Can You Make an AI Audio Tour Guide? Scripts, Voices and Licensing

Write a fact-checked script from sources you may use, turn it into audio with a text-to-speech service whose plan includes a commercial licence, and host the files behind QR codes or in an app. Licensing applies to every layer: the script's sources, the voice, any music, and any cloned voice, which needs the speaker's written consent.

The voice is the cheap part. A 12-stop route costs a few dollars in speech generation and takes minutes to render. What takes days is research, fact-checking and paperwork, and those are where a tour business gets hurt: a wrong date read aloud to a hundred visitors, a music track the licence never covered, a guide's cloned voice still on sale after they leave. AI speeds up the drafting and the audio. It does not remove the checking.

Follow me on Instagram@sagnikteaches

Three ways to build one and what each leaves you to license

A small walking-tour company with a self-guided route through a historic quarter is the running example: 12 stops, about 90 minutes on foot, and roughly 18 minutes of audio in total. That comes from a planning figure of 225 words a stop, so 2,700 words, read at about 150 words a minute. The company can build it three ways.

Connect on LinkedInSagnik Bhattacharya

Do it yourself suits five to twenty stops and needs no app, although automatic playback by location usually does. A hosted platform saves time and moves some licence questions to its terms, which is why the last section lists ten questions to ask before you upload a script. Your own tour descriptions will also need translating if you serve several languages; the tutorial on tour descriptions in several languages covers that workflow.

Subscribe on YouTube@sagnikteaches

The licence map: six layers, six different questions

Think of an audio guide as a stack. Each layer has its own owner and its own terms, and a gap in any one can undo the rest.

Translation is not a separate layer: it inherits the licence of the text it came from. If the English script uses wording you lack permission for, the Spanish version does too. Keep the map in a single document, because insurers, platforms and marketplaces all ask the same questions eventually.

Write the script from a claim log, not from ChatGPT's memory

Left alone, a chatbot writes fluent history that mixes true facts, plausible guesses and local legends. The cure is to put the facts in first. Before drafting anything, build a claim log: one row per fact you intend to say, with where it came from and whether a second source agrees. Here is a fragment for two stops, using a made-up row of mill cottages and a made-up market (illustrative).

Only verified rows go into the prompt. This one asks for stop 4 and tells the model what to do when it runs short of facts.

Write the audio script for stop 4 of a self-guided walking tour. Use only the verified notes below.
If you need a fact that is not in the notes, write [NEEDS FACT] instead of guessing.
Style: spoken English, 200 to 230 words, sentences under 25 words, no lists, no brackets, numbers written as
words. Start by telling the listener where to stand and what to look at. End with the direction to stop 5.

Verified notes for stop 4 (Weaver's Row, invented for this example):
- A row of 14 cottages, built in the 1840s for mill workers (two sources).
- Each had two rooms up and two down; the families shared one pump in the yard.
- The mill closed in 1961 and the cottages were sold off one by one.
- Stop 5 is the covered market, about 150 metres along the lane; turn left at the pump.

The result (illustrative) reads well, which is exactly why it needs a second look.

Stand at the cobbled corner and look along the terrace of low brick cottages on your right. This is Weaver's Row. Fourteen homes were built here in the eighteen-forties for workers at the mill. Each had two rooms upstairs and two down, and every family shared one pump in the yard. Picture fourteen households sharing it. The mill closed in nineteen sixty-one, and the cottages were sold off one by one. Look at the doors: you can still tell which owners kept the original paint. Locals say the lane was named after the first weaver to move in. Now follow the lane towards the pump. Turn left there, and the covered market is about a hundred and fifty metres ahead.

Two sentences did not come from the notes. "You can still tell which owners kept the original paint" is an invented detail that may be false on the day. "Locals say the lane was named after the first weaver" is the unverified legend from the claim log, which the model picked up from its own memory despite the instruction. Cut the first and decide on the second: either keep it with the "it is said" framing or drop it. The opening also invents the scene ("cobbled", "brick", "on your right"), because the prompt asked it to say where to stand. Those are claims about the street, so list them in the log as site details to confirm on the walk. The draft is a little short of the target length, which is better than padding it with guesses. If you want a second set of prompts for describing routes and itineraries, the tutorial on AI for tour guides has some.

