Big Idea

The argument in one line.

OpenAI's GPT Live-1 speech-to-speech model removes the stitched speech-to-text-to-speech pipeline that made older voice agents sound robotic, and an AI coding agent can now wire it to a live phone number without writing code.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

Someone building or reselling AI phone receptionists who wants to know whether speech-to-speech finally beats a stitched speech-to-text pipeline.
A non-technical builder who wants to see an AI coding agent wire a voice stack together without writing code.
Anyone evaluating LiveKit, Telnyx, or GPT Live-1 as pieces of a voice-agent stack and wants to see them connected live.

SKIP IF…

You need a production-ready receptionist script today. This is a fast proof-of-concept, not a hardened call-flow build.
You want a deep technical explanation of the GPT Live-1 API itself rather than a click-through setup guide.

TL;DR

The full version, fast.

OpenAI's GPT Live-1 is a true speech-to-speech model, meaning it listens to raw audio instead of a transcript, so it catches pauses, laughs, and tone the way a stitched-together speech-to-text-to-speech stack never could, and it now runs reliably enough on a real phone number to replace it. The walkthrough uses Codex, OpenAI's coding agent, to connect GPT Live-1 to LiveKit (which bridges the model to a phone line) and Telnyx (which provides the number), all through plain-language prompts instead of manual code. The result is a working AI receptionist that takes a live call, and a second prompt adds a tool call so the agent can look up a real-time gold price mid-conversation, proving the stack can do more than talk.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 02:29

01 · What's New? (Demo)

Plays a call built on GPT Live-1 back to back with an older three-model voice-agent setup to show the difference in latency and tone.

02:29 – 03:02

02 · What I'm Covering

States the plan: build a receptionist voice agent on a real phone number using Codex, LiveKit, and the new GPT Live-1 model, no coding required.

03:02 – 03:52

03 · Codex Setup

Points to OpenAI's Codex desktop app as the tool that will build the entire system from prompts, and covers its free and paid tiers.

03:52 – 06:10

04 · LiveKit Setup

Introduces LiveKit as the platform that connects the voice model to a phone number, explains the free vs. cloud tiers, and recommends cloud so the agent stays reachable when the computer is off.

06:10 – 07:07

05 · Telnyx Phone Number Setup

Picks Telnyx as the phone number provider for non-US numbers, noting LiveKit itself can sell US numbers directly.

07:07 – 09:34

06 · Prompting Codex

Opens the Codex desktop app, creates a project folder, and dictates a single voice prompt describing a plumbing-company receptionist built on LiveKit and GPT Live-1.

09:34 – 12:52

07 · Building the Voice Agent

Watches Codex research LiveKit's docs and API, generate LiveKit API keys, and wire the OpenAI key into the project while narrating each authentication step.

12:52 – 14:45

08 · Connecting Telnyx to Codex

Creates a Telnyx account, buys a number, generates a Telnyx API key, and hands it to Codex so it can finish linking the phone number to the LiveKit agent.

14:45 – 16:36

09 · Live Call Test

Calls the finished agent live: it introduces itself as a plumbing company's AI receptionist and works through a leaking-shower-head scenario that escalates into a flooding emergency.

16:36 – 17:13

10 · Results & Refinement Notes

Reacts to the call, noting the agent correctly told the flooded caller to move to safety and call emergency services, and flags that its 'demo agent' framing would need removing for real use.

17:13 – 18:18

11 · Adding Tool Calls

Dictates a second prompt telling Codex to add a function call that hits a public API for the live gold price and reads the result back mid-conversation.

18:18 – 19:15

12 · Live Test: Function Call

Calls the agent again and asks for the live gold price; it triggers the new tool call and reads back a real, timestamped quote.

19:15 – 19:32

13 · Final Thoughts

Wraps with a call to subscribe and points to a companion video building websites with GPT-6 Astra.

Atomic Insights

Lines worth screenshotting.

