tsukumi
1a44c9d437
Merge branch 'dev' of github.com:litagin02/Style-Bert-VITS2
2025-03-26 20:09:40 +09:00
tsukumi
2aa503233a
Improve: Add option to generate AIVM/AIVMX files directly when running convert_onnx.py, add license information
2025-03-26 20:01:17 +09:00
tsukumi
810ca43615
Improve: Disable enable_cpu_mem_arena for CPU inference only to prevent excessive memory consumption during the inference session of the BERT model
2024-12-22 09:17:31 +09:00
tsukumi
405c5deddd
Improve: Adjusted the default Execution Provider options
2024-12-22 07:04:20 +09:00
tsukumi
08c439e88e
Refactor: Use I/O Binding during BERT inference and always release memory after inference
2024-12-22 06:49:43 +09:00
tsukumi
15af441cba
Improve: Convert ONNX version of BERT models to FP16
...
I have found that half-precision has little effect on speech synthesis quality and, depending on the environment, can reduce file size and memory usage by half, so I have decided to use FP16.
2024-12-19 14:51:30 +09:00
tsukumi
ebd249bca8
Fix: spm.model is missing
2024-12-19 13:16:25 +09:00
tsukumi
2833fb5eeb
Fix: The conversion script for English BERT to Fast Tokenizer was incorrect, so tokenization was not performed correctly
2024-12-19 11:10:33 +09:00
litagin02
8d476175a7
hathc run style:fmt
2024-11-17 12:44:27 +09:00
litagin02
a6bc65271c
hatch fmt
2024-11-17 12:41:05 +09:00
tsukumi
20b0b0908b
Fix: Ensure memory is cleared on unload
2024-11-10 07:06:20 +09:00
tsukumi
9723f75f99
Fix: g2p process fails in PyTorch-independent environment
2024-11-08 02:56:05 +09:00
tsukumi
7793e9da43
Remove: Unused import
2024-09-28 13:50:10 +09:00
tsukumi
c792bf9a0d
Improve: Make ONNX inference tests more rigorous
2024-09-23 12:27:29 +09:00
tsukumi
5ca6e184d7
Fix: ONNX inference of BERT models fails in some Windows environments
2024-09-23 11:33:17 +09:00
tsukumi
eb5d245903
Add: Test for ONNX inference code (without PyTorch dependency)
2024-09-23 10:04:02 +09:00
tsukumi
b7cd512978
Fix: During ONNX inference, the tensor of the return value of the tokenizer is now obtained in Numpy's NDArray format to eliminate the dependency on PyTorch
2024-09-23 07:40:17 +09:00
tsukumi
dcca37b9a8
Add: Support for speech synthesis in English and Chinese for ONNX inference in Non-JP-Extra models
2024-09-23 07:14:44 +09:00
tsukumi
0e0b4bea31
Fix: Disable optimization on InferenceSession initialization to speed up loading of ONNX models
2024-09-22 23:08:00 +09:00
tsukumi
9b27ba9777
Improve: Graphical display of ONNX conversion script logs
2024-09-18 06:11:10 +09:00
tsukumi
449927616a
Fix: Suppress many warning logs
2024-09-18 05:48:37 +09:00
tsukumi
fa7123c422
Fix: ONNX BERT model/tokenizer is not preloaded by default to avoid wasting VRAM
2024-09-18 05:11:15 +09:00
tsukumi
5e2c83c6c7
Improve: Support ONNX inference, add ONNX conversion script
2024-09-17 16:55:03 +09:00
tsukumi
9de6e04ad7
Add: Preparation for ONNX inference support ④
2024-09-17 13:25:17 +09:00
tsukumi
b9fdeb4492
Add: Preparation for ONNX inference support ②
2024-09-17 09:02:14 +09:00
tsukumi
c140c0612a
Fix: Old comment
2024-09-17 05:23:57 +09:00
tsukumi
cfd6bd5bb6
Improve: Models can now be loaded directly onto NVIDIA GPUs
2024-09-17 05:23:02 +09:00
tsukumi
0e71d7ef7c
Fix: Specifying the device_map option in extract_bert_feature()
2024-09-17 05:22:30 +09:00
tsukumi
80dd6dbc22
Improve: Use the device_map option in transformers to load BERT models directly to the GPU
2024-09-17 05:22:06 +09:00
tsukumi
bdf20e8911
Refactor: Use bert_models.transfer_model()
2024-09-17 05:21:03 +09:00
tsukumi
f84d2ed48b
Add: Implement bert_models.transfer_model()
2024-09-17 05:19:44 +09:00
litagin02
2274087da4
Fmt only (maybe)
2024-05-25 18:15:36 +09:00
litagin02
837c4397db
Merge pull request #118 from tsukumijima/dev
...
