Translate Alt Text & Captions Accessibly

How do you translate alt text and captions accurately for accessibility? Begin with the function of the image, control, graph or audio segment, then preserve its meaning for someone who cannot see or hear the original. Alt text translation should support an action when the image is a button, explain a meaningful result when it is a chart and remain empty when an image is purely decorative. This is a practical accessibility translation guide, not a claim that every translated page is WCAG conformant.

Readers searching for how to localize image alt text, accessible captions and subtitles, screen-reader button labels, translated transcripts and multilingual image descriptions often need more than a glossary. They need a method for checking the user experience, including keyboard operation, timed media, accessible names, data alternatives and reading order. Translating only what appears visually can omit an entire stream of information that assists people using screen readers.

Did you know W3C’s Web Accessibility Initiative recommends different approaches for decorative, functional, informative and complex images? WAI also distinguishes transcripts, captions and other media alternatives. This article follows the accessible-alternative job rather than the broader software product localization owner. Fictional examples demonstrate how the same visible design can produce a very different experience when its accompanying words are mistranslated.

The 50-second answer

Translate the purpose, evidence and usable action, not the artwork alone. A decorative image may have empty alt text; a functional image needs its function; a complex chart often needs a separate text explanation; captions need spoken words plus relevant sounds in time with the media. Verify the translated page through keyboard and assistive technology. The final question is whether the user can accomplish the same task and understand the same important information.

The visible failure

A learner opens a multilingual exam practice page. The visible ‘Submit answer’ label is translated, but the screen reader announces ‘blue arrow’. A second page shows a graph falling from 80 to 40 but the accessible text says only ‘colourful bars’. Both pages look translated; both lose their essential job for users who cannot rely on the image. The repair starts by naming the action or evidence instead of merely describing appearance.

This article owns the translation of accessible alternatives and timed media, not the full localization of every software interface. Examples are invented teaching scenes. WCAG conformance also depends on many implementation details outside wording; a successful text translation alone does not confer certification.

First principles: how accessibility changes translation

1. Alt text expresses the role of an image

If an icon functions as a link or action, the translated accessible name must tell a user what it does. The shape is secondary. An informative image needs the information it contributes, not a decorative inventory. In a multilingual context, a phrase that looks correct on screen can still fail in a link list, a screen reader’s speech stream or a caption track. The translator should therefore record both the visible source and the nonvisual purpose. If the user cannot infer the expected action from the alternative, revisit the source semantics before polishing the words.

Test this principle using two versions of the same scene. In version one, remove the image or mute the sound and rely only on the translated alternative. In version two, use the complete original experience. Compare the decisions a learner can make. Where the experiences differ, identify precisely what information is missing, rather than extending the text with generic description.

2. Decorative does not mean forgotten

An empty alternative can be the correct choice for a purely decorative image. Translating a decorative flourish into a full spoken sentence adds cognitive load without meaningful benefit. In a multilingual context, a phrase that looks correct on screen can still fail in a link list, a screen reader’s speech stream or a caption track. The translator should therefore record both the visible source and the nonvisual purpose. If the user cannot infer the expected action from the alternative, revisit the source semantics before polishing the words.

3. Captions must remain synchronized

Accessible captions convey meaningful spoken words and relevant audio cues aligned to media. Merely translating a script does not guarantee a usable caption track if cues appear at the wrong time. In a multilingual context, a phrase that looks correct on screen can still fail in a link list, a screen reader’s speech stream or a caption track. The translator should therefore record both the visible source and the nonvisual purpose. If the user cannot infer the expected action from the alternative, revisit the source semantics before polishing the words.

4. Transcripts have different scope

A transcript can include speech and descriptions of relevant visuals. A bare dialogue script may miss a gesture, written sign or silently displayed result that is crucial to the lesson. In a multilingual context, a phrase that looks correct on screen can still fail in a link list, a screen reader’s speech stream or a caption track. The translator should therefore record both the visible source and the nonvisual purpose. If the user cannot infer the expected action from the alternative, revisit the source semantics before polishing the words.

5. Audio description gives visual information a route

When information exists only in video pictures, a description or suitable media alternative can convey it. The translated text must be timed and chosen to avoid competing with essential dialogue. In a multilingual context, a phrase that looks correct on screen can still fail in a link list, a screen reader’s speech stream or a caption track. The translator should therefore record both the visible source and the nonvisual purpose. If the user cannot infer the expected action from the alternative, revisit the source semantics before polishing the words.

