License:
Custom-by-permission
Steward:
University of Hawaii at ManoaDataset ID:
cmwjfpeca009f07ntumc5o2m5
Release Date: 11/16/2026
Format: FLAC, TSV
Size: 540.31 MB
What it is. Audio segments and time-aligned transcripts from field recordings of casual Javanese conversation (mostly ngoko with some krama; rural cassette recordings from the 1980s). Joseph Errington (Yale University) collected the recordings and deposited them at PARADISEC as collection JJE1, "Javanese conversations 1" (items END01–END13, END15, END16: 15 items, 29 recording sides, 17.05 h). He supplied his typed transcripts for this project, and we aligned them to the recordings. This package holds the segments we trained on: the audio, the text and the timings. Contents. audio/<stage>/<PARADISEC file>/<segment_id>.flac: 2,167 clips (7.88 h) as 16 kHz mono 16-bit FLAC, lossless copies of the clips we trained on. transcripts.tsv has one row per clip: file, segment_id, stage, paradisec_file (for example JJE1-END01-A: collection, item and tape side), start_s and end_s (seconds from the start of that file), duration_s, tier, label_source and text. The clips form two sets. breadth: 213 segments (0.97 h) from our first alignment pass, used in the first training stage (once as they are and once as a 24 kbps Opus copy). breadth_insert: 1,954 segments (6.90 h) from the final pass, used once in the short second stage: GOLD 356 and SILVER 1,477 with the typed transcript as text, and 121 SILVER segments whose text is our own recogniser's output (label_source = our_recogniser). The two sets come from different alignment passes of the same recordings, so their spans overlap: 177 of the 213 breadth segments overlap a breadth_insert segment, and 88 neighbouring pairs within breadth_insert overlap one another by up to half a second. Together the clips cover 7.16 h of the recordings. Recordings. The complete recordings are at PARADISEC: Joseph Errington (collector), 1986. Javanese conversations 1. Collection JJE1 at catalog.paradisec.org.au [Open Access]. https://dx.doi.org/10.26278/K67K-2995. Registered PARADISEC users can download the files under PARADISEC's Conditions of Access. Each clip is the span from start_s to end_s of the named file, mixed down to mono and resampled to 16 kHz; we checked every clip sample for sample against our 16 kHz copies of the recordings. How the alignment was made. The transcripts were not time-aligned to the tapes. We aligned them with the MMS forced aligner (10 s windows) and compared each span with our own fine-tuned Whisper model (run locally) and, where available, a commercial hosted multimodal speech model, which received only Prof. Errington's audio (about 16 of the 17.05 hours). A segment was kept when the typed transcript agreed with at least one of the two (character error rate, CER, ≤ 0.30; GOLD additionally needs CER ≤ 0.15 and an alignment score above a null-calibrated bar); its text is then the typed transcript. Where the transcript was missing or disagreed but the two recognisers agreed with each other (CER ≤ 0.15), the text is our own recogniser's output. The hosted model's text was never used as a label. The breadth set comes from an earlier pass that kept the spans with the higher forced-alignment confidence. The typed transcripts are working transcripts, shared as they are; agreement with a recogniser is a quality check, not a guarantee that a transcript is complete or exact. Rights and terms. Prof. Errington, the collector and rights holder, gave us permission to use the recordings and transcripts on 15 September 2026 and agreed on 11 November 2026 that this material be made available on the Mozilla Data Collective. Use is for research and education, not for profit. Please credit Joseph Errington and cite the PARADISEC collection as given above, please point others to this listing rather than passing the files on, so that the credit and these terms stay with them, and please do not try to identify the people recorded. See LICENSE. Preparation code. Conversion to 16 kHz mono, forced alignment and tiering: pipelines/data_prep/errington_repair/ in the solution archive (run_alignment.sh gives the stage order). Clips longer than 30 s were excluded by the trainer; targets pass through training/text_norm.py (the scorer's normalisation and Javanese diacritic folding) at load time. Context. Lost In Transcription: Indonesian-Javanese Track (Mozilla Data Collective / DrivenData, August–October 2026); prize-finalist submission id-3228 (University of Hawaii at Manoa). Contact. Samuel Watson, University of Hawaii at Manoa — [email protected] Revision. 13 November 2026. Published by the University of Hawaii at Manoa team at the close of the Lost In Transcription competition, whose rules ask that external training data be made available on the Mozilla Data Collective. The complete recordings are held by PARADISEC: Joseph Errington (collector), 1986. Javanese conversations 1. Collection JJE1 at catalog.paradisec.org.au [Open Access]. https://dx.doi.org/10.26278/K67K-2995.
