How to write AI voice note scripts that sound like a person, with eight before-and-after rewrites
A good AI voice note script runs about 25 to 65 words, carries one idea, opens with the customer's name and ends with one clear next step. At the 155 words a minute that one text-to-speech study used as its normal rate, that is 10 to 25 seconds. Every price, date and reference number in it also goes in a text message underneath. This guide comes from Learnmind.ai, which builds WhatsApp AI agents in Dubai that reply with a voice note when the moment calls for it. It covers length, openings, sign-offs, pronunciation, numbers and what never to say out loud.
The short answer
Aim for 10 to 25 seconds. That is roughly 25 to 65 words. In a study of synthetic speech, a faster rate of 178 words a minute lowered comprehension for native and non-native listeners alike.
One idea per note. Young adults hold about three to five meaningful items in mind at once. A note that confirms a booking should not also sell a package.
Open like a person, close with a handover. Put the customer's name and the reason for the note in the first few words. End on one question or one next step, then stop.
Say numbers once, and always send them in text. Prices, times, addresses and references go in a written message under the voice note. WhatsApp's audio message has no caption field, so that text is its own message.
Some things never go in audio. Card numbers, one-time codes, test results, ID numbers, links and any offer the business has not approved. A voice note can be played on speaker in a waiting room.
Write for the ear, not the eye
A text reply is read at the customer's pace. They can skim it, read it twice and copy the price. A voice note plays at the speaker's pace, once, in order. So the script has to do the work the eye normally does: signposting, short sentences and a clear end.
WhatsApp said in March 2022 that its users send 7 billion voice messages a day on average. Your business sits in the same chat list as your customers' friends and family. A voice note that sounds like a radio advert or a phone menu stands out for the wrong reason. One that sounds like a calm, specific person fits in.
| Written for the eye | Written for the ear | Why it changes |
|---|---|---|
| "Appt: Sat 10/10, 4:00 PM, Dr. Rana, JLT branch." | "You're all set for Saturday at four, with Dr Rana at our JLT clinic." | Abbreviations, slashes and colons can be skipped or read out literally |
| A sentence with a bracket (like this one) in the middle | The aside becomes its own short sentence | A listener cannot see where a bracket closes |
| One long sentence with three clauses | Two or three short sentences | The listener has to hold the start until the point arrives |
| A bulleted list of five options | Two options by name, then "I've sent the full list below" | Five spoken options is more than most people can hold |
| "Please find below the details as per your request" | "I've put the details just below" | Office phrases sound read out, not said |
| "Kindly confirm at your earliest convenience" | "Does Saturday work for you?" | A direct question is how people talk to each other |
Terms used in this guide
- Voice note script
- The exact words an AI voice will speak, written and checked before any audio is generated.
- Text companion
- The text message sent with the voice note. It carries every number, date, address and link.
- Phonetic respelling
- Writing a word the way it sounds, such as "joo-MAY-ruh" for Jumeirah, so the voice says it right.
- SSML
- Speech Synthesis Markup Language, the W3C standard for marking up text for a speech engine. Version 1.1 became a W3C Recommendation on 7 September 2010.
- Customer service window
- On WhatsApp, the 24 hours after a customer's last message or call. Inside it a business can send free-form messages, including audio.
How long an AI voice note should be, and why one idea is the limit
A 2007 paper by Jones, Berry and Stevens in Computer Speech and Language tested speaking rate directly. Text-to-speech descriptions of eight banking products were played to 80 native speakers of Australian English and 80 non-native English speakers. The rates were 155 words a minute, which the paper calls normal, and 178. The faster rate lowered comprehension for both groups and made the message no more persuasive. Speed bought nothing.
That matters for any business whose customers may be listening in a second or third language, or in a car or a busy reception.
Real conversation is faster, so a script voiced at chatting speed can sound rushed. In a 2006 paper, Yuan, Liberman and Cieri measured 2,438 recorded English telephone conversations at an overall 196 words a minute, or 236 with silences removed. They also found that people talking with strangers tend to use longer turns and slower speech. A customer is a stranger, so calm and slightly slow is the natural register.
