The Transcripts Were Never There: A Full-Pile Audit of the Linguistic Atlas of the Gulf States Web Archive, and a Failed Calibration for ASR Re-Transcription

The Linguistic Atlas of the Gulf States (LAGS; interviews 1968–83) survives online as 1,082 informant folders of audio at the Linguistic Atlas Project; its formal products — most recently DASS2019 — have used only 64 speakers. We pre-registered a transcription-fidelity audit of the pile: re-transcribe every informant's audio with a general automatic speech recognition (ASR) ruler and score agreement with the original transcriptions. The audit's first finding is that the comparison cannot be run: the folders contain no transcriptions. The per-informant .xml is an audio segment index, not a transcript; each folder's only text is a biography; verbatim transcripts exist only for DASS2019's 64, in a separate tree. We therefore report (i) a census and index-integrity audit, and (ii) the pre-registered calibration gate on the curated anchor. The census finds 983 of 1,082 folders (90.9%) hold audio and index, totalling 70,542 segment MP3s across 5,807 reels; but the index disagrees with the files in both directions — 939 indexed segments (1.45%) absent on disk and 962 files unindexed (1,901 in all), including a column-scrambled index and a folder whose 81 indexed segments have no audio. The calibration gate fails: re-transcribing two DASS2019 speakers' pile MP3 segments (reel 1; 18,425 reference words) with Whisper large-v3 yields word agreement of 57.4% and 64.9% (pooled 61.9%), far below the pre-registered 85% fitness bar. Per the pre-registration, no fidelity claim is made for the ~850 never-audited informants: the stranded majority of a foundational dialect survey is auditable only for completeness and index integrity.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24409284
Primary Topic
Digital and Traditional Archives Management
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

The Transcripts Were Never There: A Full-Pile Audit of the Linguistic Atlas of the Gulf States Web Archive, and a Failed Calibration for ASR Re-Transcription

Atharva Chourey
Zenodo (CERN European Organization for Nuclear Research)
Digital and Traditional Archives Management
preprint

The Transcripts Were Never There: A Full-Pile Audit of the Linguistic Atlas of the Gulf States Web Archive, and a Failed Calibration for ASR Re-Transcription

Atharva Chourey
preprint en

Abstract

The Linguistic Atlas of the Gulf States (LAGS; interviews 1968–83) survives online as 1,082 informant folders of audio at the Linguistic Atlas Project; its formal products — most recently DASS2019 — have used only 64 speakers. We pre-registered a transcription-fidelity audit of the pile: re-transcribe every informant's audio with a general automatic speech recognition (ASR) ruler and score agreement with the original transcriptions. The audit's first finding is that the comparison cannot be run: the folders contain no transcriptions. The per-informant .xml is an audio segment index, not a transcript; each folder's only text is a biography; verbatim transcripts exist only for DASS2019's 64, in a separate tree. We therefore report (i) a census and index-integrity audit, and (ii) the pre-registered calibration gate on the curated anchor. The census finds 983 of 1,082 folders (90.9%) hold audio and index, totalling 70,542 segment MP3s across 5,807 reels; but the index disagrees with the files in both directions — 939 indexed segments (1.45%) absent on disk and 962 files unindexed (1,901 in all), including a column-scrambled index and a folder whose 81 indexed segments have no audio. The calibration gate fails: re-transcribing two DASS2019 speakers' pile MP3 segments (reel 1; 18,425 reference words) with Whisper large-v3 yields word agreement of 57.4% and 64.9% (pooled 61.9%), far below the pre-registered 85% fitness bar. Per the pre-registration, no fidelity claim is made for the ~850 never-audited informants: the stranded majority of a foundational dialect survey is auditable only for completeness and index integrity.

Zenodo (CERN European Organization for Nuclear Research)
Digital and Traditional Archives Management
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon