Text-to-Speech for Customer Service: IVR and Phone Prompts

In customer service, text-to-speech is used mostly for the voice callers hear before they reach a person: phone menus (IVR), greetings, hold messages, after-hours notices and outage announcements. It gives you one consistent voice across every prompt, in every language you support, and lets you change a single message in minutes instead of booking a new recording. The quality of the result depends far more on how the prompts are written than on the voice you choose.

This guide covers how to write prompts callers can follow, the audio formats phone systems expect, handling numbers and names, multilingual menus, disclosure rules, and what it costs.

Where does text-to-speech fit in customer service?

Use	Pre-generated or live?	Typical update frequency
Main greeting and phone menu	Pre-generated	A few times a year
Hold and queue messages	Pre-generated	Monthly or per campaign
Opening hours, holiday closures	Pre-generated	Several times a year
Outage or incident notice	Pre-generated on demand	When something breaks
Voicemail and after-hours greetings	Pre-generated	Rarely
Personalized messages (order status, appointment times)	Generated per call or per batch	Constantly

Most teams start with the first five, which are just a set of audio files uploaded to the phone system. Personalized messages need an integration, covered below.

How do you write IVR prompts callers can actually follow?

Callers can't scroll back. Everything has to work in a single listen, often on a phone speaker in a noisy room.

Rules for phone menus

• Say the option before the key. "For billing, press 1" lets callers recognize their need first and then hear the number. "Press 1 for billing" makes them hold a number in memory while waiting to find out what it's for.
• Keep it to four or five options. More than that and callers forget the first ones. Use a second level if you must.
• Put the most common reason first. Look at why people actually call.
• Don't make people sit through marketing. A promotion before the menu adds time to every single call.
• Always offer a way to repeat and a way out. "To hear these options again, press star. To speak to someone, press 0."
• Leave a pause after the menu. Give callers a couple of seconds to decide before the system times out.

Before and after

Before:

"Thank you for calling. Please listen carefully as our menu options have recently changed. Press 1 if you would like to inquire about an existing order, the status of a delivery, or a return. Press 2 for…"

After:

"Thanks for calling Northwind. For an order, delivery or return, press 1. For billing, press 2. For technical help, press 3. To speak to someone, press 0."

The second version is half as long and every option takes one breath.

How do you get numbers, dates and names read correctly?

Phone prompts are full of things a voice can read in more than one way. For pre-generated prompts, write them out exactly as they should be spoken.

Written	Problem	Write it as
0800 123 4567	Could be read as "eight hundred" or a huge number	"oh eight hundred, one two three, four five six seven"
1-800-555-0100	Grouping and the "1"	"one eight hundred, five five five, zero one zero zero"
03/04	March 4 or April 3?	"the fourth of March"
9am–5pm Mon–Fri	Abbreviations and dashes	"nine a.m. to five p.m., Monday to Friday"
Order AB12	"Ab twelve"?	"A, B, one, two"
€1,250.50	Fine in most voices, but check	"one thousand two hundred and fifty euros fifty"

Brand names, product names and place names should go in the pronunciation dictionary once, so every prompt says them the same way. This matters more on the phone than anywhere else, because a mispronounced company name in the greeting is the first thing every caller hears.

What audio format do phone systems need?

Telephone audio is narrowband: traditional phone lines carry roughly 300 to 3,400 Hz and sample at 8 kHz. Most phone systems accept one or more of these:

• WAV, 8 kHz, 16-bit, mono (PCM)
• WAV, 8 kHz, mono, μ-law (common in North America and Japan) or A-law (common in Europe and most other regions)
• MP3, which some cloud phone platforms accept and convert themselves

Leelo exports MP3 and WAV (export options). Export WAV, then use your phone platform's upload tool or an audio converter if it needs 8 kHz μ-law or A-law. Two practical tips:

• Choose a voice that survives narrowband. Voices with a clear mid-range carry well over the phone. Very breathy or very deep voices lose detail. Test candidates by playing them through an actual phone call.
• Keep music subtle. Hold music competes with speech once both are squeezed into phone bandwidth. If you build hold messages with a music bed, keep the music well under the voice (adding sounds and music).