The same study found speaking rate rises sharply across turns of one to seven words. Very short phrases are spoken more slowly per word. A script built from three-word sentences runs longer than its word count suggests, so time it.
| Target length | Words at 155 a minute | Words at 178 a minute | What fits |
|---|---|---|---|
| 10 seconds | About 26 | About 30 | A confirmation, a thank-you, "yes, we have it" |
| 15 seconds | About 39 | About 45 | One answer plus one question |
| 20 seconds | About 52 | About 59 | A short explanation and a next step |
| 25 seconds | About 65 | About 74 | The upper limit for a first reply |
| Over 30 seconds | 78 or more | 89 or more | Split it into two notes, or send text |
The word counts are our arithmetic from the study's two rates, which were measured for English. Arabic has its own rhythm, so we treat Arabic timing separately in Arabic AI voice notes.
The 10 to 25 second target is our recommendation, not a platform rule. A first reply should feel like a quick personal note, not a voicemail. WhatsApp lets people play voice messages at 1.5x or 2x speed, and in 2022 extended that to forwarded messages. A customer who speeds your note up has told you it was too long.
There is a technical reason too. Meta's documentation says a voice message shows a play icon only if the file is 512 KB or smaller; above that, the customer sees a download icon. Voice messages must be mono OGG files with the OPUS codec, and every audio file has a 16 MB ceiling.
By our arithmetic, a file encoded at 32 kilobits a second reaches 512 KB at a little over two minutes. A good script never gets close. For the sending side, see voice notes through the WhatsApp Business API.
One idea per note
Nelson Cowan's 2010 review of working memory puts the limit at "3 to 5 meaningful items in young adults". For longer verbal chunks, "such as idioms or short sentences", young adults recall only three or four. That is the budget for a voice note. Spend it on one thing.
A 60-word note holding a price, a date, a location, a deposit and a cancellation rule asks for five chunks at once. One idea means one decision for the customer: confirm, choose, come in or reply. If the script needs the word "also", split it or move the second idea into the text companion.
When a voice note is the wrong format
| Send a voice note when | Send text instead when |
|---|---|
| The customer sent a voice note first | The customer wrote a long, precise message and expects a precise answer |
| Warmth matters: a welcome, a thank-you, an apology for a delay | The content is a list: prices, a menu, a timetable |
| A short explanation would sound cold in writing | The customer will need to copy something: an address, a reference, a link |
| The customer asked you to explain something | The customer said they cannot listen right now |
| The conversation is inside WhatsApp's 24-hour customer service window | The window has closed: outside it, only pre-approved templates can be sent |
We compare the three channels in voice note, text or phone call.
How to open and close a voice note
Openings: a greeting, a name, a reason
Michal Marmorstein studied openings in Hebrew WhatsApp chats, in a 2023 paper in Studia Fennica Linguistica. Most openings were a minimal greeting that the other person did not return. The greeting worked as "a design feature of the message", framing what followed as a social encounter instead of starting cold with content.
A business voice note needs the same frame in its first two or three seconds: a greeting, the customer's name and the reason for the note. Then straight to the answer.
| Opening | Example | Use it when |
|---|---|---|
| Name and reason | "Hi Sara, it's Layla from the clinic, about your Saturday booking." | Most replies, and every first voice note in a chat |
| Thanks first | "Thank you for waiting, Priya." | Any reply that comes after a noticeable gap |
| Mirror the greeting | "Wa alaykum as-salam, Khalid." | The customer opened with "As-salamu alaykum" |
| Never | "Hello valued customer, this is an automated message regarding your enquiry." | Never. It sounds like a phone menu and says nothing |
Mirror the customer's language as well as their greeting. A customer who wrote in Arabic should not hear an English voice note back. If the persona has a name, such as "Layla from the clinic", keep it the same in text and voice.