Licensing
Custom terms, by permission of the rights holder (Joseph Errington, Yale University): research and education use, not for profit; credit Joseph Errington and cite PARADISEC collection JJE1.
https://dx.doi.org/10.26278/K67K-2995Use is for research and education, not for profit. Credit Joseph Errington and cite the collection: Joseph Errington (collector), 1986. Javanese conversations 1. Collection JJE1 at catalog.paradisec.org.au [Open Access]. https://dx.doi.org/10.26278/K67K-2995. Please point others to this listing rather than passing the files on. The complete recordings are held by PARADISEC, which handles access to them (registered users, under PARADISEC's Conditions of Access).
Commercial use; attempts to identify the people recorded.
Intended Use
Training and evaluation of speech-recognition models for Javanese, and research on Javanese conversation. The package holds the Errington segments in the training data of a Lost In Transcription IN-JV submission.
What it is. Audio segments and time-aligned transcripts from field recordings of casual Javanese conversation (mostly ngoko with some krama; rural cassette recordings from the 1980s). Joseph Errington (Yale University) collected the recordings and deposited them at PARADISEC as collection JJE1, "Javanese conversations 1" (items END01–END13, END15, END16: 15 items, 29 recording sides, 17.05 h). He supplied his typed transcripts for this project, and we aligned them to the recordings. This package holds the segments we trained on: the audio, the text and the timings.
Contents. audio/<stage>/<PARADISEC file>/<segment_id>.flac: 2,167 clips (7.88 h) as 16 kHz mono 16-bit FLAC, lossless copies of the clips we trained on. transcripts.tsv has one row per clip: file, segment_id, stage, paradisec_file (for example JJE1-END01-A: collection, item and tape side), start_s and end_s (seconds from the start of that file), duration_s, tier, label_source and text. The clips form two sets. breadth: 213 segments (0.97 h) from our first alignment pass, used in the first training stage (once as they are and once as a 24 kbps Opus copy). breadth_insert: 1,954 segments (6.90 h) from the final pass, used once in the short second stage: GOLD 356 and SILVER 1,477 with the typed transcript as text, and 121 SILVER segments whose text is our own recogniser's output (label_source = our_recogniser). The two sets come from different alignment passes of the same recordings, so their spans overlap: 177 of the 213 breadth segments overlap a breadth_insert segment, and 88 neighbouring pairs within breadth_insert overlap one another by up to half a second. Together the clips cover 7.16 h of the recordings.
Recordings. The complete recordings are at PARADISEC: Joseph Errington (collector), 1986. Javanese conversations 1. Collection JJE1 at catalog.paradisec.org.au [Open Access]. https://dx.doi.org/10.26278/K67K-2995. Registered PARADISEC users can download the files under PARADISEC's Conditions of Access. Each clip is the span from start_s to end_s of the named file, mixed down to mono and resampled to 16 kHz; we checked every clip sample for sample against our 16 kHz copies of the recordings.
How the alignment was made. The transcripts were not time-aligned to the tapes. We aligned them with the MMS forced aligner (10 s windows) and compared each span with our own fine-tuned Whisper model (run locally) and, where available, a commercial hosted multimodal speech model, which received only Prof. Errington's audio (about 16 of the 17.05 hours). A segment was kept when the typed transcript agreed with at least one of the two (character error rate, CER, ≤ 0.30; GOLD additionally needs CER ≤ 0.15 and an alignment score above a null-calibrated bar); its text is then the typed transcript. Where the transcript was missing or disagreed but the two recognisers agreed with each other (CER ≤ 0.15), the text is our own recogniser's output. The hosted model's text was never used as a label. The breadth set comes from an earlier pass that kept the spans with the higher forced-alignment confidence. The typed transcripts are working transcripts, shared as they are; agreement with a recogniser is a quality check, not a guarantee that a transcript is complete or exact.
Rights and terms. Prof. Errington, the collector and rights holder, gave us permission to use the recordings and transcripts on 15 September 2026 and agreed on 11 November 2026 that this material be made available on the Mozilla Data Collective. Use is for research and education, not for profit. Please credit Joseph Errington and cite the PARADISEC collection as given above, please point others to this listing rather than passing the files on, so that the credit and these terms stay with them, and please do not try to identify the people recorded. See LICENSE.
Preparation code. Conversion to 16 kHz mono, forced alignment and tiering: pipelines/data_prep/errington_repair/ in the solution archive (run_alignment.sh gives the stage order). Clips longer than 30 s were excluded by the trainer; targets pass through training/text_norm.py (the scorer's normalisation and Javanese diacritic folding) at load time.
Context. Lost In Transcription: Indonesian-Javanese Track (Mozilla Data Collective / DrivenData, August–October 2026); prize-finalist submission id-3228 (University of Hawaii at Manoa).
Contact. Samuel Watson, University of Hawaii at Manoa — [email protected]
Revision. 13 November 2026.
Package. 2,167 clips / 7.88 h as 16 kHz mono FLAC, with their transcripts. Package sha256: b1aa9717cc6bfc6e758613eb2b3eddc6a9accca038a34cdb3050557aae4557de.