How to Create a Multilingual Museum Audio Guide

A museum audio guide fails in a particular way. The recording can sound polished, the QR code can scan, and the script can still leave a visitor standing in front of the wrong case, hearing an object described in words that no longer match its label.

Vois gives a museum team a local place to turn an approved script into reviewed audio, with 100+ voices and Pro support for 600+ languages. The studio can preserve a language version, audition pronunciation in context, arrange the audio, and export it without scattering the production work across several services. It does not replace curatorial, cultural, accessibility, or language review.

The cure is not a longer script. It is a better system. Treat every stop as a small, versioned publication with a stable identity, a source record, named reviewers, language variants, and a clear handoff to the device in a visitor's hand. That makes a multilingual audio tour maintainable when exhibitions change.

Start a museum audio guide with a stop ledger, not a route

A route is temporary. Objects move between galleries, galleries close, and visitors rarely follow a prescribed path. Give every audio stop a route-independent ID before anyone writes narration.

A small stop ledger can live in a spreadsheet, collection system, or content repository. What matters is that it has one row per stop and a history people can inspect. Include:

• a stable stop ID, such as ARC-0142
• accession or collection reference, where appropriate
• displayed title, maker or culture field, and current gallery context
• source-script version and approved date
• language, reviewer, and audio-file version
• pronunciation record and any visitor-facing content note
• QR or app destination that resolves from the stable ID

Do not name recordings stop-03-final.mp3. That name only makes sense on the day someone exports it. A file such as ARC-0142_en_v04.wav says which object, language, and revision it belongs to. The same ID can serve a QR code, an app deep link, a printed guide, and a staff support note without pretending the visitor must stand in one exact place.

For a room introduction, use a separate stable ID such as GALLERY-ARC-INTRO. Keep object narration and room orientation distinct. If an object moves, its interpretation can travel with it while the gallery introduction changes on its own schedule.

Write object-level scripts that visitors can follow

An audio guide should add something the label cannot. Begin each script by identifying the object in the visitor's view, then offer one carefully sourced interpretive thread. A short visual anchor helps a visitor confirm they found the right thing: a material, a visible detail, the position within a case, or the relationship to a neighboring object.

Build the source script as a record, not just prose. We recommend fields for:

1. Object identification: the displayed title, date or period if approved for display, and stop ID.
2. Interpretation: the specific curatorial point this stop is making, with source notes for factual claims and quotations.
3. Description and visitor context: details that help a listener orient themselves, plus any content note the museum decides is appropriate.
4. Production notes: pronunciation, pacing, language-specific direction, reviewer names, and version state.

Keep the distinction between evidence and interpretation visible. A curator may approve a concise statement of what is known, while a different sentence frames a scholarly interpretation or introduces uncertainty. Do not flatten those into one authoritative-sounding monologue. When a collection record changes, the team should be able to find the exact sentence and language versions that need another look.

Read the draft aloud before translation. Phrases that work on a wall label can become tangled when heard. Short sentences, intentional paragraph breaks, and specific nouns make the narration easier to follow without resorting to a simplified version of the collection story.

Put curatorial and cultural review before translation

Translation cannot repair an unclear or unapproved source script. Lock the source language only after a curatorial reviewer has checked factual claims, dates, attributions, quotations, object names, and any unresolved uncertainty.

Then give each target-language script its own review path. A language reviewer can assess meaning, register, grammar, pronunciation, and whether a phrase would be understood as intended. When an object concerns a living, source, or otherwise closely connected community, involve the people and relationships your institution already trusts for that material. They should not be asked to rubber-stamp a finished recording. Bring them into the review while the wording can still change.

The ICOM Code of Ethics for Museums is a useful professional reference for this work. It frames museum practice around documented collections, professional responsibility, and respect for communities. It does not turn cultural review into a one-size-fits-all checklist, which is exactly the point. Define who can approve what for your collection, then make that decision traceable in the stop ledger.

A practical status flow is simple: draft -> curatorial approved -> translation reviewed -> ready for audio -> audio approved -> released. A translated recording is not finished because it exists. It is ready only when its language and interpretation have both been reviewed.

Create a pronunciation record before you generate audio

Names are where a guide earns or loses trust quickly. Build a pronunciation record alongside the script instead of correcting names one export at a time. For each name, place, term, or collection-specific word, record the canonical spelling, a plain-language pronunciation cue, applicable language, context, and reviewer.

A note like "stress the second syllable" can be more useful to a narrator than a vague request to "say it naturally." Keep the context too. A word can be pronounced differently when it is a family name, a place, or a term in another language.