How do you set up a phone menu in several languages?

1. Offer the language choice in each language. "For English, press 1. Para español, oprima dos. Pour le français, appuyez sur trois." Each phrase should be spoken by a native voice in that language.
2. Translate prompts, don't just transliterate them. Phone conventions differ: Spanish IVRs often use "oprima" or "marque" rather than a literal "press."
3. Cast one voice per language and keep it. Leelo has voices in more than 80 languages; for example, audition several Spanish voices and pick the one your callers will find most natural.
4. Keep the menu structure identical across languages, so key 2 means billing in every language and support staff can troubleshoot easily.

Do you have to tell callers the voice is AI?

Disclosure is good practice, and in some cases it's a legal requirement.

• Outbound calls: in the US, the Federal Communications Commission ruled in 2024 that AI-generated voices count as "artificial" voices under robocall rules, which generally require the recipient's prior consent.
• AI systems that converse with people: the EU AI Act requires that people be told when they're interacting with an AI system, unless it's obvious.
• Inbound menus: pre-recorded menus are a familiar format, but if a synthetic voice could be mistaken for a live agent, make it clear early.

The rules differ by country and change quickly, so check with your legal team. This is general information, not legal advice.

Should your brand use a stock voice or its own?

• Stock voice: fastest. Pick one from the library and use it everywhere.
• Designed voice: create a voice from a description ("friendly, calm, mid-thirties, clear, unhurried") so no other company sounds like you.
• Cloned voice: if your company already has a familiar voice, such as a staff member who recorded the old greetings, voice cloning can recreate it from a short recording, with that person's written consent. Agree in advance what happens if they leave.

How do you keep prompts up to date?

The main reason teams switch to text-to-speech is updates. A holiday closure used to mean booking studio time; now it's a line edit.

Worked example: a holiday closure

1. Add a new line under the "opening hours" line, with the same speaker.
2. Change the text: "We're closed on Monday the twenty-fifth. We'll reopen on Tuesday at nine a.m."
3. Regenerate only that line in the same voice and settings, so it matches the rest of the menu.
4. Export and upload it to the phone system with a schedule for the closure dates.

Automating with the API

If messages depend on live information, such as an outage notice or a queue that's longer than usual, use the API to generate audio from text in your own systems and webhooks to know when a file is ready. For speech generated during a live call, test the delay end to end with your phone platform before you commit.

What does it cost to voice a phone system?

Leelo uses credits: 1 credit is 1 word in a standard voice, 1.5 for premium voices, and 5 for your own designed or cloned voices. Failed jobs return their credits. A worked example:

• A typical IVR: 40 prompts averaging 15 words is about 600 words, so 600 credits with a standard voice. That fits in the Free plan's one-time 1,000 credits, which is enough to build and test a complete menu.
• The same menu in four languages: about 2,400 credits.
• With a designed brand voice: about 3,000 credits for one language.
• API access starts on the Pro plan ($99 a month, 600,000 credits, 3 seats).

Commercial use of the audio, which includes playing it to your customers, is allowed on paid plans. See pricing for all plans, and the phone menus use case page for more examples.

Frequently asked questions

Can I use text-to-speech for my company's phone menu?

Yes. Write the prompts, generate them in one voice, export WAV, and upload them to your phone system. Most cloud and on-premises phone systems accept WAV files, sometimes after conversion to 8 kHz.

What's the best voice for an IVR?

A clear, mid-range voice at a slightly slower pace than normal conversation, because phone audio removes detail. Test two or three candidates over a real phone call, not on studio headphones, and pick the one that's easiest to understand.

How long should a phone greeting be?

As short as possible: the company name, an optional one-line message, then the menu. Every extra second is paid by every caller, every time. If you need to announce something, put it after the menu or in the queue message.

Do I need to re-record the whole menu when one option changes?

No. With a line-by-line studio you edit and regenerate only the changed prompt in the same voice and settings, so it matches the rest. Then you replace that single file on your phone system.

Can callers tell it's an AI voice?

Some will, especially on longer messages. That's fine if the prompts are clear and the voice is pleasant; callers mostly care about getting to the right place quickly. Where laws or your own policies require it, tell callers the voice is automated.