The queue that looks like transcription is usually a review queue
A transcript backlog rarely means the transcription step is slow. It usually means nobody defined what “done” looks like, so every file sits in a review state that has no owner, no gate, and no exit condition. A long queue is the symptom. The cause is an SLA written against the wrong artifact.
If your SLA says “transcript delivered within X days,” you have measured the input, not the output. The publishable artifact is a speaker-labeled transcript, a reviewed caption file, or an approved summary. Each one has a different review burden and a different person who can sign it off. Until the SLA names that artifact and that person, the queue will keep growing at the review stage, not the transcription stage.
This is the same handoff problem that shows up between the live room and the post-event pipeline. If the room-to-chat handoff is where hybrid events break, the room-to-archive handoff is where content pipelines break. The fix is structural: define the artifact, name the owner, set the gate, and write down what happens when Q&A makes attribution ambiguous.
Define the deliverable before the deadline
Four artifacts come out of a recorded session, and they are not interchangeable:
- Raw transcript. Machine output with timestamps. Useful as an input. Not publishable. Section508.gov is explicit that current auto-captioning technology does not meet the minimum standard for pre-recorded media, and that captions must be edited when starting from auto-captions. The same logic applies to a raw transcript you intend to publish.
- Speaker-labeled transcript. The raw text plus verified speaker attribution. This is the artifact most Q&A sessions actually need, because the value of a panel transcript is knowing who said what.
- Reviewed caption file. A timed-text file in a format your player accepts. WebVTT is intended for marking up external text track resources in connection with the HTML
<track>element and provides captions or subtitles for video content. TTML2 is a content type representing timed text media for interchange among authoring systems. Both are interchange formats; neither is a quality guarantee. - Approved summary. A short editorial product, usually 150–400 words, that requires a human read of the transcript and a decision about what not to include.
Each artifact needs its own gate. A raw transcript can clear in hours. A speaker-labeled transcript for a four-person panel with audience Q&A may need a full review pass. A reviewed caption file needs a timing and readability check. An approved summary needs an editorial owner who can say no.
The practical move is to stop writing one SLA. Write four lines, one per artifact, each with a named role and a pass/fail condition. The role matters more than the deadline. A deadline with no owner is a wish.
Make Q&A the stress test, not an afterthought
Prepared remarks are easy. One speaker, one microphone, predictable vocabulary. Q&A is where the queue stalls, because it combines the three conditions that break attribution:
- Overlapping speech. Section508.gov advises avoiding overlapping voices whenever possible, and notes that speech over 180 words per minute may be too fast for captions. A heated Q&A exchange routinely exceeds both.
- Off-mic audience questions. The question is often the most valuable part of the segment and the least intelligible.
- Disputed speaker labels. Diarization collapses two people sitting close together into one voice, or splits one person into several. A panel-editing guide describes exactly this failure with a mono mix of four microphones, and recommends correcting labels by ear on disputed passages before doing anything else.
Two operational rules make Q&A survivable:
Rule 1: The moderator repeats the question into a live microphone. This is a production instruction, not a transcription instruction. A conference transcription guide recommends having a moderator repeat audience questions into a microphone to improve multi-speaker accuracy. It also gives you a clean audio source for the question, which shortens review time.
Rule 2: Disputed labels have a named resolver and a time limit. The SLA should say who resolves a disputed speaker label, how they are contacted, and what happens if they do not respond within the review window. The default should be to publish with a neutral label such as “Audience member” rather than hold the entire file. Holding the file is how a two-day review becomes a long queue.
Section508.gov states that when multiple speakers can be understood, all of the speech must be included in the captions, taking priority over perfect synchronization. That is a useful editorial principle for Q&A: completeness of speech beats frame-perfect timing. It also means the reviewer cannot quietly drop the messy exchange to make the file easier to finish.
Set quality gates from captioning guidance, not from vibes
An SLA needs a definition of “good enough to publish.” Captioning guidance gives you concrete gates you can put in a checklist:
- Punctuation, capitalization, and spelling. Section508.gov lists these as required edits when starting from auto-captions, along with checking names, places, and terms.
- Speaker identification. The DCMP Captioning Key treats speaker identification as part of equivalent access, not a nice-to-have. For a panel transcript, this is the gate that matters most.
- Sound effects and meaningful non-speech audio. Section508.gov says captions must include all important sounds. For a conference recording, that usually means laughter, applause, a door, or a microphone feedback event that changes the meaning of a moment.
- Readability and timing. The DCMP Captioning Key gives a minimum caption duration of 40 frames and a maximum of 6 seconds, with background music notation as an exception. Those constraints affect how fast a caption file can be reviewed, because a reviewer has to check timing, not just words.
- On-screen context. BBC subtitle guidance says that if onscreen graphics are not easily legible, subtitles must include text within those graphics that provides contextual information, including the speaker’s identity, what they do, and the organizations they represent. For conference recordings with lower-third name cards, this is a real review item.
None of these gates require a specific vendor or a specific turnaround time. They require a reviewer with a checklist and the authority to pass or fail the file.
What the SLA should actually say
A useful post-event content SLA has six fields per artifact. It fits on one page and can be pasted into the event runbook.
