A Welsh-language advocate walked into a museum in Cardiff and asked whether the new audio guide was available in Welsh. The duty manager said no, but that English was available, and that the museum would love to add Welsh eventually. The advocate walked out. Three days later, the museum received a polite but firm letter from the Welsh Language Commissioner's office asking for a timeline.
This is a specific kind of pressure, and it lands on more museums than people realise. Any institution operating in a region with a recognised minority or heritage language — Catalonia, the Basque Country, Wales, Scotland, Ireland, Brittany, Friesland, parts of Switzerland, parts of Belgium, most of Latin America, the Pacific, many cities with diaspora communities — faces a version of this conversation. Historically, the museum's reply was some variation of "it's not financially feasible to produce a full recording in every minority language." The advocate walks out. The museum feels bad. Nothing changes for another three years.
The financial story underneath that reply has genuinely changed. Not completely, and not for every language. But enough that the old answer is no longer the right one.
The economics that kept minority languages out
For the entire history of audio guides until roughly 2024, adding a language meant producing a language. A studio session or two with native speakers. A script translation. Recording, editing, mastering. Synchronising to the existing tour structure. Deploying to the fleet. All-in, a minority language addition typically ran €8,000 to €25,000 per language per major content cycle.
For English, French, Spanish, German, Italian, Japanese, Mandarin — the top-tier languages at major tourist museums — this math works. The volume of visitors speaking those languages makes the production cost defensible. For a language with 600,000 speakers and a few thousand museum-visiting speakers per year, the per-visitor cost approaches absurdity. So the museums skipped it. Not out of malice, just arithmetic.
What the old arithmetic actually encoded was an accessibility hierarchy: big-market languages get served, minority languages don't. Institutions knew it was uncomfortable. They also knew it was the only math that worked. A few went ahead and funded specific minority-language recordings through grants or cultural partnerships. Most quietly didn't.
What AI changes
The underlying technical shift is that generating a voice track from text in a given language is no longer a studio job. It's a compute job. The per-language marginal cost of producing a usable audio track collapses from thousands of euros to pennies per stop.
For a museum, this looks like three practical changes.
All supported languages become available by default. When a platform supports 40 or 60 languages at the underlying model level, the museum gets all of them without paying per language. Catalan is there, Basque is there, Welsh is there, Maori is there. The question stops being "can we afford to add this language" and becomes "what's the quality like, and who reviews it before it goes live."
Quality review becomes the gating step instead of production cost. A good minority-language deployment still needs a fluent speaker — ideally a native speaker, ideally someone with domain knowledge — to review the AI output. That review can catch the mistranslations, the uncolloquial phrasings, the cultural missteps. The review typically runs a few hours per gallery, not a multi-week production. But it matters, and it's not optional.
Regional variants become tractable. "Spanish" isn't one language in practical terms. Castilian, Mexican, Rioplatense, Andean, Caribbean, and so on. For heritage sites where visitors expect their local variant — or where the museum represents a specific diaspora community — the ability to deploy a Mexican-specific voice alongside a Castilian one in the same cost envelope is genuinely new.
We covered the broader language tier question in AI audio guide languages supported. This article zooms in on what "rare-language support" actually looks like in practice.
The honest tier system
Not every minority language is equally well-served by AI today. The honest taxonomy we share with prospective museums:
Tier 1 — well-served minority languages. Catalan, Welsh, Scottish Gaelic, Irish, Basque, Galician, Luxembourgish, Icelandic, Faroese, Maltese, Frisian, Breton, Asturian, Maori, Hawaiian. Written corpora are large enough that translation quality is strong. Voice models exist and sound natural. A curator review is needed but the first draft is usable. Deploy with confidence and a review loop.
Tier 2 — serviceable but needs more curator work. Occitan, Romansh, Corsican, Sardinian, some Saami languages, Quechua (major dialects), Guaraní, Nahuatl, Yiddish (Litvish and Poylish variants both supported but inconsistent). Translation quality is lower. Voice models exist but may sound uncanny or use wrong prosody. Expect a longer review loop and more per-stop editing. Often the right answer is a bilingual deployment where the minority language runs alongside a major-language parallel.
