📄 Quick Rich Transcriptions of Arabic Broadcast News Speech Data

Summary

The paper reports how ELDA and LDC produced Arabic speech transcripts at very large scale for the DARPA GALE program, where the transcripts were used to train ASR systems that feed machine translation. The core difficulty was speed. Careful transcription costs far too much effort per audio hour, so the team used LDC's Quick Rich Transcription (QRTR) guidelines. QRTR sits between plain quick transcription and careful transcription: every sentence unit gets a time stamp and a type label, but the remaining markup stays minimal. LDC recorded TV and radio broadcasts by satellite, and native speakers audited them for language, content and audio quality. ELDA's transcribers then worked in the XTrans tool, covering both Modern Standard Arabic and dialect, in Broadcast News and Broadcast Conversation genres. After an initial training stage, full double-checking gave way to a capped "Quick Verification". This brought effort down to about 20 work hours per audio hour across roughly 2050 hours of audio. The project shows that useful rich transcripts can be produced on a budget when the level of detail and the level of quality control are both chosen deliberately.

Key points

• About 2050 hours of Arabic broadcast audio were transcribed. They arrived in 24 batches of 20–300 hours each and were handled by roughly 40 transcribers, half of them working remotely.
• The transcription factor began at 30–35 and fell to the target of 20 or below. Spikes came from new hires and from batches with more dialect or overlapping speech.
• XTrans exports TDF files, with one segment per line and 13 tab-separated fields. These include speaker type and dialect, section, turn, segment, sectionType and suType.
• There are four sentence unit (SU) types: statement, question, incomplete and non-speech. Sections are labelled as news report, conversation or miscellaneous, and miscellaneous sections are left untranscribed.
• Quick Verification checks three 3-minute excerpts, taken from the start, middle and end of a file, and is capped at 18 minutes. Files that fail or run over the cap go back to the transcriber with a report.

Interesting ideas

• I1. Annotation depth is treated as a graded choice (quick, quick-rich, careful), with each tier defined by which features it records and how much checking it gets, so effort can be matched to the downstream need.
• I2. Supervision is front-loaded: newcomers get full supervisor review, then peer cross-checks, and only after that do they move to cheap spot checks, so expensive QC is spent where errors are most likely.
• I3. Spot checks are taken from fixed positions (beginning, middle, end) under a hard time cap, which makes verification cost predictable and catches drift within a file.
• I4. Utterances are labelled as statement, question, incomplete or non-speech, and the "incomplete" label covers both self-restarts and interruptions instead of forcing a full parse.
• I5. Uncertain stretches are wrapped in a dedicated marker that means "best guess", and all transcribers share one agreed spelling for each proper name.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Silver data could be labelled with an explicit depth tier, such as tags only, tags plus tree, or full frames, so that training and evaluation know how much each example can be trusted.
• I2 (learn): New witness parsers or newly induced repair rules could be audited heavily at first and then moved to sampled audits once their agreement record is stable.
• I3 (learn): When validating long silver-annotated documents, sampling sentences from these three positions could expose errors that grow over the document, such as coreference or tracking failures.
• I4 (frames): The speech-act layer could include an explicit "incomplete/interrupted" act, so that broken dialogue lines are not given full who-did-what frames.
• I5 (world): Entity mentions could carry a confidence flag and resolve to one canonical name form, which keeps the text's world model consistent and makes doubtful facts visible in explanations.