Writing for the ear: numbers, names and pace

A script that reads well on a screen can sound awkward in a pair of earbuds. Run an ear edit before you generate any audio. The changes are small and mechanical.

• Numbers and dates. Write "eighteen seventy-four", not "1874", and "a hundred and fifty metres", not "150 m". Services read digits inconsistently, and a year can come out as a quantity.
• Parentheses and abbreviations. "The Market Hall (est. 1874) on St. Anne's St." becomes "The market hall opened in eighteen seventy-four, on Saint Anne's Street."
• Sentence length. One idea per sentence and under about 25 words. If you run out of breath reading it aloud, so will the voice.
• Place names. List every local name and spell hard ones phonetically in a pronunciation sheet. Most services let you respell a word or add a pronunciation rule; check which yours offers, then listen to every name.
• Directions. Finish each clip with the next move ("turn left at the pump") so a listener never stands in front of a sign waiting for audio to end.

Pace needs measuring rather than guessing. A narration pace of about 150 words a minute is a sensible planning figure, but each voice and speed setting differs. Generate stop 4, check the real running time, and divide. If it comes to 85 seconds for 205 words, your voice runs at about 145 words a minute, and you can plan the rest of the route from that number. Then compare the clip lengths with the walking time between stops. If visitors listen while they walk, a 90-second clip on a four-minute walk leaves two and a half minutes of silence, so either add a short "on the way" paragraph or accept the quiet on purpose.

Fact checks that catch what an AI invents

Most errors in an AI-drafted tour fall into a few types, and each has a cheap test.

• Superlatives. "Oldest", "first", "tallest", "only". Any of these needs two sources or comes out.
• Dates and ages. Check the year, then check what happened: a building may have been built, burnt down and rebuilt, and the plaque may give only one of those dates.
• Names and titles. Spell-check against an official source, not against the model's spelling.
• Distances and directions. Walk them. A script that says "turn left at the pump" fails when the pump was removed last spring.
• Legends. Keep them if they are good, and label them: "it is said that".
• Opening hours and prices. These change; send listeners to the venue's own page.

Here is a realistic mistake in the form it usually takes (illustrative). The script says a bridge "was built in eighteen eighty". The plaque on the bridge says "rebuilt 1880". A visitor on the tour reads the plaque aloud while your audio is playing. The fix is not to apologise; it is to keep the claim log, find the original date, change the clip and re-render only that stop. That is why the log records which stop each fact belongs to.

Ask a local volunteer, a museum curator or a long-standing guide to walk the finished route with the audio. One person who knows the place will catch things no source list does. Never publish a stop with a [NEEDS FACT] left in it.

Text-to-speech options and what their terms say

Four services are enough to show how much the terms differ. The table records what their own pages said when I read them in October 2026, and it covers licensing facts, not sound quality, which you should judge by listening to your own script.

The last row is a useful warning. Audio files you have already rendered keep playing, but if you need to fix a fact in 2027, you will have to re-render, and the replacement voice will not sound the same. Check the provider's deprecations page before you commit to any service, and keep your scripts and pronunciation sheet in your own files. The same documentation says its voices are optimised for English, so test every other language with a fluent listener before it goes live.

Now the sums for the 12-stop route. At about 5.7 characters a word, 2,700 words is roughly 15,400 characters. At the Google rates above, that costs about $0.06 with standard voices, $0.46 with Chirp 3 HD and $2.46 with Studio voices, and the free monthly allowances may cover a first render entirely. On ElevenLabs Starter, the page's 30 minutes covers the 18-minute English tour with room for about one partial re-render; a second language would need another month or a bigger plan. Voice cost is a rounding error next to the day you will spend on checking. For plan details and limits on one of these services, see the tutorial on ElevenLabs plan limits for business voiceovers.

Voices: stock, cloned and hired, with the consent wording

A stock voice from a speech provider is licensed under that provider's terms for your plan. Keep a dated screenshot of the terms and your plan for each batch you render, since both change. Add one honest line at the start of each guide, for example: "This guide is narrated with an AI-generated voice. The script was written and checked by our own team." It meets the OpenAI rule quoted above, and if you have listeners in the EU it fits the transparency direction of Article 50 of the EU AI Act. That article puts a duty on speech providers to mark synthetic audio so machines can detect it, and a separate duty on anyone publishing a deep fake to disclose it. There, a deep fake means AI-generated or manipulated audio, image or video that resembles existing people, places, objects or events and would falsely appear authentic. A generic synthetic narrator is not that, but a clone of a named guide could be.