Tier 3 — limited support today. Many indigenous languages of the Americas beyond the ones listed above, smaller African languages, many Pacific languages beyond Maori and Hawaiian, Tibetan, many South Asian regional languages, regional dialects without standard written form. The AI either produces poor output or isn't trained on the language at all. The real path here is often a traditional recording with AI as a supporting tool — or a community-led custom voice project where speakers record the track and the AI handles the rest of the platform.
Tier 4 — not currently supported. A long tail of languages with very small digital corpora, oral-only traditions, or no standardised written form. AI tools won't serve these today. What matters for a museum in this position is to design the platform so that when support arrives — and it often arrives within a year or two — the language can be added without re-procuring the system.
The tier is moving. A language that was tier 3 two years ago may be tier 2 now. Any vendor claiming uniform quality across all 40+ supported languages is either not testing carefully or is hoping you won't.
How to actually deploy a rare-language guide
The procedural shape we've seen work:
Decide the scope. Full minority-language coverage of the permanent collection, one curated highlights tour, or a specific temporary exhibition. Start narrower rather than broader. A 15-stop highlights tour in Welsh, deployed well, is more useful than a full 150-stop tour in Welsh that nobody reviewed.
Recruit a reviewer. A fluent speaker with context. For Welsh, that might be a curator who is a Welsh speaker, or a partner from the Welsh Language Commissioner's network, or a community advisor from a local cultural organization. Pay them if they aren't staff. The review work is real.
Generate the first draft. The AI produces the full tour in the target language, in a voice chosen for fit.
Review in place. The reviewer listens through the guide at the actual museum, with the actual objects in front of them. A lot of translation errors aren't visible on paper — they're visible when you realise the voice is referring to something the object doesn't actually show, or using a register that's wrong for the setting.
Iterate. Revisions go back through the platform. A well-designed system regenerates the specific segments that changed, not the whole tour.
Go live with a version number. Minority-language guides benefit from being explicitly versioned. "Welsh v1, reviewed by X on date Y" — visible somewhere in the app metadata. This is both an accountability mechanism and a cultural signal. The museum is saying: we took this seriously, we versioned it, we will update it.
The reputational and strategic case
The financial case is easy because the marginal cost is close to zero. The better case is the one finance teams sometimes miss: offering minority-language coverage does real reputational work that shows up in grants, partnerships, and media coverage.
Cultural funders reward demonstrable commitment to linguistic diversity. The Welsh Arts Council, the Basque government, the Irish Department of Culture, the Catalan Institute, UNESCO — and their counterparts across dozens of regions — have explicit funding priorities around heritage and minority language preservation. A museum that can demonstrably serve visitors in the minority language of its region has a better case in those conversations than one that can't.
More practically: communities of heritage-language speakers often visit museums specifically to see their language represented. A museum that does this well becomes a place that community feels welcome in. That's not a small thing for the long-term relationship with the local audience, the school groups, the intergenerational visits. We covered the community angle in more depth in multilingual museum audio guide.
A practical closing note
Platforms like Musa price on per-interaction or revenue share, which is the specific shift that makes minority-language deployment economically sensible. A Welsh-language guide at a museum with a few hundred Welsh-speaking visitors a year stops being a fixed-cost project that has to be justified and becomes a near-zero-marginal-cost service layer that the museum can simply offer. The Welsh speakers who use it pay for what they use. The museum isn't amortising a studio bill.
The museums doing this well right now tend to share three things: an institutional decision that minority-language coverage is part of the visitor experience rather than a checkbox, a named reviewer (sometimes paid, sometimes a community advisor) who owns the quality of the minority-language content, and a willingness to deploy in a tiered way — start with one tour, learn, expand. None of those things are expensive. All of them are choices.
If you're working through this question for a specific region or language, we're happy to walk through what's realistically available — including the honest tier your target language sits in, not the one the vendor pitch would prefer you believed.