ub-MOJI v26.09: Over 5,000 New Video Samples Added

We are pleased to announce the release of the latest version of the ub-MOJI Dataset, version 26.09.

A major expansion: 4 new participants bring ~7x more samples

This release welcomes 4 new participants (participant IDs 23–26), and with them the dataset has grown dramatically:

	v25.09	v26.09	Change
Syllable video samples	765	5,394	+4,629
Continuous video samples (sequences + words)	183	1,295	+1,112
Total video samples	948	6,689	+5,741

That’s roughly 7x more video samples than the previous release. The 4 new participants alone account for 5,536 of these new samples — 4,469 syllable video clips and 1,067 continuous video recordings (sequences/words). The remaining growth comes from filling in missing videos and correcting annotations for existing participants (see below).

Dataset restructuring

The dataset is now split into two subsets, each with its own metadata file:

ub-MOJI/
├── isolated/                # single kana characters
│   ├── metadata.parquet
│   └── videos/
└── continuous/               # continuous videos
    ├── metadata.parquet
    └── videos/
        ├── sequences/
        └── words/

You can now load either one directly with the datasets library, instead of downloading everything at once:

from datasets import load_dataset

isolated_dataset = load_dataset("kanglabs/ub-MOJI", "isolated", revision="v26.09")
continuous_dataset = load_dataset("kanglabs/ub-MOJI", "continuous", revision="v26.09")

Stable IDs for every annotated segment

Previously, syllable video clips were named after the recording they came from, e.g. ka_018_202310_t001.mp4.

Each segment in annotations.toml now carries a unique annotation_id, and syllable clips are renamed to match, e.g. ann_08ef49579212.mp4. This ID stays the same even if its label or segment boundaries are corrected later, so you can reliably track a specific clip across dataset versions. The original participant, date, and take information is still available via the annotation_id and source_video columns in isolated/metadata.parquet.

Other fixes

• Missing videos have been added and annotations corrected for 44 sequence samples, and the corresponding isolated video clips have been regenerated to match.
• Outdated class names in videos and annotations have been corrected.
• Missing metadata for the zi class has been fixed.
• fps and duration values for newly added samples have been fixed, and the BibTeX citation entry has been corrected.

For the full list of changes, please refer to the CHANGELOG.

Acknowledgments

We would like to express our sincere gratitude to the following organizations for their cooperation in the dataset recordings:

• Nakano Volunteer Center, Nakano-City Social Welfare Conference (中野区社会福祉協議会 中野ボランティアセンター)
• 中野区聴覚障害者福祉協会

We would also like to thank all the participants who took part in the recordings.