这个sdk_version: 3.47.1?

[Artrajz]

上面没看到有写版本号，只能靠猜了，你可以填这些版本号试一下"version": "1.1.0_transition" "version": "1.1.1"

[drawgamer]

上面没看到有写版本号，只能靠猜了，你可以填这些版本号试一下1.1.0_transition 1.1.1

2025-10-05 13:03:26 [INFO] Loading BERT model: R:\MyDownload\vits-simple-api-win
dows-cpu-v0.6.16\data\bert/deberta-v2-large-japanese-char-wwm [in model_handler.
load_bert:169]
2025-10-05 13:03:27 [INFO] Success loading: R:\MyDownload\vits-simple-api-window
s-cpu-v0.6.16\data\bert/deberta-v2-large-japanese-char-wwm [in model_handler.loa
d_bert:180]

我刚才试了下3.47.1虽然可以播放日语了，但是发音是机械化的棒读，发现加载的模型不是bert-base-japanese-v3但模型网站上是bert-base-japanese-v3的，哪个版本号对应bert-base-japanese-v3？

[drawgamer]

上面没看到有写版本号，只能靠猜了，你可以填这些版本号试一下"version": "1.1.0_transition" "version": "1.1.1"

加了 "version": "1.1.1",这个后可以正常播放bert-base-japanese-v3的语音了，但是日文汉字的仮名不对？
比如 弱い 正常是发 よわい 但是现在却念成了 じゃくい 了，而且是所有汉字都读错仮名！

[drawgamer]

@Artrajz 我把HF上的tokenizer_config.json文件放到bert-base-japanese-v3但是提示
2025-10-05 13:36:00 [INFO] Loading BERT model: R:\MyDownload\vits-simple-api-win
dows-cpu-v0.6.16\data\bert/bert-base-japanese-v3 [in model_handler.load_bert:169
]
2025-10-05 13:36:00 [ERROR] Failed loading R:\MyDownload\vits-simple-api-windows
-cpu-v0.6.16\data\bert/bert-base-japanese-v3. You need to install fugashi to use
MecabTokenizer. See https://pypi.org/project/fugashi/ for installation. [in mod
el_handler.load_bert:183]
2025-10-05 13:36:00 [INFO] Trying to download. [in model_handler.load_bert:184]
2025-10-05 13:36:01 [INFO] File already exists and verified successfully! [in mo
del_handler._download_model:161]
2025-10-05 13:36:01 [INFO] Loading BERT model: R:\MyDownload\vits-simple-api-win
dows-cpu-v0.6.16\data\bert/bert-base-japanese-v3 [in model_handler.load_bert:169
]
2025-10-05 13:36:01 [ERROR] Failed loading R:\MyDownload\vits-simple-api-windows
-cpu-v0.6.16\data\bert/bert-base-japanese-v3. You need to install fugashi to use
MecabTokenizer. See https://pypi.org/project/fugashi/ for installation. [in mod
el_handler.load_bert:183]
2025-10-05 13:36:01 [INFO] Trying to download. [in model_handler.load_bert:184]
2025-10-05 13:36:03 [INFO] File already exists and verified successfully! [in mo
del_handler._download_model:161]
2025-10-05 13:36:03 [INFO] Loading BERT model: R:\MyDownload\vits-simple-api-win
dows-cpu-v0.6.16\data\bert/bert-base-japanese-v3 [in model_handler.load_bert:169
]
2025-10-05 13:36:03 [ERROR] Failed loading R:\MyDownload\vits-simple-api-windows
-cpu-v0.6.16\data\bert/bert-base-japanese-v3. You need to install fugashi to use
MecabTokenizer. See https://pypi.org/project/fugashi/ for installation. [in mod
el_handler.load_bert:183]
2025-10-05 13:36:03 [INFO] Trying to download. [in model_handler.load_bert:184]
2025-10-05 13:36:04 [INFO] File already exists and verified successfully! [in mo
del_handler._download_model:161]
2025-10-05 13:36:04 [ERROR] Failed to load R:\MyDownload\vits-simple-api-windows
-cpu-v0.6.16\data\bert/bert-base-japanese-v3 after 3 retries. [in model_handler.
load_bert:191]

[drawgamer]

我试了把日文汉字都改成仮名，但是发音的音调不对，看来必须用日文汉字并且正确分词才能正常发音。

[drawgamer]

@Artrajz 现在是不是必须手动安装fugashi？但是我WIN7下不懂要怎么编译。
另外VITS的模型倒是可以正常分词读对日文汉字发音，就bert vits2模型不行。

[Artrajz]

在vits-simple-api-windows-cpu-v0.6.16文件夹打开一个cmd，然后运行命令py310\python -m pip install fugashi，然后重新启动看看行不行

[Artrajz]

bert-base-japanese-v3确实少了个tokenizer_config.json

[drawgamer]

在vits-simple-api-windows-cpu-v0.6.16文件夹打开一个cmd，然后运行命令py310\python -m pip install fugashi，然后重新启动看看行不行

我成功手动安装了！终于完美发音了，实在是太跌宕起伏了。

R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\py310\Scripts>pip install fuga
shi
Collecting fugashi
Downloading fugashi-1.5.1-cp310-cp310-win_amd64.whl.metadata (7.5 kB)
Downloading fugashi-1.5.1-cp310-cp310-win_amd64.whl (513 kB)
Installing collected packages: fugashi
WARNING: The scripts fugashi-build-dict.exe, fugashi-info.exe and fugashi.exe
are installed in 'R:\MyDownload\vits-simple-api-windows-cpu-v0.6.16\py310\Script
s' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warn
ing, use --no-warn-script-location.
Successfully installed fugashi-1.5.1