OpenAI's GPT Live-1 is a true speech-to-speech model: it listens directly to audio instead of reading a transcript, so it reacts to laughs, pauses, and tone like a real person would.
Older voice agents stitch together three separate models (speech-to-text, a text model, text-to-speech), which adds latency and gives every reply a robotic quality.
Speech-to-speech models have existed for a while, but they rarely held up over a real phone line until now, where latency and dropped calls finally stopped being a problem.
An AI coding agent can build and deploy a full voice-agent stack, connect a phone number, and add live tool calls from plain-language prompts, with no manual coding required.
LiveKit is the same infrastructure OpenAI uses to run ChatGPT's own Advanced Voice Mode, so building on it puts a hobbyist project on the identical stack as OpenAI's own production voice product.
LiveKit's free tier covers one agent and one connected phone number, enough to build and test an entire voice-agent stack before paying anything.
A phone number from a telephony provider like Telnyx costs about a dollar a month, and test calls run a couple of cents each, so a full working demo costs under five dollars to build.
Tool calls turn a voice agent from a conversation into a system: adding a live gold-price lookup took one follow-up prompt and worked on the very next call.
The agent handled an escalating flood scenario mid-demo by telling the caller to move away from the danger and contact emergency services, showing basic safety reasoning built into the base response.
Self-hosting a voice-agent platform is free but only runs while the host computer stays on, which is why a cloud tier is the practical choice for an agent meant to answer calls at 2 a.m.
Naming a specific tool's CLI in a prompt, rather than just the product name, is what gets an AI coding agent to install the right tooling instead of guessing.
A demo AI receptionist still needs a human fallback: the system in the video introduced itself as a demo agent and had limits it couldn't work around yet.

Takeaway

Voice Agents Now Sound Human And Ship Fast

THE VOICE STACK

A true speech-to-speech model removes the stitched pipeline that made older voice agents sound robotic, and a coding agent can now assemble the whole phone-answering stack from plain prompts.

01What's New? (Demo)

A true speech-to-speech model reacts to tone, pauses, and laughter because it processes raw audio, not a transcript, which is why it sounds distinctly less robotic than older voice agents.
Stitching together speech-to-text, a language model, and text-to-speech adds latency at every hop, and that compounding delay is what makes older voice agents feel stilted on a live call.

02What I'm Covering

A phone-based AI receptionist is built from three separate pieces, the voice model, a real-time routing platform, and a phone number provider, and each can be swapped independently.

03Codex Setup

A coding agent can build and deploy a full technical system from spoken or typed instructions, which means fluency with prompting matters more than fluency with any one dashboard.
Free tiers of coding-agent tools come with usage caps, so budget for an eventual upgrade if a build session runs long.

04LiveKit Setup

Free tiers of infrastructure platforms are often enough to build and fully test a working prototype before any money changes hands.
Self-hosted infrastructure is free but only available while the host machine stays powered on, so anything meant to be reachable around the clock belongs on a cloud tier instead.
Choosing infrastructure that a major AI lab already runs its own production product on is a reasonable way to bet on reliability without doing the vetting yourself.

05Telnyx Phone Number Setup

International phone numbers and US numbers can come from different providers, and picking the wrong one adds an unnecessary extra signup step.
A working phone number for a voice-agent prototype costs about a dollar a month, so hardware cost is negligible next to the model and infrastructure choices.

06Prompting Codex

Naming the exact tool by its formal name, such as a CLI rather than just the product name, in a prompt to a coding agent gets more reliable setup behavior than a vague reference.
A single, specific prompt describing the use case, the stack, and the integrations needed is enough to kick off an entire multi-service build.

07Building the Voice Agent

A coding agent can read a platform's own documentation live and translate it into working setup steps, which removes the need to manually read API docs.
Sensitive credentials like API keys should be pasted into a secrets file the agent points you to, not directly into a chat window, so they never pass through the model.

08Connecting Telnyx to Codex

Connecting a third-party service to an AI-built system is usually just generating an API key and handing it to the coding agent, no manual integration code required.
Adding a small balance to a telephony account before testing avoids a failed call breaking the flow of a build session.

09Live Call Test

Testing a new voice agent with an escalating, high-stakes scenario, not just a scripted happy path, is what reveals whether it can actually reason, not just recite.
A convincing voice agent still needs a verified safety response for emergencies, since callers will occasionally test edge cases like this by accident.

10Results & Refinement Notes

A first working build will carry leftover scaffolding, like a 'this is a demo agent' disclaimer, that needs a follow-up prompt to remove before it's presentable.
Conversational refinement is an ongoing pass, not a one-shot outcome; plan for a second and third prompt round after the first working version.

11Adding Tool Calls

A voice agent that can only talk back is not useful for real business tasks; tool calls are what turn conversation into action.
Adding a new capability to an already-working voice agent is a single follow-up prompt, not a rebuild, when the underlying agent was built by a coding agent in the first place.

12Live Test: Function Call

Verifying a new tool call with a real, checkable answer, like a live price with a timestamp, is a fast way to confirm the integration is pulling live data and not a cached or hallucinated value.

Glossary

Terms worth knowing.