6. Interactive states carry meaning

A button may be expanded, disabled, required, selected or invalid. Translating its name without preserving state and help text can leave a user unable to complete a task. In a multilingual context, a phrase that looks correct on screen can still fail in a link list, a screen reader’s speech stream or a caption track. The translator should therefore record both the visible source and the nonvisual purpose. If the user cannot infer the expected action from the alternative, revisit the source semantics before polishing the words.

7. Complex graphs need more than a few adjectives

A chart may have axis units, trends, uncertainty and exceptional points. A concise alt plus adjacent detailed explanation or data table can communicate more than a long list of colours. In a multilingual context, a phrase that looks correct on screen can still fail in a link list, a screen reader’s speech stream or a caption track. The translator should therefore record both the visible source and the nonvisual purpose. If the user cannot infer the expected action from the alternative, revisit the source semantics before polishing the words.

8. Language and directionality affect speech

A mixed-language page may require correct language metadata so names and phrases are pronounced appropriately. A right-to-left interface may also reorder mixed scripts if identifiers are not protected. In a multilingual context, a phrase that looks correct on screen can still fail in a link list, a screen reader’s speech stream or a caption track. The translator should therefore record both the visible source and the nonvisual purpose. If the user cannot infer the expected action from the alternative, revisit the source semantics before polishing the words.

9. Meaningful order matters

When navigation, step instructions or table headings are spoken in the wrong order, even accurate strings can become misleading. Structure and text must be tested together. In a multilingual context, a phrase that looks correct on screen can still fail in a link list, a screen reader’s speech stream or a caption track. The translator should therefore record both the visible source and the nonvisual purpose. If the user cannot infer the expected action from the alternative, revisit the source semantics before polishing the words.

10. Task completion is the strongest final test

A translated label succeeds when the intended reader can identify the action and finish the task without guessing from colour, visual position or the untranslated page. In a multilingual context, a phrase that looks correct on screen can still fail in a link list, a screen reader’s speech stream or a caption track. The translator should therefore record both the visible source and the nonvisual purpose. If the user cannot infer the expected action from the alternative, revisit the source semantics before polishing the words.

A ten-step translation and accessibility review method

Step 1: Find the user task

Identify the action or information the nonvisual or nonauditory reader needs. Make the task a complete sentence before drafting any alternative. For a concrete example, consider decorative divider: Floral ornament separating paragraphs. The draft must preserve the distinction between the medium and the task. Record what the source actually establishes, what the user needs to do and what should remain silent. Then test in the target language rather than judging the English string in isolation.

Quality gate: compare the alternative with its original context, ask another person to describe its purpose without seeing the screen and test whether a changed source would produce an appropriately changed translation. If the same target sentence would be used for opposite states or different actions, it lacks the necessary specificity. Repair that exact gap and document the change.

Step 2: Classify the content

Decide whether each image is decorative, informative, functional, textual or complex. Choose the corresponding accessible alternative. For a concrete example, consider search magnifier: Magnifying glass button opens search. The draft must preserve the distinction between the medium and the task. Record what the source actually establishes, what the user needs to do and what should remain silent. Then test in the target language rather than judging the English string in isolation.

Step 3: Map sources and relationships

Record image context, audio clip, dialogue timestamps, button state, route order, nearby text and existing semantics. For a concrete example, consider informative photograph: Photograph shows a cracked walkway after flooding. The draft must preserve the distinction between the medium and the task. Record what the source actually establishes, what the user needs to do and what should remain silent. Then test in the target language rather than judging the English string in isolation.

Step 4: Protect critical data

List all numbers, names, units, codes, warnings and cause-and-effect links that must survive translation. For a concrete example, consider step-free campus map: Map routes wheelchair users through north entrance. The draft must preserve the distinction between the medium and the task. Record what the source actually establishes, what the user needs to do and what should remain silent. Then test in the target language rather than judging the English string in isolation.