Closings: the hardest part to get right
Jana Tschannen of the University of Basel studied how people end voice messages, in a 2023 paper in the Bulletin suisse de linguistique appliquée. The data was 233 voice messages and 567 text messages from 42 WhatsApp chats of German and Swiss users, collected in 2018 and 2020. The analysis is qualitative, so it shows patterns rather than percentages.
The paper starts from a recognised problem: online memes mock how people end voice messages with a trailing "and... yeah...". Users do not simply stop recording, Tschannen found. They wind down with words and intonation, and some invent endings that exist only in voice messages. Three details are directly useful for scripts.
- Dead air is common. In longer messages the recording often runs on for one to two seconds of silence after the last word.
- A question hands the turn back. One 14-second message ends by making clear there is one thing that needs an answer, passing the ball to the other person.
- Farewells cluster on the switch. A farewell word often ended someone's first voice note after a run of text messages. In the chat the paper shows, later voice notes dropped it.
Our rules follow from that. End on the next step, said once. Ask at most one question, and make it the last sentence. Trim the audio so it stops within half a second of the last word. Use a short farewell on the first voice note in a text conversation, and drop it in a quick back-and-forth.
Never end with "let me know if you have any questions". The customer already knows they can reply.
Sign-offs that fit the business
| Business | Sounds like a person | Sounds like a template |
|---|---|---|
| Aesthetics clinic | "See you Saturday, Sara. Anything before then, just message here." | "We value your business and look forward to serving you." |
| Luxury car rental | "It'll be ready at ten. Enjoy the drive." | "Please do not hesitate to contact us." |
| Real estate | "I'll send the floor plan now. Shall I hold Thursday?" | "Looking forward to hearing from you at your convenience." |
| Home services | "He'll message you when he's ten minutes away." | "Your request has been logged and will be processed." |
A human sign-off names something concrete: a day, a time, a person or an action. A template sign-off names nothing. For sector detail, see AI voice notes for clinics and AI voice notes for luxury brands.
Fillers, pauses and saying names right
Fillers: one at most, and never before a number
Real speech has small words that signal what comes next: "so", "right", "okay", "actually". A script with none can sound read out. A script with five sounds like a nervous presenter.
Our rule is one discourse marker per note at most, and never in the opening line. "So" works at the start of the answer. "Right" works before the next step. "Um" and "uh" almost never belong in a business script, because an "um" before a price sounds like doubt about the price.
Pauses: punctuation first, markup second
The first tool for pauses is the script itself: a full stop where a person would breathe, a new sentence before the key fact, a comma after the customer's name. Then listen on a phone speaker, because that is where it will be heard.
Where a speech engine supports SSML, markup gives finer control. The W3C specification's break element takes strengths
from "none", "x-weak", "weak" and "medium" to "strong" and "x-strong", or an exact time. The prosody element sets rate
from "x-slow" to "x-fast" or as a percentage. The emphasis element takes "strong", "moderate", "reduced" or "none".
Support varies by engine; for say-as, the specification says SSML "does not enumerate the possible values". Judge the
audio, not the markup.
Saying names right
Nothing sounds more like a machine than a customer's own name said wrong. SSML has two tools for this. The phoneme
element gives an exact pronunciation, and engines "SHOULD support" the International Phonetic Alphabet for it. The sub
element swaps in "the string to be spoken instead of the enclosed string", which is how "JBR" becomes "J B R".
Most businesses do not need phonetic alphabets. They need a pronunciation sheet: one row per name, place and product the voice will say often.
| Written | Respelling for the voice | Hint type | Note |
|---|---|---|---|
| Jumeirah | joo-MAY-ruh | Respelling or phoneme | A place name said every day; get it right once |
| Ras Al Khaimah | ras al KHAI-mah | Respelling or phoneme | The "kh" is one sound, not "k" then "h" |
| Siobhan | shi-VAWN | Respelling or phoneme | Irish spelling defeats English reading rules |
| JBR, DIFC | J B R, D I F C | Alias | Letters, not words |
| AED 1,250 | twelve hundred and fifty dirhams | Alias, or write it out | A currency code may be read as three letters |
| Mohammed, Mohamed, Muhammad | One agreed pronunciation | Respelling | Several spellings of one name should not produce several pronunciations |
Illustrative respellings, ours. Check every entry with the person, or with someone who says the name every day.