Speech-to-speech (S2S): A model architecture that processes and generates raw audio directly, without converting speech to text and back, so it preserves tone, pauses, and emotion that a text-based pipeline loses.
LiveKit: An open-source real-time voice and video infrastructure platform that connects an AI model to a phone line or app, either self-hosted for free or run on a paid cloud tier.
Telnyx: A telecom provider used to buy and manage phone numbers and route calls into a voice-agent platform.
Codex: OpenAI's coding agent, used here to write and deploy an entire voice-agent stack from plain-language instructions instead of manual code.
Tool call: A function a voice agent can trigger mid-conversation to pull outside data, such as a live price or a calendar slot, instead of only generating spoken replies.

Resources

Things they pointed at.

00:00toolLiveKit ↗

03:02toolCodex (OpenAI) ↗

06:22toolTelnyx ↗

19:10toolGoldAPI

Quotables

Lines you could clip.

00:35

“this right here is an AGI moment for voice AI”

bold, quotable claim stated plainly, works as a cold open→ TikTok hook↗ Tweet quote

01:37

“the model itself is actually listening to your voice instead of reading a transcript”

clean one-line explainer of the core technical shift→ IG reel cold open↗ Tweet quote

16:11

“please move everyone away from the water and stairs and contact local emergency services on 000 if there's any immediate danger”

unscripted proof the agent can reason about safety, not just recite a demo pitch→ newsletter pull-quote↗ Tweet quote

19:10

“the latest spot quote from GoldAPI is $4,314.50 per troy ounce, updated about 30 seconds ago”

concrete proof the new tool call pulls real, fresh data→ TikTok hook↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphor

hey sophie we are looking for a new accountant okay i'll check what next steps might be helpful here and what kind of help would you need from the team um really everything obviously the bookkeeping for our business any sort of tax advisory would be good as well we can help with your bookkeeping and both business and individual tax returns And an accountant can review your situation.

Could I get your name, please? That's OpenAI's new GPT Live1 model released this week in the API running on a real phone number. If you're not in the voice AI space, this might not mean much to you.

But I spent three years selling AI phone agents to real businesses with receptionists and customer support systems for over $12 ,000 each. And this right here is an AGI moment for voice AI. To compare this, here's the previous best voice AI setup that we could get.

Hey Alex, we just recently stopped working with our recent accountant and we are looking to get a new accountant. Yeah, we definitely help out with that. Would you be able to give me a quote at all for bookkeeping?

The main thing is that First Chat is a free console where they look at your books and give you a quote based on that. Would we be able to book an appointment to learn more? Yeah, you can set that up.

Can I get your name? So there we have it. Doesn't sound too bad, responds relatively quick, but it still has that robotic quality.

And that agent is three separate models glued together. Speech to text, which turns your voice into words, then an AI model that writes the reply, then text to speech, which reads it out aloud. Now, meanwhile, the big AI labs have been working on something called speech to speech.

ChatGPT voice mode, which you may have already used, is already speech to speech, which is why it sounds so human and so good. The model itself, is actually listening to your voice instead of reading a transcript.

So for example, when you laugh, cough or pause, it knows exactly what's happening and just acts and responds like a real person would. The problem forever though, is that speech to speech has never worked properly through the API and especially not on a real phone number like we just saw. The latency was always really bad, calls kept breaking and it was not something that you could ever put in front of a real customer.

But now that has finally changed with the GPT Live One API. Let's take another listen so you can directly compare it to that previous call hey Sophie what we're just looking for a new accountant great thanks Brendan what's the best number for the team to reach you on could I just learn more about the services you guys do do bookkeeping just before we lock anything in yeah of course we're a Melbourne accounting firm and we do offer bookkeeping as well as tax and BAS and payroll what sort of help would you like to start with?

Okay, so there we have it. Today, I'm going to go through step -by -step how to set up this brand new model yourself on a real phone number that you or anyone else can call and talk to. This is an incredible service to sell to other companies.

These systems can take on complex receptionist roles, answer calls 24 -7, never miss a single call and instantly push the data from every call into all of the systems that a company already uses. And the best part is that I'm going to use the new GPT -6 Astra model to build the entire voice system for us.

So you don't need any technical knowledge at all to follow along and get this fully built. Let's get started. Okay.

So getting started in building out one of these phone agents, the first thing that I want you to download is going to be Codex. So Codex is OpenAI's coding agent. And this is what we're going to be using to fully create everything to do with this voice agent.

