IVR Voice Gen

Combining GPT-Live-1 with pre-generated IVR prompts

Facts about GPT-Live-1 verified 11 September 2026 against the official documentation linked at the foot of this page.

The short version

Give the conversation to GPT-Live-1 and keep everything fixed as pre-generated audio. Greetings, menu options, opening hours, hold announcements, outage notices and legal disclaimers are the same words on every call — they do not need a model listening and speaking in real time, and they are the parts callers are least forgiving about getting wrong.

This is not a compromise between two competing products. The two halves answer different questions: what does this caller want is a conversation, and read me back my reference number is a readout. A phone system that treats both as the same problem overpays for one of them and gets the other wrong.

Which part goes where

Decision table for a typical inbound support line
Moment in the call Use Why
Greeting and main menu Pre-generated audio The same words on every call. Generating once costs nothing per call, plays back identically, and the wording can be approved once and left alone.
Out-of-hours and holiday messages Pre-generated audio Fixed text, often legally worded, played to callers who cannot be helped anyway. Opening a billed voice session to read it out is spend with nothing in return.
Reading back a code, reference or IBAN Pre-generated audio or TTS with per-character control Accuracy matters more than fluency, and a cascaded path leaves you an intermediate text layer where digit grouping can be enforced. A speech-to-speech model gives you no text to correct.
Order or delivery status lookup GPT-Live-1 with a backend The caller asks in their own words and the answer comes from a tool call. This is what delegation is for, and the conversation can continue while the lookup runs.
Rescheduling an appointment GPT-Live-1 with a backend Several rounds of negotiation with interruptions and corrections. A menu tree cannot express it and a caller will not tolerate one that tries.
Open-ended questions GPT-Live-1 No fixed script exists, so there is nothing to pre-generate. Full duplex earns its cost here: the model can listen through a pause and respond to an interruption mid-sentence.

The call flow

One inbound call, with the billed voice session opening only for the part that needs it.

Hybrid IVR and GPT-Live-1 call flow An inbound call arrives over a carrier or SIP trunk and enters a fixed IVR tree built from pre-generated audio. Calls that the menu resolves end there with no voice-session charge. Calls that need a conversation open a GPT-Live-1 session, which delegates reasoning and tool calls to a backend model, and then ends or transfers to a human agent. Inbound call Carrier or SIP trunk Fixed IVR tree Pre-generated audio — greeting, menu resolved conversation Call ends No session charge GPT-Live-1 session $0.05/min, per second Backend model Billed separately Closing readout Code — pre-generated End or transfer Human agent if needed
The two shaded boxes are pre-generated audio; the billed voice session covers only the middle of the call.

Audio format notes

Telephony systems generally expect 8 kHz mono, either μ-law encoded WAV or 16-bit PCM, and that constraint applies to the pre-generated half of the system just as much as the live half. The convenient part is that both halves want the same thing: OpenAI documents that GPT-Live supports raw G.711 μ-law and A-law audio at 8 kHz over WebSocket, so when the provider stream already uses that codec and sample rate the audio can be forwarded without a conversion step.

Sources: Twilio: <Play> audio formats, Asterisk documentation, OpenAI: Telephony and SIP.

One practical consequence: generate the fixed prompts at the telephony sample rate rather than downsampling 24 kHz web audio at play time. Resampling artefacts land hardest on exactly the content that matters most here — sibilants and digit boundaries in a spoken reference code.

The cost angle

A pre-generated prompt is a one-off cost. Generate the menu once and it plays for every call that year at no per-call charge, whether that is a hundred calls or a million. A live voice session bills $0.05 a minute every time, and the meter runs through the silences as well as the speech.

That makes the arithmetic simple for the fixed parts of a call. Thirty seconds of greeting and menu, played on every one of ten thousand monthly calls, is roughly $250 a month of voice time if a live model reads it — for words that never change. The cost calculator works the same sum for your own call length and volume, including the backend model cost the voice rate does not cover.

Generate the fixed half

IVR Voice Gen is text-to-speech built for the pre-generated side of this diagram: menu options, opening hours, outage notices, and the readouts where digit strings, reference codes, dates and email addresses have to land in a form the caller can write down. It is in early access.

Generate your IVR menu prompts

Questions

Does IVR Voice Gen integrate with GPT-Live-1?

No. IVR Voice Gen generates spoken prompts as audio files and streams TTS for agents that compose their responses at runtime. GPT-Live-1 is a speech-to-speech model that speaks in its own voice and is called directly from OpenAI's API. The two can sit in the same phone system — one playing the fixed prompts, the other handling conversation — but there is no integration between them, and this page is independent analysis rather than a product announcement.

What audio format does the pre-recorded side need?

Telephony systems generally expect 8 kHz mono, either μ-law encoded WAV or 16-bit PCM. The live side lines up with this: OpenAI documents that GPT-Live supports raw G.711 μ-law and A-law audio at 8 kHz over WebSocket, so when your provider stream uses the same codec and sample rate, the bytes can be forwarded without conversion. Keeping both halves of the system on the same format removes a resampling step from the hot path.

Can GPT-Live-1 read a verification code correctly?

We do not have an answer we can stand behind. OpenAI describes strong alphanumeric handling, but the public documentation does not publish accuracy figures for digit strings, reference codes or IBANs, and we have not run an independent test of our own. Until one of those exists, treat it as unproven: put anything the caller has to write down on the pre-generated side, where the reading is fixed and you can hear it before it ships.

How does a call move from the menu to a live session?

The menu tree stays in your telephony layer — your IVR platform, or a media path through a provider such as Twilio, Telnyx, LiveKit or Daily/Pipecat. When the caller picks an option that needs a conversation, that layer opens a live session and bridges the audio into it. Inbound calls have a documented SIP path; outbound calling is not supported through the live sessions endpoint and has to go through a partner integration.

Related

Sources

IVR Voice Gen is not affiliated with OpenAI. GPT-Live is a trademark of OpenAI. This site does not integrate GPT-Live-1; the pages in this section are independent analysis based on the public documentation linked above.