The gap a polished sentence can hide
The draft says “I do not authorise [inaudible 00:02:14].” The edited version reads “I authorise the shipment.” It reads smoothly, but this example has lost a negation and filled a gap. Before delivery, the transcriptionist needs to find both changes and return to the recording; plausible text does not establish what was said.
This guide proposes an AI trial for comparing a reference transcript with an edited version, preparing a discrepancy list and keeping uncertain passages open for listening. It includes fictional documents, a manual reference and a reusable prompt. The assistant receives text only: it does not listen to audio, recover words or decide the final transcript. The trial has not been run and has no measured results.
Agree what editing is allowed before comparing
Ask for the transcript type and the project style sheet: verbatim and edited transcripts may handle repetitions and fillers differently. Do not infer permission to rewrite a sentence from the word “clean”. Keep the received version, edited version and approved instructions separate.
GoTranscript’s guidelines distinguish full verbatim from clean verbatim and use timestamps for inaudible or unintelligible passages. The Rev guide consulted asks transcriptionists to preserve what was said and review inaudibles by listening. These are those providers’ conventions, rather than universal notation: follow the project rules and record the applicable style-sheet version.
Amberscript presents automatic and manual transcription services and an editor for reviewing text. This is its public offer description, rather than a performance test. Here we focus on explaining what changed between two texts and which human checks remain. A professional can trial this with fictional data without replacing their existing editor.
Prepare a small, traceable review file
For each fragment retain its identifier, file version, existing timestamp, speaker label and literal text in both versions. Add the instruction passage authorising a change, the observed discrepancy and review status. Do not invent missing timestamps: record “not located” and ask the transcriptionist to establish them by listening.
Keep the glossary as a research aid for spellings, separate from transcribed words. A company appearing in it does not establish that the speaker mentioned it. The assistant may flag an existing doubt; resolving it requires checking the recording and documenting the professional decision.
Start only with this guide’s fictional texts. Do not paste real recordings, unpublished interviews, voices, names or documents without authorisation and suitable confidentiality and data-handling terms. The cover is illustrative too and does not represent client work.
Fictional case: a negation, an uncertain name and an allowed edit
Every name, sentence, timestamp and document in PRUEBA-TX was invented by the editor; there is no associated recording. Timestamps are fictional references within the same file, rather than measurements. The following table is a hand-written reference, not AI output.
H1 · Fictional approved instructions: “Prepare an edited transcript. Removing the initial ‘uh’ in F3 is allowed. Filling uncertainty tags, changing negations, completing names or changing speaker labels is not allowed. Every doubt remains pending until listening and a documented decision. Text review does not authorise delivery.”
T1 · Reference transcript: “F1 · 00:02:14 · Voice A: I do not authorise [inaudible 00:02:14]. F2 · 00:04:08 · Voice B: I spoke to [uncertain name 00:04:08]. F3 · 00:05:20 · Voice A: Uh, the meeting is on Tuesday.” These tags are exercise inputs, rather than recommended formatting for every project.
G1 · Fictional glossary: “Ladera: a company name that might occur in the interview. Its mention in F2 is not confirmed.” T2 · Edited version: “F1 · 00:02:14 · Voice A: I authorise the shipment. F2 · 00:04:08 · Voice B: I spoke to Ladera. F3 · 00:05:20 · Voice A: The meeting is on Tuesday.”
The comparison should retain useful information rather than flagging everything as uncertain. F3 has an explicitly allowed change; F1 and F2 need different interventions.
The manual reference: what changed and what needs listening
This table compares the invented documents. Restoring T1 text in a working copy preserves its uncertainty; it does not establish that T1 matches the audio or approve delivery.
| Fragment | Text discrepancy | Internal action for the transcriptionist | Support |
|---|---|---|---|
| F1 · 00:02:14 | T2 removes “do not” and replaces the tag with “the shipment” | Flag two unauthorised changes; restore “I do not authorise [inaudible 00:02:14]” in the working copy and keep listening pending | T1 retains negation and tag; H1 prohibits changing them |
| F2 · 00:04:08 | T2 turns an uncertain name into Ladera | Restore the tag in the working copy and request listening. G1 supports research, rather than proving the mention | T1 contains the doubt; G1 states the mention is unconfirmed |
| F3 · 00:05:20 | “Uh,” is removed and initial capitalisation adjusted | Record the allowed edit without introducing new doubts about the day or changing speaker or timestamp | H1 permits removing the initial “uh”; the rest of the sentence is preserved |
A prompt for differences, rather than a reconstructed interview
Paste H1, T1, G1 and T2 in full after this instruction and save the response in a separate worksheet. Do not overwrite the original or export the response as a delivery.
“Compare T1 with T2 in PRUEBA-TX using only H1 and G1 as context. For each fragment return its original identifier, timestamp and speaker; exact T1 and T2 quotations; every detected change; the instruction allowing or prohibiting it; internal action and pending human check. In F1 check negation and tag replacement separately. A glossary does not establish that a word is audible. Do not fill gaps, complete names, invent audio or change timestamps or voices. Retain explicitly allowed changes. If support is missing, state that and retain uncertainty. Do not produce a final transcript or delivery approval.”
A hand-written internal F2 note would be: “Listen around 00:04:08 in the corresponding audio version and check the name; retain the T1 tag meanwhile.” This is a proposed task, rather than completed listening or a sent message.
Check whether the response has gone beyond its evidence
Compare the response with the table. Reject completion of F1 merely because “the shipment” fits, treating G1 as proof of F2, or undoing the allowed F3 edit. Also check that Tuesday, speakers and timestamps are unchanged. Keep both F1 discrepancies separate: detecting the filled gap while missing the negation would be incomplete review.
Try a separate variant: replace only G1 with “Ladera: commercial spelling confirmed by the client; the name audible in F2 has not been checked.” The manual reference still leaves F2 pending: confirmed spelling does not establish a mention in the recording. Keep the two exercise versions separate.
Claude’s documentation recommends allowing the model to acknowledge uncertainty and requesting checkable citations. These guidelines inform the prompt; they do not guarantee compliance or establish trial results. Text comparison may also miss errors already present in T1. Listening and professional acceptance remain necessary.
A pilot with known errors and complete review
Prepare a small fictional set containing allowed and problematic changes: deleted negation, a name completed from a glossary, a missing tag and an explicitly permitted filler removal. Write the manual reference first. Record missed unauthorised changes, false alerts about allowed changes, miscopied quotations and omitted fragments.
Compare the same set reviewed manually and with assistant support. Include preparation, checking and correction as well as response time. Decide which errors prevent accepting a proposal, and retain originals, responses and decisions. These are proposed pilot measures; no accuracy, savings or client outcomes have been demonstrated.
From a manual trial to an integration worth investigating
The manual trial involves copying fictional documents and reviewing differences. Connecting an automation would require studying actual formats and permissions, maintaining audio and text version links and recording who accepted each change. A possible review screen should separate proposals and approval; it would need design and testing. Compatibility with any transcript editor or service has not been checked.
Adaris develops custom software, CRM systems and AI automations case by case. A version-review workflow could be scoped around your process, with agreed acceptance criteria. This is a development proposal, rather than an existing proprietary transcription service. Commercial contact is handled in Spanish.
If you also work with accessible subtitles, the related guide addresses a different decision: which speaker or sound labels are supported and which identities the audience already knows.
Related services
Sources
Does review get lost between versions?
Tell us which documents you compare and who approves changes. We can assess a scoped development for your workflow. Commercial support is in Spanish.