So once again, to reiterate, you don't have to be technical at all to follow along with this. We're not going to be doing any coding or technical setups ourselves. We are very simply going to be prompting this AI to build the entire system for us.

It's going to be able to literally put the agent on the phone number and then all we need to do is call it and we'll have the entire thing set up for us so just make sure that you've got codex installed on your computer it is a desktop application and as for the pricing of codex you can get started for completely free on their free plan which will give you limited codex access but just be aware that you might hit the usage limits if you start talking to the agent too much and then you might be forced to upgrade to their go or plus accounts and also keep in mind that these are australian prices and not us dollars so just keep that in mind next up we now need to create an account with LiveKit.

So LiveKit is a voice agent development platform and it's what we're going to be using to connect this new OpenAI GPT Live 1 model to and it's also going to help us to connect our phone number into the system as well. So we are going to be using a separate platform for the phone number but we do need a platform in the middle to connect our model and our phone number together and that's going to be happening on LiveKit.

And this is also the exact same stack that OpenAI themselves are using. You can see that they are a customer of LiveKit and the speech -to -speech AI that is running on that Chachapiti app, which you may be familiar with, is a system running on LiveKit.

So we are getting the best system possible right here. And one really cool thing about LiveKit is that we can actually self -host it as well. So because LiveKit is fully open source, we can self -host these voice agents and run them locally on our computer.

But of course, the downside of that is if you turn your computer off, then nobody's gonna be able to call your agent. So LiveKit does have a LiveKit cloud option as well. So we can very easily use this.

And this is gonna be my recommended way for doing this so that we can turn off our computers. and these voice agents will still run continuously. And this is the setup that I would recommend if you are going to be building these out for other companies to have these systems run in more of a secure cloud that you can be confident that it's going to be running reliably as much as possible.

Now, getting started with LiveKit, there are a few different plans. You can get started with the completely free plan, and this allows you to build one agent and connect a phone number as well. So you can get started in building out and following along with this entire video just by running on this free plan.

But of course, if you do scale this up and start to build other agents, for other companies. But of course, they do have these other plans.

So if you do have multiple agents and multiple needs, you will be able to upgrade and handle all of that all on LiveKit's cloud. And once again, to be clear, these plans are for LiveKit cloud. Of course, if you are going to be self -hosting it, then you will not need to pay a monthly subscription for the cloud service because you're not going to be using it.

And just to make things easier, I'll have Codex and LiveKit linked down below in the description and pinned at the top of the comments down below. So you do have easy access to these platforms. But once you are on this page, the only that you need to do is essentially just create an account.

So you can click on this start building button and then you'll just go through a typical signup process to create an account. And then you'll land on a dashboard like this. Now don't get too overwhelmed.

We're not actually going to be doing too much work in this interface itself because we're going to be getting Codex to essentially manage and do all of the work here for us. But we'll come back to this in just a minute. The last thing that we do need is the actual phone number itself.

And one of my preferred providers for this is Telnyx. They essentially just have a big database of a bunch of phone numbers in which we can purchase and once we've purchased the phone number through Telnyx we're going to be able to get Codex and finally we should be able to call this Telnyx phone number and it will connect directly to our LiveKit built voice agent.

I also just wanted to quickly add that if you're just looking to get a US phone number you can actually purchase that directly through LiveKit itself. If you want to get any international numbers outside of the US then I would recommend using Telnyx. In my case I bought an Australian phone number and so that's why I'm using Telnyx in this use case.

But if you're just looking to get a US phone number, I would say it's probably going to be the easiest just to purchase it directly through LiveKit and you can skip the entire Telnyx setup. So to get started on Telnyx, you will just need to hit the login button, sign up for an account, and we'll come back in just a minute on actually getting this all connected to Codex so it does everything for us.

Okay, so we are now in the Codex desktop app. Make sure that you have signed into your account and make sure you are in the Codex option. If you look at the top right, make sure you are in Codex and not the ChatGPT mode.

Once we are in this Codex mode, we are now going to create a folder in which we're going to be doing all of this work in. Now, it should prompt you on this main part of the screen to create a folder. And if it doesn't, you should be able to click into...

this text and select a new folder or create a new project in which we'll automatically create a new folder. In my case, I've named the folder live kit demo. And now as simple as it sounds, I am literally just going to talk to Codex with my voice and tell it exactly what I'm looking for.

I'm going to tell it that I want a voice agent. I'm going to give it the use case for the voice agent to be a receptionist, maybe for an accounting company or something else. And I'm going to tell it that I want the stack to be an agent that's built on live kit.

