v0.0.7: Onnx Model for Mandarin&Japanese&English Singing Voice and automatically identify AP and EP.

Code version: v0.0.7

• Language used: Chinese Mandarin, Japanese Romanization (Mandarin&Japanese)
• Using the dictionary: ds-zh-pinyin-lite/japanese_dict_full/ds_cmudict-07b
• Scope of application: Singing voice
• Non acoustic automatic recognition: AP, EP (not recommend)
• Release date: 2025-12-30

---

Migrate CSV Transcription Files

The completed diffsinger dataset has been created. Use this script to migrate the CSV format annotations of the old dictionary to the new dictionary.

python scripts/migrate_dict.py csv [TRANSCRIPTION_CSV] \
    --source-dict [SOURCE_DICT] \
    --target-dict [TARGET_DICT] \
    [--save-path [OUTPUT_CSV] ] \
    [--overwrite]

- `TRANSCRIPTION_CSV`: Path to CSV file with transcriptions.

- `--source-dict`: Path to source dictionary file.

- `--target-dict`: Path to target dictionary file.

- `--save-path`: Path to save migrated file (defaults to original file).

- `--overwrite`: Overwrite existing file (optional flag).

---

Data contributors

• 白烁
• 风羽翼Tsubasa
• 烛曦遥Haruka
• 夜燐Yarin
• 芸青岩
• 绮萱

---

Full Changelog: v0.0.6...v0.0.7

v0.0.6: Onnx Model for Mandarin&Japanese&English Singing Voice and automatically identify AP and EP.

Code version: v0.0.6

• Language used: Chinese Mandarin, Japanese Romanization (Mandarin&Japanese)
• Using the dictionary: opencpop-expression/japanese_dict_full/ds_cmudict-07b
• Scope of application: Singing voice
• Non acoustic automatic recognition: AP, EP (not recommend)
• Release date: 2025-12-17

---

Onnx infer

pip install -r requirements_onnx.txt

python onnx_infer.py --onnx_path xxx --wav_folder xxx_wav --language zh ...

args:
- --onnx_path / -m: Path to Onnx models.
- --wav_folder / -wf: Input folder path. (default: segments)
- --out_path / -o: Path to the output label.
- --language / -l: Designated language, zh ja en yue. (default: zh)
- --non_speech_phonemes / -np: Non speech phonemes. (default: AP, optional AP,EP)
- --pad_times / -pt: The number of times to pad blank audio before reasoning. (default: 1)
- --pad_length / -pl: The max length of blank audio on the pad before inference. (default: 5)
- --dictionary / -d: Custom dictionary path.

---

Data contributors

• 白烁
• 风羽翼Tsubasa
• 烛曦遥Haruka
• 夜燐Yarin
• 芸青岩
• 绮萱

---

Full Changelog: v0.0.5...v0.0.6

v0.0.5 fixed: Onnx Model for Mandarin&Japanese&English Singing Voice and automatically identify AP and EP.

Code version: v0.0.5 fixed

• Language used: Chinese Mandarin, Japanese Romanization (Mandarin&Japanese)
• Using the dictionary: opencpop-expression/apanese_dict_full/ds_cmudict-07b. txt
• Scope of application: Singing voice
• Non acoustic automatic recognition: AP, EP (not recommend)
• Release date: 2025-12-8

---

onnx infer

pip install -r requirements_onnx.txt

python onnx_infer.py --onnx_path xxx --wav_folder xxx_wav --language zh ...

args:
- --onnx_path / -m: Path to Onnx models.
- --wav_folder / -wf: Input folder path. (default: segments)
- --out_path / -o: Path to the output label.
- --language / -l: Designated language, zh ja en yue. (default: zh)
- --non_speech_phonemes / -np: Non speech phonemes. (default: AP, optional AP,EP)
- --pad_times / -pt: The number of times to pad blank audio before reasoning. (default: 1)
- --pad_length / -pl: The max length of blank audio on the pad before inference. (default: 5)
- --dictionary / -d: Custom dictionary path.

---

Data contributors

• 白烁
• 风羽翼Tsubasa
• 烛曦遥Haruka
• 夜燐Yarin
• 芸青岩
• 绮萱