Loading datasets...
Release Date: 11/10/2026
Format: FLAC, TSV
Size: 499.55 MB
What it is. 981 speech segments (6.60 h) cut from 96 YouTube videos in Javanese / colloquial Indonesian (vlogs, comedy, theatre, semi-spontaneous talk) whose YouTube licence field was "Creative Commons Attribution license (reuse allowed)" when checked. Labels are pseudo-labels: the output of our own fine-tuned Whisper models, run locally; a segment was kept where two independently seeded models agreed at ≥ 0.8 (1 − WER) and the label was ≤ 200 tokens and ≤ 3 words/s (+2 words). Captions were not used. Licence. Audio: CC BY 3.0 per source video (YouTube CC-BY), attributed to each uploading channel in videos.tsv and in the per-row columns of manifest.tsv (video URL, channel id, title where captured, licence field and check timestamp). Labels: our model output, released under CC BY 4.0. Contents. audio/ holds the segments as 16 kHz mono FLAC; manifest.tsv has one row per segment (file, duration, pseudo-label text, video id/URL, segment index and start/end offsets in the source video's audio, channel, licence evidence); videos.tsv lists the 96 source videos. Clips longer than 30 s were excluded from training by the trainer; they are included here unchanged. Targets passed through the competition's own text normaliser at training time; the text in manifest.tsv is as distributed by the source. Context. Lost In Transcription: Indonesian-Javanese Track (Mozilla Data Collective / DrivenData, Sept–Oct 2026); prize-finalist submission id-3284 (University of Hawaii at Manoa). Published at the close of the competition to satisfy the rule that external training data be made available on the Mozilla Data Collective. Contact. Samuel Watson, University of Hawaii at Manoa — [email protected] by the University of Hawaii at Manoa team at the close of the Lost In Transcription competition to satisfy the rule that external training data be made available on the Mozilla Data Collective. The package holds the audio of the subset we trained on, a manifest (file, duration, transcript), a datasheet, the licence and SHA-256 sums of every file. The data are re-published under their original open licence with attribution to the original creators (see DATASHEET.md and LICENSE).
Licensing
CC BY 3.0 per source video (YouTube Creative Commons Attribution licence); pseudo-labels CC BY 4.0
https://creativecommons.org/licenses/by/3.0/CC BY 3.0 per source video: attribute the uploading channel listed for each segment (videos.tsv / manifest.tsv). The labels are machine-generated pseudo-labels, not human transcripts (CC BY 4.0).
Any use that the YouTube Creative Commons Attribution licence or the uploaders' terms forbid; presenting the pseudo-labels as human transcripts.
Intended Use
Training and evaluation of speech-recognition models; ASR research. Re-publication of an openly licensed corpus subset.
What it is. 981 speech segments (6.60 h) cut from 96 YouTube videos in Javanese / colloquial Indonesian (vlogs, comedy, theatre, semi-spontaneous talk) whose YouTube licence field was "Creative Commons Attribution license (reuse allowed)" when checked. Labels are pseudo-labels: the output of our own fine-tuned Whisper models, run locally; a segment was kept where two independently seeded models agreed at ≥ 0.8 (1 − WER) and the label was ≤ 200 tokens and ≤ 3 words/s (+2 words). Captions were not used.
Licence. Audio: CC BY 3.0 per source video (YouTube CC-BY), attributed to each uploading channel in videos.tsv and in the per-row columns of manifest.tsv (video URL, channel id, title where captured, licence field and check timestamp). Labels: our model output, released under CC BY 4.0.
Contents. audio/ holds the segments as 16 kHz mono FLAC; manifest.tsv has one row per segment (file, duration, pseudo-label text, video id/URL, segment index and start/end offsets in the source video's audio, channel, licence evidence); videos.tsv lists the 96 source videos. Clips longer than 30 s were excluded from training by the trainer; they are included here unchanged. Targets passed through the competition's own text normaliser at training time; the text in manifest.tsv is as distributed by the source.
Context. Lost In Transcription: Indonesian-Javanese Track (Mozilla Data Collective / DrivenData, Sept–Oct 2026); prize-finalist submission id-3284 (University of Hawaii at Manoa). Published at the close of the competition to satisfy the rule that external training data be made available on the Mozilla Data Collective.
Contact. Samuel Watson, University of Hawaii at Manoa — [email protected]
Package. 981 files / 6.6 h. Package sha256: f2574e4363c1269ab2cc5ec6bf7454c0d778d54efb4f354a1f3f5ee8a3127777.