Step 5: Write for spoken clarity

Prefer concise natural accessible names and plain descriptions. Avoid filling essential labels with adjectives that do not help a task. For a concrete example, consider email error: Input rejected because email lacks @. The draft must preserve the distinction between the medium and the task. Record what the source actually establishes, what the user needs to do and what should remain silent. Then test in the target language rather than judging the English string in isolation.

Step 6: Check redundancy

Do not repeat adjacent real text unnecessarily. Do not remove a meaningful image description simply because its caption looks similar. For a concrete example, consider speaker change: Teacher and student alternate lines. The draft must preserve the distinction between the medium and the task. Record what the source actually establishes, what the user needs to do and what should remain silent. Then test in the target language rather than judging the English string in isolation.

Step 7: Synchronize media

Verify caption timing, speaker attribution and relevant sound cues against the actual recording at usable speed. For a concrete example, consider caption time: Warning spoken at 12 seconds. The draft must preserve the distinction between the medium and the task. Record what the source actually establishes, what the user needs to do and what should remain silent. Then test in the target language rather than judging the English string in isolation.

Step 8: Inspect structure

Test focus order, form label association, language metadata and right-to-left display on the rendered page. For a concrete example, consider transcript of visual experiment: Video silently shows gauge reaching 12. The draft must preserve the distinction between the medium and the task. Record what the source actually establishes, what the user needs to do and what should remain silent. Then test in the target language rather than judging the English string in isolation.

Step 9: Run the task independently

Have a reviewer use the translated experience without the source screen. Ask which choices were available and what evidence they received. For a concrete example, consider mixed-language title: French book title appears in English page. The draft must preserve the distinction between the medium and the task. Record what the source actually establishes, what the user needs to do and what should remain silent. Then test in the target language rather than judging the English string in isolation.

Step 10: Release with change control

Keep language, media revision, reviewer notes and unresolved questions so a new source clip or redesigned button triggers an appropriate update. For a concrete example, consider automatic caption error: Speaker says fifteen, recognizer writes fifty. The draft must preserve the distinction between the medium and the task. Record what the source actually establishes, what the user needs to do and what should remain silent. Then test in the target language rather than judging the English string in isolation.

Twenty worked examples of accessible translation

The following records are original fictional exercises. Source means the teaching situation rather than a quotation from a real product. Misleading versions show errors intentionally. Safer renderings are illustrative English descriptions; the final target-language wording must be tested in its actual interface.

Worked case 1: Decorative divider

Source: Floral ornament separating paragraphs. Misleading translation: Elaborate photograph of elegant leaves. Safer explanatory rendering: Empty alt text because the image conveys no new content.

The critical distinction is this: Adding descriptive speech to decoration interrupts navigation. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, read the page with the ornament omitted and see whether any information disappears. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Now transfer the method: imagine a second version in which the important field changes. The translated alternative should change in that precise place while retaining the rest of its meaning. This protects against memorizing a polished phrase that is no longer true. Record the original wording, the changed field and the reason for the change. Good accessibility translation is a repeatable, testable relation between two experiences rather than a single fluent sentence.

Worked case 2: Submit icon

Source: Unlabelled arrow button submits a quiz. Misleading translation: Blue triangular arrow. Safer explanatory rendering: Submit answer.

The critical distinction is this: The control’s function rather than its geometry is relevant. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, navigate to the button using speech output and submit the answer. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 3: Search magnifier

Source: Magnifying glass button opens search. Misleading translation: Little circle with handle. Safer explanatory rendering: Search.

The critical distinction is this: A person who cannot see needs the operation. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, list page buttons using assistive technology and identify Search. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 4: Home logo

Source: Brand logo linked to the homepage. Misleading translation: Picture of the company emblem. Safer explanatory rendering: Go to homepage.

The critical distinction is this: A linked image provides navigation, not merely brand decoration. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, inspect the link destination without surrounding visual context. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 5: Informative photograph

Source: Photograph shows a cracked walkway after flooding. Misleading translation: A dangerous, ruined place nobody should visit. Safer explanatory rendering: Cracks visible in the walkway after flooding.

The critical distinction is this: A description must not invent an unsupported prohibition. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, compare image evidence with words of the translation. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 6: Graph reduction

Source: Graph shows attendance falling from 80 to 40. Misleading translation: Colourful bars in two shades. Safer explanatory rendering: Attendance fell from 80 to 40 across the shown periods.