For customers' own names, there is a safer rule. If the customer has sent a voice note saying their name, use that pronunciation. If nobody can confirm an unusual name, leave it out of the audio and use it in the text companion. A missing name costs nothing. A mangled one costs the warmth the voice note was for.
Numbers, prices and dates: say them once, send them in text every time
This is the rule we never bend. Every number in a voice note also appears in the text companion, and most numbers in the companion never appear in the voice note. There are four reasons, and each would be enough on its own.
- Memory. A price, a date and a reference use up most of the three to five items Cowan's review says people can hold.
- The customer may read it instead. Meta's documentation says that if a user has turned on automatic voice message transcripts, the message arrives with a text transcription. A misheard price in a transcript is still a wrong price.
- Some customers cannot listen. On the web, WCAG 2.2 Success Criterion 1.2.1, at Level A, asks for a text alternative to prerecorded audio. It serves anyone "having trouble understanding audio information for any reason". A chat is not a web page, but the principle carries.
- A written price is a declared price. The UAE government portal describes Federal Law No. 15 of 2020 on Consumer Protection as protecting the right to goods and services "at the declared price". It also says invoices must be in Arabic, with other languages optional.
A voice note is not an invoice, and nothing we read treats it as one. Our reading is simpler. A price the customer can scroll back to and screenshot protects both sides. A price heard once at 2x speed protects nobody.
On WhatsApp the companion has to be a separate message. Meta's audio message object has three fields, id,
link and voice, and no caption. With voice set to true, the note shows the business's profile
image with a microphone icon.
| In the data | In the voice note | In the text companion |
|---|---|---|
| AED 1,250 | "twelve hundred and fifty dirhams" | AED 1,250 |
| AED 2,350,000 | Leave it out: "I've sent the price below" | AED 2,350,000 |
| 11/10/2026 | "Sunday the eleventh of October" | Sunday 11 October |
| 18:00 | "six in the evening" | 6pm (18:00) |
| +971 58 519 0783 | "plus nine seven one, five eight, five one nine, zero seven eight three", or better, leave it out | +971 58 519 0783 |
| Booking ref BK-4827 | Leave it out | Ref: BK-4827 |
Dates deserve special care. "11/10" is 11 October to most of the world and 10 November to anyone used to month-first dates. A weekday and a month name cannot be misread, spoken or written.
Two checks catch almost every number error. The voice never contains a number the text lacks. Every number in the voice matches the text exactly, including words like "about" and "from".
What the text companion costs on WhatsApp
The business: a Dubai clinic whose business number sends 2,400 service messages in October 2026.
Inside that: 300 voice notes and 300 text companions, all inside customer service windows.
Meta's charge: the first 1,000 are free; the other 1,400 cost USD 0.0157 each, so USD 21.98 for the month.
The companions' share: at most 300 x USD 0.0157 = USD 4.71.
Illustrative example, at Meta's UAE service rate from 1 October 2026. See our WhatsApp Business API pricing guide.What never to put in a voice note
Write every script as if a stranger could hear it.
A voice note can play on speaker in a car, a lift or a waiting room. It can be forwarded, and turned into text on the customer's phone. Anything private, exact or binding belongs in writing, or with a person.
