JPN_Romaji_Test2_Plus

Commercial Use: Not Approved

No warranties are made regarding fitness for purpose, functionality, accuracy, usefulness, or absence of defects.

• Release Date: Mar. 22, 2025
• Code version: v1.0.3
• Model version: Test2_Plus
• Languages used: Japanese
• Dictionaty used: Japanese-extention (Romaji prefered) (Please use phonemes supported by hparams.yaml）
• Applicability: Singing Voice
• Training Data: See data_providers.md
• Hyperparameters:See hparams.yaml

---

• 发布日期：2025.3.22
• 代码版本：v1.0.3
• 模型版本：Test2_Plus
• 使用语言：日语
• 使用词典：Japanese-extention（推荐使用Romaji）（请以hparams.yaml支持的音素为准）
• 适用范围：歌声
• 训练数据：参见data_providers.md
• 超参数：参见hparams.yaml

---

Compared to Test2
Thanks to @wolfgitpr for config providing. The size of the model has been significantly reduced.

Details：

model:
  hidden_dims: 128
  down_sampling_factor: 3
  down_sampling_times: 5
  channels_scaleup_factor: 1.3

optimizer_config:
  total_steps: 100000
  weight_decay: 0.1
  lr:
    backbone: 0.001
    head: 0.001
  freeze:
    backbone: False
    head: False

loss_config:
  losses:
    weights: [8.0, 0.1, 0.01, 0.1, 2.0, 1.0, 3.0]
    enable_RampUpScheduler: [False,False,False,True,True,True,True]
  function:
    num_bins: 10
    alpha: 0.999
    label_smoothing: 0.08
    pseudo_label_ratio: 0.3

gradient_clip_val: 0.5

The rest remains unchanged.

SOFA_model_JPN_Test2

Commercial Use: Not Approved

No warranties are made regarding fitness for purpose, functionality, accuracy, usefulness, or absence of defects.

• Release Date: Feb 16, 2025
• Code version: v1.0.3
• Model version: Test2
• Languages used: Japanese
• Dictionaty used: Japanese-extention (Romaji prefered) (Please use phonemes supported by hparams.yaml）
• Applicability: Singing Voice
• Training Data: See data_providers.md
• Hyperparameters:See hparams.yaml

---

• 发布日期：2025.2.16
• 代码版本：v1.0.3
• 模型版本：Test2
• 使用语言：日语
• 使用词典：Japanese-extention（推荐使用Romaji）（请以hparams.yaml支持的音素为准）
• 适用范围：歌声
• 训练数据：参见data_providers.md
• 超参数：参见hparams.yaml

---

Compared to Test1
The original annotations have been reviewed and optimized to ensure a more stable and consistent annotation style.

In terms of model performance, there appears to be some improvement in the boundary delineation between consonants and vowels, though the enhancement may be relatively modest.

SOFA_model_JPN_Test1

• Release Date: May 5, 2024
• Code version: v1.0.2
• Model version: Test1
• Languages used: Japanese
• Dictionaty used: Japanese Romanji
• Applicability: Singing Voice
• Training Data: See data_providers.md
• Hyperparameters:See hparams.yaml
• Commercial Use: Not Approved

---

• 发布日期：2024.5.5
• 代码版本：v1.0.2
• 模型版本：Test1
• 使用语言：日语
• 使用词典：Japanese Romanji
• 适用范围：歌声
• 训练数据：参见data_providers.md
• 超参数：参见hparams.yaml
• 商业用途：未经批准