The critical distinction is this: Trend and quantities carry the instructional job. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, check axis units, period order and accompanying data. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 7: Step-free campus map

Source: Map routes wheelchair users through north entrance. Misleading translation: Building floor plan. Safer explanatory rendering: Step-free route: use north entrance and lift.

The critical distinction is this: Generic descriptions cannot replace practical navigation. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, try the route without seeing the picture. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 8: Required field

Source: Asterisk beside name marks a required field. Misleading translation: A red star. Safer explanatory rendering: Name, required.

The critical distinction is this: Shape or colour should not be the only required-field cue. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, tab through the form and listen for required status. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 9: Email error

Source: Input rejected because email lacks @. Misleading translation: Invalid entry. Safer explanatory rendering: Enter an email address in the required format.

The critical distinction is this: Helpful errors identify the problem and next action. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, try another invalid input and confirm the explanation still helps. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 10: Disclosure state

Source: Course details currently collapsed. Misleading translation: Read more. Safer explanatory rendering: Course details, collapsed; expand for information.

The critical distinction is this: The reader must understand current state. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, test control before and after opening it. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 11: Speaker change

Source: Teacher and student alternate lines. Misleading translation: One unlabeled voice. Safer explanatory rendering: Identify teacher and student where the distinction matters.

The critical distinction is this: Speech attribution changes who made the claim. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, listen to captions without looking at faces. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 12: Alarm cue

Source: Alarm rings before instructions. Misleading translation: Nothing until next spoken sentence. Safer explanatory rendering: [Alarm sounding] ahead of the response.

The critical distinction is this: Important non-speech audio explains action. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, mute the clip and use captions to identify the cause. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 13: Caption time

Source: Warning spoken at 12 seconds. Misleading translation: Warning subtitle appears at 27 seconds. Safer explanatory rendering: Align translated warning to its actual moment in audio.

The critical distinction is this: Mistimed instruction can be confusing or unsafe. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, play at ordinary and reduced speed. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 14: Subtitles are not complete captions

Source: Translated speech omits siren. Misleading translation: Full captions completed. Safer explanatory rendering: Translated subtitles exist; add siren cue for accessibility captions.

The critical distinction is this: Deliverable scope should not be overstated. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, compare the two tracks with sound off. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 15: Transcript of visual experiment

Source: Video silently shows gauge reaching 12. Misleading translation: Transcript includes only dialogue. Safer explanatory rendering: Transcript describes the displayed gauge value 12.

The critical distinction is this: Visual evidence needs a textual route. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, read the transcript without the video. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 16: Audio described sign

Source: Video sign reads: west exit closed. Misleading translation: Person stands near door. Safer explanatory rendering: Sign: West exit closed; use east stairs.

The critical distinction is this: The missing sign changes how someone navigates. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, check the description timing against audio. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 17: Mixed-language title

Source: French book title appears in English page. Misleading translation: Read all words as though English. Safer explanatory rendering: Preserve title and French-language context.

The critical distinction is this: Speech pronunciation can be garbled by wrong language tagging. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, listen to mixed-language phrase using assistive output. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 18: Right-to-left code

Source: Arabic label followed by LTR model A-107. Misleading translation: Reverse model number characters. Safer explanatory rendering: Preserve model A-107 in stable logical order.

The critical distinction is this: Bidirectional layout can break identity strings. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, copy and paste model code from rendered page. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 19: Automatic caption error

Source: Speaker says fifteen, recognizer writes fifty. Misleading translation: Fifty samples tested. Safer explanatory rendering: Fifteen samples tested.

The critical distinction is this: A single mistranscribed number changes the claim. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, listen to original audio and compare digits. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Worked case 20: Translated PDF reading order

Source: Visually formatted two columns read in alternating lines. Misleading translation: Accessible finished PDF. Safer explanatory rendering: Preserve logical tagged reading order.

The critical distinction is this: Visual perfection cannot substitute for structured reading. If a reviewer compares only individual words, the misleading version may seem close to the source. But a translation must also preserve the user’s ability to infer what matters. In this example the visual or auditory content performs a particular job: identifying an operation, explaining evidence, signaling a state or providing a navigable path. A correct alternative preserves that job without claiming more than the original establishes.

