IVR Text to Speech: Voices, Formats and Cost (2026)

IVR text to speech is a practical way to produce phone menu prompts. Render fixed prompts once as telephony audio, use live synthesis only for text that changes per caller, and check two things before launch: per-character billing and the consent rules for AI voices on outbound calls.

Static prompts and dynamic text

An IVR, or interactive voice response system, plays audio at set points in a call. Most of that audio is fixed: the greeting, the main menu and the line that offers a human agent. Some of it is dynamic, such as an order number or an account balance that differs for every caller.

The two kinds of prompt suit different workflows. Fixed prompts can be rendered once, saved as audio files and uploaded to the phone system. Dynamic text needs synthesis on every call, and that is where per-character billing starts to matter. Plivo's IVR guide draws the same line, noting that TTS suits many prompts and frequently changing options.

Recorded voice actor or text to speech?

Both options work. The right one depends on how often the script changes, how many languages you need and whether the voice must match an existing brand. The table lists the published price points we could verify. The voice actor figures come from two sellers' rate cards. They are estimates, not union scale.

Prices as of October 17, 2026.

| Option | Best suited to | Published price point | |---|---|---| | Voice actor, Voice Crafters rate card | A brand voice recorded once and kept | IVR and on-hold: $50 to $200 per finished minute, $100 to $200 minimum | | Voice actor, VoiceBros marketplace | Small prompt packs | Standard five-prompt IVR pack: $80 to $250; English, Spanish and French pack: $200 to $600 | | Text to speech, Onda Lite on Starter | Frequent script edits and many languages | About $0.03 for a five-prompt pack of 100 words (derived) |

A recorded actor gives you one consistent human performance for a script that rarely changes. Text to speech is faster to revise: change a line, re-render it and replace one file. Neither option is cheaper in every case, so price each prompt set by its volume and how often it changes.

Telephony audio formats

Phone networks carry narrowband audio, so the file format matters more than it does on a website. Amazon Connect recommends "8 KHz .wav files that are less than 50 MB and less than 5 minutes long." It also notes that higher-rate files are down-sampled to 8 kHz "because of PSTN limitations," which can reduce quality.

| Provider | Telephony options listed | Sample rate | |---|---|---| | Amazon Polly | mu-law, A-law, PCM | 8000 Hz for mu-law and A-law; 8000 or 16000 Hz for PCM | | Google Cloud Text-to-Speech | MULAW and ALAW, with a WAV header | 8-bit companded G.711 samples | | Azure AI Speech | Raw and RIFF 8-bit mu-law and A-law; 16-bit PCM | 8 kHz | | Gemini API speech generation | audio/mulaw and audio/alaw | Configurable; 8000 in Google's examples |

Before you render with any provider, confirm the file format your phone platform accepts, and convert the file if it needs something else.

What IVR text to speech costs

Most providers bill per character. Telephony platforms can add their own per-minute charges on top. Twilio's <Say> verb is a useful example because it bills in 100-character blocks, so short phrases round up.

Prices as of October 17, 2026. Twilio's pricing page says it is current as of August 2026 and quotes per 100 characters. The per-million figures below are our arithmetic.

| Service and tier | Price per 1M characters (derived for Twilio) | |---|---| | Twilio Say, Standard | $8 | | Twilio Say, Neural | $32 | | Twilio Say, Generative | $130 | | Amazon Polly list, Standard | $4 | | Amazon Polly list, Neural | $16 | | Amazon Polly list, Generative | $30 |

Rendering once is the cheapest pattern for a prompt that repeats. Take a 200-character prompt, which is two Twilio billing blocks, played 10,000 times. That is 2 million characters: $64 through Twilio Neural, or $32 at Polly's Neural list price, before any free tier (derived). Rendering the prompt once costs 200 characters, about $0.0032 at Polly Neural (derived).

A 20-second prompt of about 50 words is roughly 300 characters, or three Twilio blocks, which comes to about $0.0096 on Neural (derived). Inside Amazon Connect, the default Polly Neural and Standard voices are free, while Generative voices and custom Brand Voices are charged, as of October 17, 2026.

SSML support in IVR platforms

SSML, the speech markup language, lets you control pauses and pronunciation. Support varies by platform, and an unsupported tag can break a prompt.