- Artifact type. Raw transcript, speaker-labeled transcript, reviewed caption file, or approved summary.
- Owner. A named role, not a team. “Content lead” is a role. “Marketing” is a queue.
- Review gate. The checklist the owner uses to pass or fail the artifact. Reference the quality gates above.
- Q&A attribution rule. Either “moderator repeats questions” or “audience questions are labeled ‘Audience member’ unless the speaker is identified on the recording.” Pick one and write it down.
- Escalation contact. Who resolves a disputed speaker label, and within what window. If no response, the default label applies.
- Publish / no-publish criteria. What must be true before the artifact goes live, and what must be true before it is withheld. Silence on this field is how unreviewed captions get published.
The SLA should also include a de-scope clause. Some sessions may ship without full speaker attribution; some must not ship at all until reviewed. Naming both prevents the quiet drift toward publishing whatever is finished first.
Copyable artifact: one-page post-event content SLA card
Paste this into the runbook and fill in the bracketed fields before doors open.
POST-EVENT CONTENT SLA CARD
Event: [event name] Date: [date] Version: [n]
ARTIFACT 1: RAW TRANSCRIPT
Owner: [role]
Gate: machine output reviewed for file integrity and timestamp continuity
Q&A rule: n/a
Escalation: [role] within [window]
Publish criteria: internal use only; not published
ARTIFACT 2: SPEAKER-LABELED TRANSCRIPT
Owner: [role]
Gate: speaker labels verified on all disputed passages; names checked against
speaker list; punctuation and capitalization corrected
Q&A rule: moderator repeats audience questions; unresolved labels default
to "Audience member"
Escalation: [role] within [window]; default label applies after window
Publish criteria: all speakers labeled or defaulted; no [unclear] markers in
the first 200 words
ARTIFACT 3: REVIEWED CAPTION FILE
Owner: [role]
Gate: timing checked against DCMP duration guidance (min 40 frames, max 6s,
music exception); sound effects included; on-screen names and titles
captured; line breaks at logical pauses
Q&A rule: same as Artifact 2
Escalation: [role] within [window]
Publish criteria: file validates in target player; no unreviewed auto-caption
text remains
ARTIFACT 4: APPROVED SUMMARY
Owner: [role]
Gate: editorial read of speaker-labeled transcript; quotes checked against
transcript; no attribution to a speaker who did not say it
Q&A rule: audience questions paraphrased unless the asker is identified
and consented
Escalation: [role] within [window]
Publish criteria: owner sign-off recorded in [system]
DE-SCOPE CLAUSE
Sessions that may ship without full attribution: [list]
Sessions that must not ship until reviewed: [list]
Reversible fixes when the queue slips
If the queue is already behind, tighten in this order. Each step is reversible and testable before the next event.
- Stop measuring “transcript delivered.” Start measuring the publishable artifact. If the dashboard says transcripts are on time but nothing is published, the metric is wrong.
- Add the moderator-repeat rule to the speaker briefing. It costs one sentence in the pre-event call and removes the most common Q&A attribution failure.
- Name the disputed-label resolver. One person, one window, one default. This is the single highest-leverage change because it converts an open-ended review into a bounded one.
- Separate the raw transcript from the publishable artifact in the queue. Raw transcripts can clear fast. Publishable artifacts need the gate. Mixing them makes the whole queue look slow.
- De-scope explicitly. If a session cannot be reviewed before the publish window, say so in the runbook and publish nothing from it. A missing transcript is a known gap. An unreviewed one is a liability.
The long queue is not a transcription problem. It is an undefined-artifact problem. Define the artifact, name the owner, set the gate, and write down what happens when Q&A makes attribution ambiguous. The queue gets shorter because the exit condition becomes visible.
FAQ
How long should a post-event transcript SLA be?
There is no universal benchmark in the sources reviewed here. University of Iowa guidance notes that paid captioning turnaround can be three to four business days, and that some services complete captions for content of 30 minutes or less within 24 hours. That is a vendor range, not an event benchmark. Set your SLA from your review capacity, not from a vendor’s marketing page.
Can we publish auto-generated captions if we label them as auto-generated?
Section508.gov states that current auto-captioning technology does not meet the minimum standard for pre-recorded media, and University of Iowa policy states that auto-generated captions do not meet accessibility requirements and must be reviewed and edited before publishing. Labeling does not fix the quality gap. Treat auto-output as an input to the review gate.
What if a speaker disputes their transcript after publication?
The SLA should include a correction path: who receives the dispute, what the correction window is, and whether the published artifact is versioned. A reversible fix is to publish with a visible correction log rather than silently editing the file. The log is the audit trail.
Do we need both WebVTT and TTML?
Only if your players require different formats. WebVTT is intended for external text tracks used with the HTML <track> element. TTML2 is designed for interchange among authoring systems. Pick the format your distribution platforms actually accept, and validate the file before the publish gate.
How do we handle Q&A where the audience question is inaudible?
Use the moderator-repeat rule as the primary fix. If the question is still unintelligible, Section508.gov guidance allows a descriptor such as “unintelligible” when an entire section has no understandable speech. Do not invent the question. Mark it and move on.