To verify the repair, read all content linearly through screen-reader output. Ask whether a reader who does not see the original could perform the relevant task or explain the important fact. If that is not possible, inspect the source, the accessible name, surrounding text and structural markup together. A semantic fault cannot always be fixed by changing a noun; sometimes a separate data table, label, transcript or correctly timed caption cue is required.

Twelve common failures and their repairs

1. Meaning replaced with visual texture

A ‘blue triangle’ appears where the real purpose is ‘Submit answer’. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit decorative divider: Adding descriptive speech to decoration interrupts navigation. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

2. Decorative noise

Every separator receives a long translated description even when content is unchanged. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit submit icon: The control’s function rather than its geometry is relevant. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

3. Unexplained graph

A numerical trend is translated as ‘a chart’. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit search magnifier: A person who cannot see needs the operation. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

4. Dropped sound cue

A siren or doorbell explaining the scene vanishes from captions. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit home logo: A linked image provides navigation, not merely brand decoration. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

5. Late captions

Critical instructions display only after the described action is over. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit informative photograph: A description must not invent an unsupported prohibition. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

6. Speaker attribution loss

A quote is assigned to the wrong participant. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit graph reduction: Trend and quantities carry the instructional job. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

7. Visible-only validation

The translator never checks accessible names, state or keyboard order. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit step-free campus map: Generic descriptions cannot replace practical navigation. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

8. Unreviewed automatic transcription

Numbers or specialist words are accepted because the generated sentence sounds natural. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit required field: Shape or colour should not be the only required-field cue. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

9. Broken mixed-language pronunciation

Incorrect language metadata makes a meaningful name unrecognizable in speech. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit email error: Helpful errors identify the problem and next action. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

10. Directionality damage

An LTR model code is reordered in an RTL sentence. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit disclosure state: The reader must understand current state. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

11. Reading-order collapse

An exported file mixes two columns in a non-sensical spoken sequence. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit speaker change: Speech attribution changes who made the claim. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

12. Role confusion

An audio description, transcript and caption track are treated as equivalent products. Repair this by identifying the user’s missing action or information and comparing it with the authoritative source. For a practical check, revisit alarm cue: Important non-speech audio explains action. Do not make the target text longer just to sound thorough. Add only what the use case actually needs, and verify that the added material does not fabricate a state, number, visual fact or permission.

Practice and transfer: twelve short tasks

Practice 1: Decorative divider

Try this without looking back: the source is ‘Floral ornament separating paragraphs’, and an initial translator wrote ‘Elaborate photograph of elegant leaves’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘Empty alt text because the image conveys no new content’. Adding descriptive speech to decoration interrupts navigation. Your test is to read the page with the ornament omitted and see whether any information disappears. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Practice 2: Submit icon

Try this without looking back: the source is ‘Unlabelled arrow button submits a quiz’, and an initial translator wrote ‘Blue triangular arrow’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘Submit answer’. The control’s function rather than its geometry is relevant. Your test is to navigate to the button using speech output and submit the answer. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Practice 3: Search magnifier

Try this without looking back: the source is ‘Magnifying glass button opens search’, and an initial translator wrote ‘Little circle with handle’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘Search’. A person who cannot see needs the operation. Your test is to list page buttons using assistive technology and identify Search. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Practice 4: Home logo

Try this without looking back: the source is ‘Brand logo linked to the homepage’, and an initial translator wrote ‘Picture of the company emblem’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘Go to homepage’. A linked image provides navigation, not merely brand decoration. Your test is to inspect the link destination without surrounding visual context. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Practice 5: Informative photograph

Try this without looking back: the source is ‘Photograph shows a cracked walkway after flooding’, and an initial translator wrote ‘A dangerous, ruined place nobody should visit’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘Cracks visible in the walkway after flooding’. A description must not invent an unsupported prohibition. Your test is to compare image evidence with words of the translation. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Practice 6: Graph reduction

Try this without looking back: the source is ‘Graph shows attendance falling from 80 to 40’, and an initial translator wrote ‘Colourful bars in two shades’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘Attendance fell from 80 to 40 across the shown periods’. Trend and quantities carry the instructional job. Your test is to check axis units, period order and accompanying data. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Practice 7: Step-free campus map