| Never in audio | Why | Do this instead |
|---|---|---|
| Card numbers, bank details, CVV codes | Overheard, forwarded, transcribed | A secure payment link in text |
| One-time codes and passwords | No legitimate reason to say them | Never ask for them, never send them |
| Test results, diagnoses, treatment advice | Personal, and clinical advice needs a clinician | Hand over to the clinic team |
| Emirates ID, passport or licence numbers | Identity details should not be spoken | Ask for documents in text; confirm receipt without reading them back |
| Addresses, links and email addresses | Nobody can tap or copy audio | Text companion, with a map pin or link |
| Prices or promises the business has not approved | A warm voice sounds like a commitment | Approved facts only; anything else goes to a person |
| Terms and conditions in full | Too long to hear, and they need to be read | One sentence in voice, the full terms in writing |
| A sales pitch nobody asked for | Outside the 24-hour window it cannot be sent at all | Opt-in and an approved template; see WhatsApp marketing consent in the UAE |
Sounding human is not the same as pretending to be human
The goal of a script is warmth and clarity, not making a customer believe a member of staff recorded it. The persona can have a name and a personality. If a customer asks whether they are talking to a person, the answer should be true.
The EU now writes this into law. Article 50(1) of the AI Act requires providers to make sure people are told when they are interacting with an AI system, unless that is obvious. Article 50(2) requires providers of systems that generate synthetic audio to mark it in a machine-readable format. Article 50(4) requires deployers to disclose deep fakes, which the Commission defines to include AI audio resembling existing persons.
The European Commission says these rules apply from 2 August 2026, with a grace period for marking until December 2026 for generative systems already on the market. Whether the AI Act reaches a particular Dubai business is a question for a lawyer. As a design rule, it costs nothing to follow. Never clone a real employee's voice without their written agreement, and never script fake hesitations to suggest a person is thinking on the spot.
Eight before-and-after script rewrites
Each rewrite starts from the kind of reply a business sends when it reads a text answer aloud. Timings use 155 words a minute; spoken figures make a "before" run longer still. Every business, name and price is invented.
1. Aesthetics clinic: a price question at 11pm
Customer: a voice note asking the price of full-leg laser hair removal.
Before, 53 words, about 21 seconds: "Hello and thank you for contacting the leading aesthetics clinic in Dubai. Our full legs laser package is AED 1,800 for six sessions, or AED 400 per session. New clients get 20% off until the end of the month, terms and conditions apply. Please let us know if you have any other questions."
After, 39 words, about 15 seconds: "Hi Sara, thanks for your voice note. Full legs is done as a course, so there's a single-session price and a package price. I've sent both just below. Would you like me to hold a patch test for Saturday?"
Text companion: Full legs: AED 400 a session, or AED 1,800 for six. Patch test: free. Saturday: 11am, 2pm or 4pm.
What changed: the name comes first, the boast and the conditional discount are gone, both prices moved to text, and it ends on one question.
Illustrative example.2. Luxury car rental: a weekend quote
Customer: "How much for the black SUV in your story, Friday to Sunday?"
Before, 51 words, about 20 seconds: "Dear Sir, thank you for your interest. The rental for 3 days is AED 4,500 plus a refundable security deposit of AED 5,000. It includes 250 km per day, with extra km at AED 10, plus tolls and fuel. Kindly share your Emirates ID, passport and driving licence copies to proceed."
After, 42 words, about 16 seconds: "Hi Omar, good news, the black SUV is free Friday to Sunday. I've put the price and the deposit below so you have them in writing. If it works for you, I can hold it while you send your documents. Shall I?"
Text companion: Black SUV, Friday 9 to Sunday 11 October: AED 4,500. Refundable deposit: AED 5,000. 250 km a day, then AED 10 a km. Tolls and fuel extra. To book: driving licence plus Emirates ID or passport.
What changed: "Dear Sir" is gone, five numbers left the audio, and the document request moved to text, where the customer can act on it.
Illustrative example.3. Salon: a booking confirmation
Customer: "Can I book gel hands and feet with Maria on Sunday afternoon?"
Before, 29 words, about 11 seconds: "Your appointment has been successfully booked for gel manicure and pedicure on 11/10/2026 at 16:00, Shop 12, Building 4, Al Barsha. Cancellations within 24 hours will be charged 50%."
After, 31 words, about 12 seconds: "Hi Priya, you're booked in with Maria on Sunday at four, gel on hands and feet. Come ten minutes early and we'll have tea ready. Everything's written below. See you Sunday."