Cloning your best guide's voice deserves care. It saves re-recording when facts change, and it creates a legal dependency on one person. ElevenLabs requires its Creator plan or above for professional cloning, together with a verification step, and OpenAI's documentation describes custom voices built from a consent recording. Even so, the provider's check does not replace your own paperwork. Ask the guide to sign something plain before any recording. Illustrative wording, to be reviewed by a solicitor:

I agree that [company] may record my voice and create a synthetic copy of it, to be used only for the audio tours named in this agreement, in the languages listed. The copy may be used until [date or event]. I may withdraw consent by written notice, after which [company] will stop using the copy within [period] and delete it from the speech provider's systems. [company] will pay me [fee]. The copy will not be used for any other purpose or sold to a third party. [CHECK: solicitor]

Each blank there is a negotiation. Term, exit and deletion matter most: what happens to published tours the day the guide leaves? A cleaner arrangement for a small team is a stock voice and a human guide who introduces the tour in their own voice. Do not make an AI voice imitate a recognisable person who has not agreed, and label dramatised speech by historical figures as dramatisation. If you want the legal angle on cloned voices in more depth, the tutorial on AI voice cloning and the legal rules for business use goes through it.

Music, sound effects and the royalty-free trap

Music is where good intentions meet licence terms. "Royalty-free" describes how you pay, not whether the music is free: one music library's own guidance says it means no ongoing royalty payments after licensing, and that a library's licence tier decides your rights. The same guidance warns about tiers labelled "personal" that do not cover commercial use. A paid guide is commercial use.

A second trap comes from the same guidance: a track can be sold royalty-free and still be registered with a performing-rights society, in which case you can still owe public performance fees. Whether audio played from a visitor's own phone in the street counts as public performance is a question for the licensor, in writing. Two more questions belong on the same list, because library licences differ on them: is there a cap on downloads or listeners for a paid product, and is streaming to the public treated differently from a file the visitor downloads?

Creative Commons tracks need the same reading. Their own explanations say attribution (BY) requires credit to the creator, NonCommercial (NC) allows only use "not primarily intended for or directed towards commercial advantage or monetary compensation", and ShareAlike (SA) means adaptations must be shared on the same terms. A paid tour will almost certainly not count as noncommercial, so NC tracks are out. For BY, put the credit in the stop's transcript page and, if the licence asks for it, in the audio.

Before you buy or download anything, ask the licensor four questions and save the replies: does the licence cover a paid audio guide; does it cover playback in public places; is there a cap on downloads or listeners; and what are the term and territory? AI music tools have their own terms, so check whether your plan allows commercial use and who owns the output. My advice for a first version is to publish with the voice only. A guide without music is not worse, and you can add licensed music later without redoing the research.

Script sources: Wikipedia, guidebooks, plaques and AI drafts

Facts are one thing and wording is another. In most places the date a building opened is free for anyone to state, but a guidebook's sentence about it is the author's. Keep a source-licence column in the claim log so that you know at a glance which rows you may quote and which you may only learn from. Here is a short version (illustrative).

Wikipedia's reuse guidance says its text is available under Creative Commons Attribution-ShareAlike and that reusers must credit the authors, indicate changes and release modified versions under the same licence. A script adapted from an article could therefore carry that licence into a commercial guide, so treat the article as a pointer to primary sources. The reuse page spells it out.

The AI drafts raise a quieter question: who owns them? A national copyright office's January 2025 report on AI outputs concluded that prompts alone do not give users enough control to count as authors, while a human's creative selection, arrangement and modification of AI output can be protected, and human-written material stays protected. The practical reading for a tour business is to rewrite substantially, keep your drafts and edits, and not assume you can stop a competitor copying an AI script you barely changed. It is also why the claim log and your own research are your real assets. If you plan to work with museums or heritage sites, expect them to ask for this evidence; the tutorial on an AI policy for museums shows what such organisations ask about audio scripts and labels.