Try this without looking back: the source is ‘Map routes wheelchair users through north entrance’, and an initial translator wrote ‘Building floor plan’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘Step-free route: use north entrance and lift’. Generic descriptions cannot replace practical navigation. Your test is to try the route without seeing the picture. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Practice 8: Required field

Try this without looking back: the source is ‘Asterisk beside name marks a required field’, and an initial translator wrote ‘A red star’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘Name, required’. Shape or colour should not be the only required-field cue. Your test is to tab through the form and listen for required status. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Practice 9: Email error

Try this without looking back: the source is ‘Input rejected because email lacks @’, and an initial translator wrote ‘Invalid entry’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘Enter an email address in the required format’. Helpful errors identify the problem and next action. Your test is to try another invalid input and confirm the explanation still helps. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Practice 10: Disclosure state

Try this without looking back: the source is ‘Course details currently collapsed’, and an initial translator wrote ‘Read more’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘Course details, collapsed; expand for information’. The reader must understand current state. Your test is to test control before and after opening it. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Practice 11: Speaker change

Try this without looking back: the source is ‘Teacher and student alternate lines’, and an initial translator wrote ‘One unlabeled voice’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘Identify teacher and student where the distinction matters’. Speech attribution changes who made the claim. Your test is to listen to captions without looking at faces. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Practice 12: Alarm cue

Try this without looking back: the source is ‘Alarm rings before instructions’, and an initial translator wrote ‘Nothing until next spoken sentence’. Identify what the user cannot do or know from the first rendering. Form an accessible alternative that is clear in the target context. Then explain whether any part of the source should remain untranslated, be left empty, or be supplied elsewhere as a longer description.

A suitable repair is ‘[Alarm sounding] ahead of the response’. Important non-speech audio explains action. Your test is to mute the clip and use captions to identify the cause. Afterward create a new example with a different number, role or button state; a durable method must still work. The goal is independent recognition of accessible meaning, not merely copying the answer key.

Frequently asked questions

Does every image need descriptive alt text?

No. Decorative or redundant images often need empty alt text; informative and functional images need alternatives selected for their particular context. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

Should alt text always name colours?

Only when colour is materially important. Describe the meaning it carries, or preserve that meaning through another accessible channel. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

Is a translated transcript the same as captions?

No. Captions require synchronization and meaningful audio coverage; a transcript is a different reading experience. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

What does a functional icon need?

A name that tells the reader what the button or link does, ideally aligned with the visible operation. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

Can automatic captions go live without review?

Important numbers, names, speaker changes and relevant sounds should be verified against the original recording. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

Do charts need long alt text?

Complex charts often need a concise accessible alternative and a separate detailed explanation or data representation. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

What if the visual control uses a single glyph?

The glyph may stay visible, but the assistive label must identify the action. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

Why preserve focus and reading order?

A keyboard or screen-reader user needs a logical sequence to understand instructions and move between controls. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

How do language tags affect translation?

Correct metadata can help speech output pronounce embedded languages and distinguish different content segments. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

Do captions need non-speech sounds?

Yes, when such sounds matter to understanding the media. Irrelevant background noise need not dominate. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

Can a decorative image receive nonempty alt text?

If it conveys a genuine function or information in context, it is not merely decorative. Determine the actual role. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

What is the acceptance test?

Ask whether someone can complete the target-language task using keyboard and assistive technology without visual guessing. A useful check is to focus on the user decision: what would someone using the translated alternative believe, select or understand? Preserve that decision path with a documented reason for your wording. When a technical implementation is involved, seek review from an accessibility specialist rather than claiming compliance from wording alone.

Sources and connected learning routes

W3C WAI: image alt decision tree; W3C WAI: transcripts; W3C: Understanding WCAG 2.2.

Next, read Vocabulary Learning Hub, How English Works, school rubric translation and rights and attribution translation. Those guides own separate teaching jobs. This one remains the accessible-alternatives and captions owner.

Final exercise: choose a short real page containing an image, a functional control and a video. Build a source-to-target decision table, perform the keyboard and audio tests, and explain which changes improved actual task completion. Consult the W3C guidance again if there is disagreement over whether a graphic is informative or decorative. Reliable translation grows through visible checks, not guesses.

Top Ways to Translate Correctly | Translate Alt Text, Captions and Accessible Instructions Without Losing User Actions