---

DiffSinger training notebook:

current supported data format:

• lab + wav (NNSVS format)
• csv + wav (DiffSinger format)
• ds (DiffSinger .ds file) [broken]

NOTE:

• your_speaker_folder's folder name will be used as spk_name so please be careful about your file naming
• colab notebook primarily uses python; thus space in file name or folder path may be invalid
• for an in-depth guide for SVS training and/or labeling, please see SVS Singing Voice Database - Tutorial
• it is advised to edit your data using SlurCutter for a more refined data for your pitch model
• please visit DiffSinger Discord for any help and questions regarding model production

Zip file format examples:

#single speaker (lab + wav | ds + wav)
your_zip.zip:
    |
    |
    your_speaker_folder:
        |
        |
        data_1.wav
        data_1.lab (or.ds)
        .
        data_2.wav
        data_2.lab (or.ds)
        .
        data_3.wav
        data_3.lab (or.ds)
        .
        ...

#single speaker (csv + wav)
your_zip.zip:
    |
    |
    your_speaker_folder:
        |
        |
        wavs (folder named "wavs" containing all the wavs)
        .
        transcriptions.csv

#multi speaker (lab + wav | ds + wav)
your_zip.zip:
    |
    |
    your_speaker_folder_1:
        |
        |
        data_1.wav
        data_1.lab (or.ds)
        .
        data_2.wav
        data_2.lab (or.ds)
        .
        data_3.wav
        data_3.lab (or.ds)
        .
        ...
    your_speaker_folder_2:
        |
        |
        data_1.wav
        data_1.lab (or.ds)
        .
        data_2.wav
        data_2.lab (or.ds)
        .
        data_3.wav
        data_3.lab (or.ds)
        .
        ...

#multi speaker (csv + wav)
your_zip.zip:
    |
    |
    your_speaker_folder_1:
        |
        |
        wavs (folder named "wavs" containing all the wavs)
        .
        transcriptions.csv
    your_speaker_folder_2:
        |
        |
        wavs (folder named "wavs" containing all the wavs)
        .
        transcriptions.csv

---

Vocoder finetuning notebook:

current supported data format:

• wav

NOTE:

• it is suggested to use manual segmented audio for cleaner segments (though there's minimal difference when using the auto segmentation)
• zip file format can consist of any type of files, even subfolders. data extraction will only account .wav that are within the zip into the training set

---

SOFA training notebook (wip):

current supported data format:

• lab + wav (NNSVS format)

NOTE:

• this notebook is still a rough draft, please either don't use it at all or use it with caution....

---

MFA inference notebook:

current supported data format:

• wav
• txt + wav (wav with text transcription)
• lab + wav (wav with text transcription)

NOTE:

• this notebook is the alternative force alignment option, but SOFA should works better on singing data
• zip must have no subfolders

---

Plans (update might not be in order):

• [notebook] improve SOFA notebook, add inference
• [notebook] fix .ds training option

---

Credits:

• openvpi for DiffSinger fork and more

• UtaUtaUtau for nnsvs-db-converter

• Kei for the original notebook

• MLo7 for the repo's content

• PixPrucer for an in-depth SVS guide

• haru0l for the base pretrain with embeds

• AgentAsteriski for the current local GUI