Text companion: Sunday 11 October, 4pm (16:00), gel manicure and pedicure with Maria. Shop 12, Building 4, Al Barsha, map pin attached. Changes less than 24 hours before: 50% charge.
What changed: a slash date that a voice might read as "eleven slash ten" became a weekday, and the cancellation rule moved to writing.
Illustrative example.4. Real estate: an overseas buyer at 2am
Customer: "Is the 2-bed in the Marina still available, and what's the service charge?"
Before, 52 words, about 20 seconds: "Yes, the 2 bedroom apartment in Dubai Marina is still available. The asking price is AED 2,350,000 and the service charge is AED 18 per square foot per annum, on 1,250 square feet. The return is approximately 6 to 7% based on current rental yields. Would you like to schedule a viewing?"
After, 40 words, about 15 seconds: "Hi James, yes, the two-bed in the Marina is still available. I've sent the price, the service charge and the floor plan below. If you're abroad, we can do the viewing by video. Would Tuesday or Wednesday evening suit you?"
Text companion: 2-bed, Dubai Marina, 1,250 sq ft. Asking price: AED 2,350,000. Service charge: AED 18 per sq ft a year, about AED 22,500. Floor plan attached.
What changed: a seven-figure price is slow to say and easy to mishear, so it went to text. The yield estimate went entirely: it is a forecast, and forecasts belong with an agent, not a script.
Illustrative example.5. Restaurant: a birthday table for twelve
Customer: a voice note asking for a table for twelve on Thursday at half past eight, for a birthday.
Before, 52 words, about 20 seconds: "We have received your reservation request for 12 guests on Thursday 15 October at 8:30 PM. For groups of 10 or more a set menu is required at AED 285 per person. A deposit of AED 1,000 is required to confirm. Please note we have a 2 hour seating policy for groups."
After, 44 words, about 17 seconds: "Hi Reem, a birthday table for twelve on Thursday at half eight, lovely. For groups that size we do a set menu, so I've sent it below with the price. Have a look, and tell me if you'd like me to hold the table."
Text companion: Thursday 15 October, 8:30pm, 12 guests. Group set menu: AED 285 per person, menu attached. Deposit to confirm: AED 1,000. Group seating: 2 hours.
What changed: the voice keeps the warmth of the occasion. The rules that could put her off are in writing, where she can share them with the group.
Illustrative example.6. Home services: an AC technician running late
Customer: "Where is the technician? I was told 1pm."
Before, 42 words, about 16 seconds: "Dear Customer, we regret to inform you that due to unforeseen circumstances our technician will be delayed. The new estimated time of arrival is between 3:00 PM and 4:00 PM. We apologise for any inconvenience caused. Your job reference number is HS-20931."
After, 38 words, about 15 seconds: "Hi Faisal, I'm sorry, Arun is running late. The job before yours took longer than planned. He'll be with you between three and four this afternoon, and he'll message when he's ten minutes away. Does that still work?"
Text companion: New arrival window: 3pm to 4pm today. Technician: Arun. Job ref: HS-20931.
What changed: a named person and a real reason replace "unforeseen circumstances". The apology sounds meant when spoken, and the reference number stayed in text.
Illustrative example.7. Nursery: a parent asking about places
Customer: "Do you have FS1 places for my daughter for September 2027? What are the fees?"
Before, 41 words, about 16 seconds: "Thank you for your interest in our nursery. Fees for FS1 are AED 42,000, payable in 3 instalments, plus a non-refundable registration fee of AED 500. Please upload the birth certificate, passport copies, Emirates ID, vaccination records and 2 passport photos."
After, 48 words, about 19 seconds: "Hi Noura, thank you for thinking of us for your daughter. We do have FS1 places for next September. The best first step is a visit, so you can see the classrooms. I've sent the fees and the document list below. Would a morning next week suit you?"
Text companion: FS1, September 2027: AED 42,000 a year, in 3 instalments. Registration: AED 500, non-refundable. Documents: birth certificate, passports, Emirates IDs, vaccination record, 2 photos.
