The script knows a name the audience has not learned yet
An off-screen voice says “I am back.” The production script names Lara, but the scene has not identified her. Adding her name to a draft accessible subtitle may reveal information the audience does not yet have. Before changing the label, return to the video and check the audio, the image and what the story has revealed at that point.
This guide proposes a manual AI trial to review the evidence behind speaker and sound labels in subtitles for Deaf and hard-of-hearing viewers, or SDH. It includes fictional notes, an editor-written reference table and a prompt. The assistant receives text only: it has not heard the scene. Label selection, timing and delivery approval remain the professional’s responsibility. The trial has not been run and has no measured results.
Keep three sources separate
The script, viewing notes and subtitle draft serve different purposes in this exercise. Keep the script as a production reference; record the reviewer’s audio and visual observations separately; preserve the exact proposed delivery text. A name in the cast list does not establish that a scene has revealed that identity, and a written action does not establish that its sound is present in the final mix.
W3C explains that captions cover speech and audio information needed to understand the content. Its guide also recommends checking automatic captions. Netflix’s Spanish style guide advises against unrevealed names when a speaker label is needed, with a specific exception for certain documentary or unscripted material. It also distinguishes sounds from visual actions. Each reference has its own scope: agree the style sheet for the job. This guide does not certify compliance or platform acceptance.
On pages read on 23 October 2026, LinguaVox describes accessible subtitling with speakers and sounds, while MultiSignes offers subtitling for Deaf viewers. These are their public service descriptions. Here we focus on a particular professional task: justifying each suggested label without filling gaps in the material with imagination.
Prepare a review sheet before asking for a response
Give each draft event an identifier and retain its internal link to the video version and segment, if already available. Keep dialogue, draft label, audio observation, visual observation, narrative information available so far, source document and unresolved question in separate fields. E1, E2 and E3 below are identifiers, not timecodes.
Explicitly mark audio as not reviewed where appropriate. Do not replace that status with no sound: they mean different things. Do not turn a recorded bang into someone slamming a door without evidence identifying its source. Reserve fields for the subtitler’s decision, the style-sheet version and the reason for a change.
Start with the fictional material below only. Do not upload unreleased footage, voices, participant names or real scripts without authorisation and suitable confidentiality and data-handling terms. The generated cover is an illustrative scene, not client work.
Fictional case: an early identity reveal and an unheard door slam
PRUEBA-SDH and all its documents, characters and events are invented by the editor. There is no associated video: the viewing observations are fictional exercise inputs too. The table below is a manual reference, not an AI-generated response.
G1 · Production-script excerpt: “E1: Lara, off-screen: I am back. E2: closes a door. E3: music starts.” It contains no approved subtitling instructions.
V1 · Fictional review notes: “E1: I am back is audible; the speaker is not visible throughout the event and their identity has not been revealed previously. No other vocal characteristics have been recorded. E2: audio not yet reviewed; the G1 action has not been checked against the video. E3: slow instrumental music is audible; no title or emotional intent identified. It interrupts dialogue and the reviewer considers it relevant; its source is not visible.”
B1 · Draft for review: “E1: [Lara] I am back. E2: [door slams]. E3: [cheerful music].” H1 · Fictional brief: “Prepare justified proposals for professional review. Do not generate a delivery file, timecodes or accessibility certification. Final wording must follow the style sheet validated by the person responsible.”
The aim is to retain supported information, locate missing evidence and leave the subtitler a well-documented decision.
What the review should flag
This table covers only the fictional case’s textual evidence check. Suggestions are internal notes awaiting a decision, rather than ready-to-publish captions.
| Event and draft | What V1 supports | Change or question for the professional | Source |
|---|---|---|---|
| E1 · [Lara] I am back. | Off-screen voice, identity not revealed; dialogue recorded | Flag the early name. Retain dialogue; request a generic identifier following the style sheet, without inventing vocal traits | G1 knows the name; V1 limits what is revealed; B1 reveals it early |
| E2 · [door slams] | Audio not reviewed; even the sound is unconfirmed | Leave the label unvalidated and request audiovisual review. Do not turn a script action into a confirmed sound or silence | G1 records an action; V1 records the outstanding check |
| E3 · [cheerful music] | Slow instrumental music, relevant according to the note; emotion unidentified | Suggest [slow instrumental music] for review. Remove unsupported emotion; add no title, composer or particular instrument | V1 supports the description; H1 requires a professional decision |
A prompt to try with these documents
Paste G1, V1, B1 and H1 in full after the prompt. Keep the response as a separate draft; do not import it over the subtitle file.
“Review the PRUEBA-SDH labels using only G1, V1, B1 and H1. For each event return its draft label, supporting observations, unchecked information, internal suggestion and a question for the subtitler. Cite the document and exact passage supporting each observation or change; state when support is missing. Distinguish script knowledge from what V1 says the audience knows. Do not invent identity, vocal traits, sounds, emotion, song title or viewing results. Audio not reviewed does not mean silence. Preserve recorded dialogue. Do not change timing, segmentation or format; do not generate a delivery file. Suggestions require a professional decision and the applicable style sheet.”
An editor-written follow-up for E2 would be: “Can the reviewer check this segment’s audio and image and describe the sound actually present, its source if identified and its relevance?” This is an internal note, not a sent message or confirmation of a door slam.
The subtle failure: a plausible label can still lack support
Compare the response with the table and revisit the quoted passages. Natural wording alone is insufficient. Check whether it keeps a name known only from the script, invents a voice’s gender without a description, treats the draft’s emotion as an observation or turns an outstanding review into an absence of sound. Correct the suggestion and adjust the prompt before repeating the trial if it fails.
Try a separate variant: replace only V1’s E2 note with “Audio reviewed: a short loud bang is audible off-screen; source unidentified. The reviewer considers it relevant.” The manual reference for this variant supports suggesting [short loud bang] for review, without attributing it to a door. Leave E1 and E3 unchanged; do not mix the two case versions. Final wording still depends on the style sheet and professional judgment.
These notes do not replace full viewing, synchronisation, readability or final-version checks. The proposal is limited to checking textual support for labels. It does not demonstrate video accessibility or a tool’s ability to hear the material correctly.
A first pilot to compare with manual review
Prepare a small set of fictional cases with manual references, including an unrevealed identity, an unidentified sound source and unreviewed audio. Record unsupported labels, omitted relevant information, incorrect citations and preparation plus review time. Compare the same cases handled manually and with assistant support; define in advance which errors prevent accepting a suggestion.
Return any label adding unrecorded information to review. Retain the original draft, response, corrections and decision owner. These are proposed pilot measures, without demonstrated accuracy, savings or accessibility outcomes.
When an integration may be worth assessing
The manual trial means copying fictional text and reviewing a response. A connected automation would require checking the team’s actual editor and formats, linking notes to the correct audiovisual version, retaining permissions and history and preventing an unapproved suggestion from being exported as an accepted delivery. Compatibility with subtitling software has not been verified.
Adaris develops custom software, CRM and automation with AI integrated case by case. A review sheet connected to your workflow could be assessed with an agreed scope and acceptance criteria. This is a development proposal, not an existing Adaris subtitling tool. If your problem is preparing script pronunciation notes for voiceover, the related guide covers that different task.
Related services
Sources
Are subtitling questions scattered across documents?
Tell us how you review labels and who approves delivery. We can assess a custom review sheet or integration using a fictional example. Commercial communication is in Spanish.