Hosting and delivery: QR codes, offline use and transcripts

The simplest delivery is one web page per stop with an audio player, the transcript, a map link and the next-stop direction. Print a QR code on each sign or on a route card, and test every code on a phone that is not on your wifi. Name files clearly (stop-04-v2.mp3) so the claim log can say which clip to re-render when a fact changes.

Two features are worth adding early. Offer a download of the whole set for visitors who will lose signal in narrow streets. And publish transcripts, since they help deaf and hard-of-hearing visitors, people who prefer to read, and search engines. Put the AI-voice line and the music credit at the top of each page.

If you sell the guide, link it to your ticketing and booking, which the tutorial on walking tour software for ticketing and guide rotas compares.

Ten questions for an audio-guide platform before you upload anything

Platforms that bundle voices, hosting and location triggers save time, and they also decide a lot of the licence picture. Ask these before you upload a script, and keep the answers.

1. Who owns the audio files and the scripts I upload?
2. Can I export every MP3 and transcript if I leave?
3. Which speech service sits underneath, and does my plan include its commercial licence for paying visitors?
4. If I clone a guide's voice, who holds the consent record and what happens to the voice if I delete it?
5. Is music included, and under what licence for a paid, publicly played guide?
6. Is my content used to train any model?
7. Are machine translations labelled, and can a fluent speaker edit them before they go live?
8. What happens to my tours if the platform closes or the voice model underneath is retired?
9. Is pricing a flat fee or per listener, and what counts as a listener?
10. Does the platform tell listeners the narrator is AI-generated?

A vendor that answers question 2 with "we will look into it" has told you something. If you are comparing platforms by price as well, add the cost of leaving to the cost of staying.

A build plan with hours, and a test walk

These are my estimates for the 12-stop English tour, assuming you know the area.

The test walk is the last gate. Walk the route with a stopwatch and a phone on mobile data. At each stop, check that the QR code scans, the audio starts within a few seconds, the clip ends before a typical walker reaches the next stop, the directions are right, and the names sound acceptable to a local. Then ask one person who has never been there to follow the tour with no help and tell you where they hesitated. Fix those points, re-render only the clips that changed, update the claim log, and note the date. After that the guide is ready to sell.

Questions tour businesses ask about AI audio guides

Can I use a free text-to-speech plan for a paid tour?

Check the plan's terms first. One provider's pricing page lists a commercial licence from its first paid plan upwards and lists none for its free plan. Keep a note of the plan and the date each batch of audio was made, because terms and plans change, and screenshot the terms page the day you render.

Do I have to tell listeners the voice is AI?

One speech provider's documentation says its usage policies require clear disclosure that the voice is AI-generated, and the EU AI Act has disclosure duties for synthetic audio that imitates real people. A single line at the start of each guide covers both and costs nothing.

Can I clone my best guide's voice?

Only with their informed, written consent covering this use, the languages, how long you can use it and what happens if they leave. The provider will usually ask for its own verification as well. Ask a solicitor to check the consent wording before you record anything.

Can I paste text from Wikipedia into my script?

That text is shared under a licence that requires credit and says adapted versions must carry the same licence, which may not suit a commercial guide. Use articles as leads to primary sources, write in your own words and confirm every fact with a second source.

How long should each stop be?

Plan about 200 to 230 words, roughly 90 seconds at a common narration pace of 150 words a minute, then time the finished audio on the real route. Walking time between stops matters as much as the clip length, so put the directions to the next stop at the end of each clip.

Further reads

• 8 Examples of AI in Tourism From Guided Tours to Group Bookings — Eight examples of AI in tourism, from guided tours to group bookings.
• AI Marketing Booklet Creator: Print a Tour Brochure From Itineraries — Print a tour brochure from your itineraries.
• How Small Tour Operators Use AI for Bookings and Itineraries — Where else small tour operators use AI day to day.
• Viator and GetYourGuide Listings: Rewrite Titles With AI — Rewrite marketplace listing titles once the audio tour exists.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ElevenLabs pricing page; OpenAI text-to-speech documentation and deprecations page; Google Cloud Text-to-Speech pricing; Amazon Polly FAQs; a national copyright office's January 2025 report on AI outputs; Creative Commons licence explanations; Wikipedia's reuse guidance; EU AI Act Article 50; music-library licensing guidance. All read in October 2026. Check each current page before you buy.