An IVR voice generator that reads numbers the way people say them.
Phone systems spend most of their time reading back things that are not words: booking references, callback numbers, verification codes, delivery windows. Those are exactly the strings ordinary text-to-speech gets wrong.
What your system sends
Your booking reference is A4-92-KJ7.
What the caller hears
Your booking reference is A four, ninety-two, K J seven.
Your booking reference is A — four — nine — two — K — J — seven.
What your system sends
Your slot is 14h30 on 03/04.
What the caller hears
Your slot is fourteen thirty on oh three oh four.
Your slot is half past two in the afternoon on the third of April.
What your system sends
Transfer to FR76 3000 6000 0112 3456 7890 189.
What the caller hears
Transfer to F R seventy-six, three thousand, six thousand, one hundred twelve, three thousand four hundred fifty-six, seven thousand eight hundred ninety, one hundred eighty-nine.
Transfer to F — R — seven six — three zero zero zero — six zero zero zero — zero one one two — three four five six — seven eight nine zero — one eight nine.
Recordings: the ordinary column is a general-purpose text-to-speech engine given the raw text as-is. The number-aware column is the same text after IVR Voice Gen's normalisation step, spoken by Gradium TTS. Both generated 2 September 2026.
What breaks a phone voice agent?
Five things, and they are always the same five. Every one is invisible in a demo built from clean sentences, and every one shows up on the first real call.
- Digit strings get grouped like quantities
- A reference number is not a number. Ordinary models read 1222 as “one thousand two hundred twenty-two”, which no caller can write down.
- Letters and digits collide
- Mixed strings like A4-92-KJ7 need per-character delivery with pauses. Run them together and the caller asks you to repeat every time.
- Punctuation disappears
- Dots, hyphens and the @ symbol carry meaning inside an email address or a URL. Dropping them produces an address that does not exist.
- Regional formats get read literally
- 14h30, 03/04 and £1,250.00 mean different things in different locales. A literal reading is wrong in every one of them.
- The caller has no screen
- On a web page a mispronunciation is cosmetic. On a phone call it is a failed task — there is no text to fall back on.
What do people generate with it?
Phone menu prompts
Main greeting, department options, hold messages, closed-for-the-day.
Appointment reminders
Date, time and address read back in the caller’s local format.
Order confirmations
Order references and delivery windows, digit by digit.
Verification codes
One-time codes delivered by voice with per-character pacing.
Outage notices
Recorded updates you can regenerate in a minute instead of a day.
After-hours messages
Opening hours and emergency numbers, correct in five languages.
How do streaming TTS models compare on speed and accuracy?
Time to first audio decides whether a call feels like a conversation. Word error rate decides whether the caller gets the right number, but the published rate is measured on ordinary transcripts, not on the digit strings and reference codes a phone system spends most of its time reading. That gap is the reason this product exists.
| Model | Time to first audio (median) | Word error rate |
|---|---|---|
| Inworld TTS-2 | 170 ms | 4.8% |
| ElevenLabs Flash v2.5 | 186 ms | 6.8% |
| Gradium TTS (beta) runs IVR Voice Gen | 214 ms | 5.5% |
| Cartesia Sonic-3.5 | 272 ms | 5.9% |
| Deepgram Aura-2 | 292 ms | 5.4% |
| Gradium TTS (default) runs IVR Voice Gen | 385 ms | 4.7% |
Figures from the Coval TTS benchmark, read 1 September 2026. Coval updates that board continuously and the order changes, so check the live figures rather than trusting this snapshot. The six models above are the streaming ones most often put in front of a phone line; the full board lists thirty. IVR Voice Gen is built on Gradium TTS.
IVR text-to-speech: common questions
Why does text-to-speech read phone numbers wrong?
Most models normalise digits as quantities before synthesis, so 1222 becomes “one thousand two hundred twenty-two”. Phone numbers, reference codes and account numbers need per-digit delivery with grouping pauses instead, which requires the model to recognise what kind of string it is looking at.
Which languages can I generate IVR prompts in?
English, French, German, Spanish and Portuguese, with the number, date and currency conventions of each locale applied automatically rather than through manual text preparation.
What audio format do Twilio and Asterisk need?
Telephony systems generally expect 8 kHz mono, either μ-law encoded WAV or 16-bit PCM. Exports are available in both alongside standard 24 kHz WAV and MP3 for web playback.
Sources: Twilio: <Play> audio formats, Asterisk documentation
Does it work with LiveKit, Pipecat or Vapi?
Yes. Prompts can be generated ahead of time as audio files for a fixed IVR tree, or streamed over WebSocket for agents that compose their responses at runtime.
Sources: LiveKit Agents, Pipecat, Vapi
How fast is it in a live call?
Time to first audio is what matters on a call, because telephony network overhead is added on top of whatever the model takes. Read on 1 September 2026, the Coval TTS benchmark put Gradium TTS, the engine this runs on, at 214 ms median time to first audio. Coval updates that benchmark continuously, so treat any single figure as a snapshot rather than a fixed specification.
Sources: Coval TTS benchmark
Can I use my own brand voice?
Voice cloning from a short reference recording is supported, so an existing recorded greeting can be extended to new prompts without booking the original voice artist again.
Get early access
Tell us the language you need and we will send you a link when your locale opens. No newsletter.