Printto TH Dataset

Thai DiffSinger dataset for multi-speaker training.


About This Dataset

The Printto TH dataset is specifically curated for training DiffSinger to sing in the Thai language. It is sourced from Printto Magicbeat and comprises 42 minutes and 39 seconds of Thai singing data. The dataset is fully labeled using PRINTmov's Thai phoneme system, ensuring precise and accurate phonetic representation for effective training.

Phonemes

Below is the PRINTmov's Thai phoneme system used in this dataset.

The spreadsheet can also be accessed from https://www.printmov.com/thai-diffsinger.html

Download

This section provides access to the sample dataset, primarily designed for multi-speaker training purposes. Please strictly follow the usage guidelines and rules provided in readme.txt within the dataset. It is strictly prohibited from using this dataset with any other person's voice without obtaining their explicit consent and permission.

The Printto TH dataset contains fully labeled Thai singing data in NNSVS format separated into 3 speakers: Printto V1, Printto V2, and Pyao (annonymous).

Please note that the dataset may contain some copyrighted lyrics. The primary intended use of this dataset is for educational purposes.

    To make it easier to manage the policy, the download link will be sent to your email provided below.

Training Results

Dataset Result

The demo below is the result of training this dataset for 140,000 steps.


Multi-speaker Training Result

Despite its small size, the dataset has been tested for multi-speaker training alongside a Japanese dataset, yielding satisfactory results.

Dictionary

Thai DiffSinger Dictionary

Thai dsdict-th.yaml is required in order to use the Thai DiffSinger Phonemizer in OpenUTAU. It uses the same phonemes for both long vowels and short vowels, so the timing adjustment should be made manually in OpenUTAU by the user.

Download dsdict-th.yaml File

You can download dsdict-th.yaml taken from Printto Magicbeat's voice library here.

Download dsdict-th.yaml