What changed: a five-item document list is unhearable, so it moved to text. The voice does what text cannot: it sounds pleased to hear from a parent.
Illustrative example.8. Hotel: a late checkout request
Customer: "Any chance of a late checkout tomorrow?"
Before, 46 words, about 18 seconds: "Dear Guest, late check-out is subject to availability and may incur additional charges. Late check-out until 14:00 is complimentary for loyalty members, otherwise AED 150 per hour. Until 18:00 it is charged at 50% of the room rate. Please contact reception on the day to confirm."
After, 35 words, about 14 seconds: "Hi Daniel, good news, I can give you checkout at two tomorrow, at no charge. If you need longer, there's a small fee, and I've put the options below. Enjoy the rest of your stay."
Text companion: Late checkout confirmed: 2pm tomorrow, no charge. Until 4pm: AED 150 an hour. Until 6pm: 50% of the nightly rate.
What changed: the answer leads. The conditional pricing that made the old reply sound like a refusal moved to text.
Illustrative example.Across all eight, the same four moves repeat: name first, numbers out, one closing question, and no letter phrases like "kindly". They work just as well for voice note appointment reminders and AI voice notes for lead follow-up.
An eight-step check before any script goes live
- Name the one ideaWrite it in five words. If you cannot, the script holds two ideas.
- Write the text companion firstEvery number, date, address, reference and link goes here before the voice script exists.
- Draft the voice script under 65 wordsGreeting and name, the answer, one next step.
- Read it aloud at a calm paceTime it. Over 25 seconds, cut or split.
- Mark every name and termAnything not on the pronunciation sheet gets a respelling, or comes out of the audio.
- Strip the numbersRemove every number the voice does not need, and check the rest match the text exactly.
- Run the never listNo card details, codes, health details, ID numbers, links or unapproved offers.
- Listen on a phone speakerIf any phrase sounds read out rather than said, rewrite it in the words you would use face to face.
How Learnmind builds voice notes that sound like your team
Learnmind is a Dubai company that builds bespoke AI systems for customer conversations. Our WhatsApp AI agent replies in under ten seconds, day or night, in the customer's own language and the brand's voice, with Arabic detected automatically. When the moment calls for it, it replies with a voice note. In the words of our WhatsApp agent page, that is "a warm spoken reply, built to sound like a person, not a script".
It also handles the voice notes customers send, which we cover in how an AI understands customer voice notes.
The voice comes from the persona. On day 1 we hold a 15-minute call and map your services, pricing logic, common questions and brand voice; you prepare nothing. On days 2 to 5 we write the persona: its name, tone, discovery questions, objection handling and escalation rules. The agent is then built and tested on real questions, and is live from about day 14. The full timeline is on how it works.
| A rule from this guide | Where it lives in a Learnmind build |
|---|---|
| Sound like your best person, not a template | The named persona, with your tone, written in days 2 to 5 |
| Only approved prices and promises | Answers come from facts the business approves |
| Mirror the customer's language | Replies in the customer's language; English and Arabic, with Arabic auto-detected |
| Know when a person should take over | Handover with the whole conversation attached, so the customer never repeats themselves |
| See exactly what was said | An operator view where the business sees every conversation |
| Change a price, change the script | Updates are made the same day |
One example from our own work. Alcaz Media, a UAE performance marketing agency, runs a Learnmind assistant on its WhatsApp. It answers leads within seconds around the clock, qualifies them, shares pricing and books strategy sessions. It handles images and voice notes, and deflects spam. The case study is on our results page.
To hear it on your own number first, the free 14-day WhatsApp AI trial runs on your business's own WhatsApp number. The trial is free; Meta bills its own message fees to you directly. For the wider picture, start with AI voice notes for business.
Frequently asked questions
How long should an AI voice note be?
Aim for 10 to 25 seconds, which is about 25 to 65 words at 155 words a minute. In a 2007 text-to-speech study, a faster 178 words a
minute lowered comprehension. Past 30 seconds, split the note or send text.
How do I make an AI voice note sound human?