• Amazon Connect supports speak, break (pauses up to 10 seconds), lang, mark, p, phoneme, prosody, say-as, sub, w and a whispered effect. Unsupported tags are ignored.
• Twilio says unsupported SSML "can cause an error and fail the <Say> instruction," so test each voice tier before launch.
• On Twilio, ElevenLabs voices accept only break and phoneme, and only for en-US.
• Character limits per <Say> are 4,000 for Basic, 3,000 for Polly (SSML tags excluded), and 5,000 for Google and ElevenLabs (SSML included).

Script and menu rules that work

Plivo's IVR guidance is vendor advice from 2023, not a study, but its checklist is simple and easy to test:

1. Keep the main menu to four to six options, with the most-used choice first.
2. Keep the greeting to eight seconds or less, and each option to about four seconds.
3. Name the task before the key: "to speak to sales, press or say 1."
4. Let callers interrupt the prompt, which is called barge-in.
5. Accept both keypad and spoken input.
6. Always offer a route to a human agent.

Rules for AI voices on phone calls

The rules depend on the direction of the call. In the United States, the FCC's Declaratory Ruling of February 8, 2024 (FCC 24-17) says the Telephone Consumer Protection Act's limits on "artificial or prerecorded" voices cover current AI technologies that generate human voices. The ruling applies to outbound calls. Callers need prior express consent, which is written in some cases, must identify the party responsible for the call, and must offer an opt-out. It does not address the inbound IVR prompts that customers hear after they dial in. Check CG Docket No. 23-362 for later FCC action.

In the EU, the Commission says the AI Act's transparency rules apply from 2 August 2026. If your system is a conversational AI agent, people must be told they are talking to an AI. Providers must also apply a machine-readable mark to synthetic audio, with a grace period to December 2026 for systems already on the market. The Commission's summary does not say whether a fixed menu of recorded prompts counts as an AI system, so treat that question as open and check with counsel. This section is not legal advice.

Building a multilingual prompt set

Many menus need the same options in several languages. OndaLabs offers 30 prebuilt voices that can be used in every supported language, and dedicated pages for 37 languages, such as Spanish and French. Have a native speaker review each script before you render it, and give each prompt a stable file name so an update replaces one file rather than the whole set.

Where OndaLabs fits

OndaLabs is a voice studio for writing, auditioning and rendering prompt audio. Its Free plan includes 2,000 credits a month, about 2 minutes, and Onda Lite costs 1 credit per character. Commercial use is included from the Starter plan at $9 a month, which carries 170,000 credits, about 189 minutes of speech as listed. OndaLabs is not a phone platform, and nothing in this article suggests it connects to one. Audition voices on the text to speech page and compare plans on the pricing page.

For related reading, see Amazon Polly alternatives for other voice providers, and SSML versus expressive tags for directing delivery in a script.

Frequently asked questions

Can you use text to speech for an IVR?

Yes. Most IVR prompts are fixed, so they can be rendered once and saved as audio files. Live synthesis suits text that changes per caller, such as balances or order numbers. Check that your phone platform accepts the file format you render.

What audio format does an IVR need?

Amazon Connect recommends 8 kHz .wav files under 50 MB and under 5 minutes long. Other platforms list mu-law or A-law encodings at 8 kHz. Confirm the requirement with your phone provider before you render.

How much does text to speech cost per call?

Cost depends on characters and on the platform's own charges. As of October 17, 2026, Polly Neural lists at $16 per 1 million characters, and Twilio's Neural Say tier is $32 per 1 million characters (derived). Rendering a repeated prompt once avoids paying for it on every call.

Is it legal to use an AI voice on phone calls?

For outbound calls in the US, the FCC's 2024 ruling says AI-generated voices fall under the TCPA's rules for artificial or prerecorded voices, so consent and opt-out requirements apply. The ruling does not cover inbound IVR prompts. Check CG Docket No. 23-362 and get legal advice for your situation.

Should IVR prompts be recorded by a voice actor or generated with TTS?

Use a voice actor when one stable brand voice matters most and the script rarely changes. Use text to speech when prompts change often or you need many languages. Many teams combine the two, recording the greeting and generating the rest.