Vois includes a pronunciation dictionary that can hold approved entries. If your team uses an AI agent to prepare the reviewed CSV for Vois, give it a narrow instruction instead of a raw command:

Prompt your agent

First, check whether Vois is available for this project and report the app or CLI status plus any import-readiness issue. Then prepare the reviewed entries in museum-terms.csv for the pronunciation dictionary, using only word, pronunciation, and context. Do not alter source scripts, generate audio, or replace an existing entry without flagging it. Report missing context, duplicate terms, and rows that need language review. Do not import anything until I approve the proposed changes.

Review and approve the update

1. Compare the reported Vois status and proposed rows with the curatorial and language-approved pronunciation record.
2. Resolve every duplicate, missing context, language-specific question, and import-readiness issue.
3. Approve a scoped import only after the review is complete, then keep the agent's status and import result with the stop ledger.
4. Audition every changed name in its actual sentence and mark the language version ready only after the listener approves it.

The approval record still belongs with the museum's content workflow. A correct dictionary entry can sound wrong in a particular voice or language, so hear it in the finished sentence rather than treating an import as the final gate.

Localize the source record, not the finished voice file

A multilingual audio tour is easiest to maintain when every language is a child of the same approved source stop. Store the source script, then create a language record for en, es, fr, or any other target language. Each record should carry the same stop ID and its own translator, reviewer, pronunciation notes, audio asset, and approval state.

Do not assume an English script can be read word-for-word in another language. The target-language reviewer should be able to reshape sentence order, examples, and explanatory context while preserving the approved meaning. Keep decisions in notes so an update does not force the next translator to rediscover why a phrase was chosen.

If you produce the recordings in Vois, audition voices in the target language rather than treating a voice choice as a global default. Vois offers 100+ voices, and Pro supports 600+ marketed languages. Those options make production flexible, but they do not replace a reviewer who can listen for clarity, tone, and correct pronunciation. For broader planning, see our guide to multilingual content at scale.

Make accessibility and QR handoff part of the stop design

An audio file alone is not an accessible visitor experience. Plan the handoff as one complete stop: a printed or on-screen object ID, a title, language selector, audio player, matching transcript, and a clear way to return to the guide index. Offer a description when it supports the specific visitor experience you are designing, and test it with the people and practices your museum uses for accessibility review.

The Smithsonian Guidelines for Accessible Exhibition Design discuss presenting essential information through more than one format. The Web Content Accessibility Guidelines 2.2 are also a practical reference when the guide lives on the web or in an app. Neither resource substitutes for testing the actual handoff in the actual gallery.

A QR code should resolve to a durable page, not directly to a fragile media filename. For example, museum.example/audio/ARC-0142 can present the current title, available languages, transcript, and audio asset. If version four replaces version three, the QR stays put. Test the code on the devices and network conditions visitors will encounter, then test the fallback: can someone enter the stop ID manually or find the same recording from the guide index?

Master and check the audio as a set

Mastering is the final polish, not a shortcut around review. Choose a delivery format with the app or guide provider, then use the same approved production settings across a language set where that makes sense. Keep the uncompressed review asset separate from the delivery asset so you can regenerate a compressed file without reopening a language decision.

The audio reviewer should listen for more than volume. Use a stop-by-stop checklist:

• the spoken object ID, title, and language match the ledger
• the audio begins and ends cleanly, without a clipped word or accidental silence
• approved pronunciations work in context
• the recording matches the reviewed script and transcript
• the voice is clear in the target delivery player
• the QR or app destination opens the intended, current stop

Listen in a quiet review setting and in a representative visitor setting if possible. A file that sounds fine at a desk may behave differently through the device, headphones, or speaker arrangement used in the gallery.

Use an update workflow that preserves the trail

Collections interpretation changes. A date is revised, a loan returns, a label is rewritten, or a translator identifies a better term. The update workflow should be predictable: change the source record, identify affected language records, reopen only the gates that changed, generate new audio, complete audio QA, and hand the approved asset to the QR or app owner.

Do not overwrite the old record without a version note. Mark what changed, why it changed, who approved it, and which audio assets replaced it. Then verify the live stop after the external handoff. This is less glamorous than recording day, but it prevents a visitor from discovering that the beautiful new guide is quietly narrating yesterday's exhibition.

A multilingual museum audio guide works when every visitor can find the right object, hear a reviewed story in the chosen language, and return tomorrow to a guide the team can still maintain. For a local workflow that keeps language versions together, see multilingual generation, then Get started with one thoroughly reviewed stop.

Sources

• ICOM Code of Ethics for Museums
• Smithsonian Guidelines for Accessible Exhibition Design
• Web Content Accessibility Guidelines (WCAG) 2.2

The Vois Team