Write for the ear: short sentences, the customer's name first, one idea and one question at the end. Use at most one filler word such
as "so", never before a number. Then listen on a phone speaker and rewrite anything that sounds read out.
Should I put prices in a voice note?
Say a price only if it is short and simple, and always repeat it in a text message underneath. Long prices, ranges and conditions
belong only in text.
How do I get an AI voice to pronounce names correctly?
Keep a pronunciation sheet with a phonetic respelling for every name, place and product the voice says often. Where the engine
supports SSML, the phoneme element gives an exact pronunciation and the sub element swaps in a spoken alias. If a customer's name is
uncertain, leave it out of the audio.
Can a business send a WhatsApp voice note to someone who has not messaged first?
No. Meta's documentation says free-form messages, including audio, can only be sent inside the 24-hour customer service window that
opens when a customer messages or calls. Outside it, only pre-approved template messages can be sent.
Do I have to tell customers a voice note was made by AI?
In the EU, Article 50 of the AI Act has applied since 2 August 2026: people must be informed they are interacting with an AI system.
Whether that reaches a given Dubai business is a legal question. Either way, if a customer asks whether they are talking to a person,
the answer should be true.
What should a business voice message script include?
A greeting with the customer's name, the reason for the note, one answer, and one next step or question. Prices, dates, addresses,
references and links go in a text companion. Nothing private goes in at all.
Does a voice note cost more than a text message on WhatsApp?
From 1 October 2026, Meta charges for service messages after 1,000 free per business number a month, at USD 0.0157 each in the UAE. A
voice note and its text companion count as two messages. For most businesses the companion adds a few US dollars a month.
Who builds WhatsApp AI agents with voice notes in Dubai?
Learnmind.ai, based in Dubai, builds bespoke WhatsApp AI agents that reply in English and Arabic and send voice notes when the moment
calls for it. A build starts with a 15-minute call and is live from about day 14.
Want WhatsApp voice notes that sound like your best person?
We build WhatsApp AI agents that speak in your brand's voice, in English and Arabic, from facts you approve. A 15-minute call is enough to map it.
Sources
- ResearchJones, Berry and Stevens, Synthesized speech intelligibility and persuasion: speech rate and non-native listeners, Computer Speech and Language, 2007
- ResearchYuan, Liberman and Cieri, Towards an Integrated Understanding of Speaking Rate in Conversation, Interspeech 2006
- ResearchNelson Cowan, The Magical Mystery Four: How is Working Memory Capacity Limited, and Why?, 2010
- ResearchMichal Marmorstein, Why say 'hi'? Framed openings in Hebrew WhatsApp messaging, Studia Fennica Linguistica, 2023
- ResearchJana Tschannen, Multimodale Beendigungsstrategien in Audio-Postings, Bulletin suisse de linguistique appliquée, 2023
- ResearchW3C, Speech Synthesis Markup Language (SSML) Version 1.1, 7 September 2010
- ResearchW3C, Understanding WCAG 2.2 Success Criterion 1.2.1
- PlatformMeta for Developers, Audio messages
- PlatformMeta for Developers, Sending messages
- PlatformMeta for Developers, WhatsApp pricing, service message charges and UAE rates from 1 October 2026, checked 27 September 2026
- PlatformWhatsApp, Making voice messages better, 30 March 2022
- PlatformWhatsApp, Introducing voice message transcripts, 21 November 2024
- LawEU AI Act, Article 50: transparency obligations
- RegulatorEuropean Commission, Quick facts: transparency rules for AI systems
- RegulatorUAE Government portal, Consumer protection, Federal Law No. 15 of 2020
- LearnmindThe Learnmind WhatsApp agent
- LearnmindHow a Learnmind agent goes live
- LearnmindLearnmind results, the Alcaz Media case study
Written by Edmund Gay, Learnmind.ai, Dubai. Research findings, WhatsApp limits, prices and legal text are as published in the sources above. The 10 to 25 second target, the rules, the respellings and the eight scripts are our own interpretation.