音声合成時に読み上げテキストの読みを表す音素列を指定する機能を追加 + 様々な改善
2024-05-15 07:37:46 +09:00
tsukumi
44693e7d6b
Refactor: improve regular expression performance during NLP
...
Improve performance by pre-compiling regular expressions that are executed many times.
2024-05-12 17:53:30 +09:00
gordon0414
7de367b76f
moved tokenizer loading inside of g2p function for code integrity.
2024-05-10 11:55:38 +09:00
gordon0414
3018309199
Optimize g2p function for improved speed and efficiency
...
- Moved global variables (ARPA, _g2p, eng_dict, tokenizer) to top-level scope to avoid redundant initialization.
- Simplified conditions for words with apostrophes.
- Streamlined __post_replace_ph function by removing redundant checks.
- Optimized __refine_syllables function with direct iteration.
- Avoided re-initialization of tokenizer in __text_to_words by moving it to global scope.
- Added performance test in __main__ to validate improvements.
- Resolves issue #104 .
2024-05-09 16:12:15 +09:00
tsukumi
2d6baa3928
Fix: changed the substitute phoneme from「ん」to 「'」 when the pronunciation can't be obtained
2024-05-06 22:17:34 +09:00
tsukumi
9d15d5763d
Refactor: run "hatch run style:fmt"
2024-04-19 04:17:27 +00:00
tsukumi
797c354b1b
Add: the ability to specify a phoneme sequence representing the reading of the text to be read during speech synthesis
...
With this feature, it is possible to specify individual readings for homonyms such as "明日 (ashita/asu)" while maintaining the Kanji representation of the read text.
2024-04-19 04:09:21 +00:00
litagin02
3d8b60c03c
Refactor
2024-03-15 12:27:46 +09:00
litagin02
73b01f7172
Add another wave dash for long symbol
2024-03-14 23:31:16 +09:00
tsukumi
07d246b98b
Refactor: run "hatch run style:fmt"
2024-03-12 18:27:08 +00:00
tsukumi
483bc68d57
Refactor: unify invoke format of open() function
2024-03-12 18:00:51 +00:00
tsukumi
74af2c831c
Add: cache_dir and revision optional arguments to load_model() / load_tokenizer()
2024-03-12 17:45:10 +00:00
litagin02
dada0af2b6
Format import using isort --profile black --gitignore --lai 2 .
2024-03-11 12:37:51 +09:00
litagin02
c776c08235
Apply black formatter
2024-03-11 09:47:47 +09:00
tsukumi
9c233630ef
Refactor: don't keep models in bert_feature module for each language
...
BERT models and tokenizers are already stored and managed in the bert_models module and should not be stored here.
In addition, since there may be situations where the user would like to use cpu instead of mps for inference when using it as a library, the automatic switching process to mps was removed.
2024-03-10 14:18:57 +00:00
tsukumi
b7d7c78203
Refactor: when pyopenjtalk_worker is called without initialization, continue processing without a worker
...
When using style-bert-vits2 as a library, the requirement to be able to launch it in multiple processes may not be necessary. Also, if the library is embedded and exe-ed using PyInstaller or similar, it is difficult to make pyopenjtalk_worker run in a separate process.
Therefore, we changed it so that the worker is used only when it is explicitly initialized.
2024-03-10 03:04:27 +00:00
tsukumi
0a74fe521b
Merge branch 'dev' of github.com:litagin02/Style-Bert-VITS2 into dev
2024-03-09 18:12:56 +00:00
tsukumi
61e2a1deae
Refactor: add type hints to style_bert_vits2.models.utils
2024-03-09 17:26:42 +00:00
