VideoText workflow guide
Rev's AI transcription produces fast first drafts that still need human review. This guide covers the most common AI transcription errors, how to correct them efficiently, and the post-processing workflow for cleaning up Rev AI output.
Choose Rev, GoTranscript, TranscribeMe, Scribie, or a custom client specification. Each guideline has distinct rules for verbatim level, speaker label format, timestamp intervals, inaudible notation, and maximum paragraph length.
Clean verbatim removes fillers, false starts, and repetitions for readability. Full verbatim preserves all spoken content. Applying the wrong level is the most common single reason transcripts fail marketplace QA — it cannot be fixed without re-reading the source audio.
A single speaker must have exactly one label format from the first occurrence to the last. Mixed formats (JOHN SMITH / John / J. Smith) require a find-and-replace pass across the full document before any other formatting work.
Rev-style: timestamps every 2 minutes or at each speaker change. GoTranscript: job-dependent — bold [00:00:00] either every 2 minutes or at every speaker change, with no timestamping on the qualification test. TranscribeMe: per-speaker-turn timestamps. Paragraph limits range from 8 lines (Rev) to roughly 500 symbols (GoTranscript).
Verify inaudible notation consistency (all [inaudible] or all [INAUDIBLE], never mixed), check bracket format for crosstalk sections, confirm verbatim level is consistent throughout, and validate that paragraph length does not exceed the client's maximum.
Clean verbatim text. Speaker names in CAPITAL LETTERS followed by a colon. Timestamps in [HH:MM:SS] format at 2-minute intervals and at speaker changes. No paragraph longer than 8 lines. Inaudible sections marked [inaudible]. False starts and fillers removed.
Clean verbatim on most jobs, full verbatim where the job says so. Bold speaker labels followed by a colon and a single space. [inaudible 00:00:00] and [unintelligible 00:00:00] for unclear audio, [crosstalk] for overlapping speech. Timestamps only where the job calls for them, in bold [00:00:00]. Paragraph breaks at topic shifts, under roughly 500 symbols.
Structured document with consistent heading-style speaker labels, paragraph breaks, correctly formatted timestamps, and clean or full verbatim as specified. Formatted for direct delivery — no manual cleanup required before attaching to the client email.
Reduce revision risk before submitting marketplace jobs. A pre-delivery formatting check catches the label inconsistency, timestamp format error, or verbatim level mismatch that triggers a QA rejection and unpaid revision.
Create a consistent formatting baseline across a team of transcriptionists working on the same client account — so reviewers spend time on content accuracy, not fixing label formats.
Turn raw AI-generated transcripts into readable interview documents: speaker structure, consistent punctuation, accurate paragraph breaks, and timestamps that let readers verify context in the source recording.
Page 1: "JOHN SMITH: Thanks for joining us." Page 3: "John: So as I was saying..." Page 7: "Speaker 1: Right, exactly." — All three refer to the same speaker. Label drift of this kind fails QA at every major transcript marketplace.
[00:05:12] on page 2, (00:05:12) on page 4, [5:12] on page 6, and 00:05:12 on page 8 — all in the same document. Most style guides require a single format, and inconsistency triggers automatic rejection on automated QA systems.
"[inaudible]", "[INAUDIBLE]", "[unclear]", and "[crosstalk]" appearing in the same file for the same type of audio problem. Rev requires "[inaudible]" in lowercase brackets; GoTranscript requires a full timestamp inside the tag and keeps [crosstalk] separate. Neither accepts a mix of styles.
Pages 1–4 are clean verbatim (fillers removed). Pages 5–8 switch to full verbatim (fillers preserved). This happens when an ASR tool applies different post-processing across transcript segments — and it fails QA because the verbatim level is supposed to be consistent throughout.
Clean verbatim by default. Speaker names in ALL CAPS followed by a colon. Timestamp format [HH:MM:SS] every 2 minutes and at speaker changes. Maximum paragraph: 8 lines. Inaudible: [inaudible]. Crosstalk: [crosstalk]. False starts and filler words removed.
Clean verbatim by default, full verbatim where the job specifies it. Bold speaker label, colon, single space (Speaker 1:, Interviewer:), with a ? prefix when the speaker is uncertain. [inaudible 00:00:00] for speech that cannot be heard and [unintelligible 00:00:00] for speech heard but not understood; [crosstalk] for overlap. Timestamps are job-dependent: bold [00:00:00] every 2 minutes or at every speaker change. Paragraphs under roughly 500 symbols. Spell out zero through nine.
Strict verbatim — all fillers and false starts preserved. Timestamps required at each speaker turn. Speaker format "SPEAKER NAME:" in caps. Specific bracket format for technical notation. Paragraph breaks only at speaker changes.
Many agencies specify hybrid rules that don't map cleanly to any standard guide — for example, Rev-style speaker labels with GoTranscript-style clean verbatim and no timestamps at all. These combinations must be documented explicitly because the formatter cannot infer them.
Rev AI reports approximately 80–85% accuracy on clear audio, which is typical for AI transcription. Accuracy drops significantly with accents, multiple speakers, technical jargon, background noise, and low-quality audio. Human transcription on Rev reaches 99%+ accuracy.
Common Rev AI errors: (1) homophones (their/there/they're, to/too/two); (2) proper nouns and names misspelled; (3) technical or domain-specific terms incorrect; (4) punctuation errors; (5) speaker label mix-ups during crosstalk; (6) missing words in fast speech or overlapping dialogue.
Best workflow: (1) Listen to the audio once while reading the AI transcript to spot obvious errors; (2) Fix proper nouns and names first; (3) Review homophones in context; (4) Check speaker labels at every speaker change; (5) Fix punctuation; (6) Verify technical terms against context. Use playback speed control to catch fast-speech errors.
Yes. Rev AI transcription costs $0.25/minute vs. $1.99+/minute for human transcription. For content where 80–85% accuracy is acceptable (internal notes, rough drafts, searchable archives), AI is cost-effective. For legal, medical, or publishable content, human transcription is recommended.
VideoText offers AI transcription using Whisper large-v3, which achieves similar or better accuracy than Rev AI on most content types, at a lower cost on monthly plans. VideoText starts free with tool-specific free access. For the same post-processing workflow, VideoText output can be edited using the Guideline Format tool to meet Rev's formatting standards.