I want that agent to be using the GPT live one API. And I want to get a phone number from plugged into this as well.

And literally by me speaking to the computer and telling it to build that, it's going to figure out how to build everything for us. And I can sit back, relax, and just watch it work. Hey Codex, I would like you to build a voice agent.

I would like to build a receptionist system for a plumbing company. If you could build this system out using LiveKit, using the new GPT Live One API to power this, that would be great. I would also like you to connect into Telnyx.

I've got a phone number on Telnyx. If you can connect into that account as well to build the entire setup. And so, literally from that incredibly basic prompt, laying out step -by -step what I wanted to build, Codex is going to be able to go on the internet, find LiveKit.

It's going to be able to look through their documentation to look at all of the LiveKit APIs. And it's going to be able to look at that code from LiveKit and understand exactly what it needs to do to literally build this voice agent on my LiveKit account. One thing that I will quickly add to my prompt is that I want it to be using my LiveKit cloud account.

We don't want it to accidentally build it and run it locally on my computer. So I will just tell it that I want it to build the agent into my LiveKit cloud account. And just to make sure, can you put this agent in my LiveKit cloud account?

I don't want this to be running locally. I want you to connect into my LiveKit cloud account and build the agent there. Just another quick note in your prompt, I'd actually recommend specifically typing LiveKit CLI, just telling it in the prompt to install the LiveKit CLI specifically.

In my prompt, I just said LiveKit, but if you want it to run as smoothly as possible, just type LiveKit CLI. That will then pick up on exactly what you're looking to install. In my case, mine was already sort of configured.

So it was able to figure that out. But to ensure that it runs as smooth as possible, just type live kit CLI. It's going to be able to install and understand exactly what we need to do just by doing that.

Okay, I have now sent that message off to Codex. We're going to be able to sit back, relax, watch it work. And it is just going to tell us step -by -step everything that we need to do.

And if we ever run into any issues, if you are following along with this, you can just tell Codex what the issue that you're having is. You can just ask it how to find something that... told you to find.

We can see now that Codex is asking us about the plumbing company that we're building the voice agent for. You can just make something up. This is just for a demo.

Okay, now just quickly, we can see that the agent actually found a LifeKit project from one of my other folders that it was able to connect into. So in your case, obviously, if you're setting up brand new, this is not going to happen for you. But in your case, what should happen is that it should tell you that it requires your account to be authenticated.

And it is more than likely going to ask you to provide some kind of an API key from LifeKit. And we can find these keys by going into the LiveKit web dashboard, heading into the settings section, which is just if we click out of here.

at the bottom in settings and then we click on api keys this is where we can create a key in the top right and we can create this key in the top right with the create key button we can name this whatever we want and once it's created i can click into this one that i just created we will then get back quite a few details and it's more than likely that codex will ask us for some of these details and you just need to follow along take the api keys and the secret and all the variables that it might be asking for and provide it directly to the agent now i don't recommend directly copying and pasting these api keys into the chat itself and Codex will more than likely not recommend you to do that either but it will give you a step -by -step on how to do it properly.

Essentially we just have to paste these keys into a secrets file that Codex has created for us this way that the keys don't actually go through any AI model and we're just putting the keys in where they should go. And once we've done this you should get that authentication into your account and you should see it come up with your account name as well just to confirm that it is properly connected.

Now another thing to note whilst this is building you can see here it's your OpenAI key has access to GPT Live 1. Now Codex might already have access to your API key if you've used it in any previous projects.

But if it doesn't have access to that API key yet, or it just seems like it hasn't found that API key and is at any point asking you for that OpenAI API key, you will of course just need to head to OpenAI's API console. So you will of course need an account with OpenAI. The same one that you're using for Codex is fine.

So on the openai .com page, you can just go to login and then click into the API platform. And then you should see at the bottom left of this dashboard, there's an API keys section. Inside of here, you can then very easily create the API key, copy it, and then provide it to the project where needed.

Okay, so our Codex agent is almost complete at building out our entire voice agent. And once it's done, I'm going to give it a live call so we can listen to exactly how good it sounds. But before we do that, if you could just quickly check that you are subscribed by scrolling down below and clicking on the subscribe button.

I recently discovered that 90 % of you watching right now are not subscribed to the channel at all, which is pretty... insane and i would really appreciate it if you were subscribed and make sure to keep watching as we're now going to finish up building this voice agent then i'm going to give it the live call and really stress test it so we can see exactly how well it sounds from the very little work that we've put into actually building it and then after this i'm also going to get codex to build an actual tool call into our voice agent so it isn't just a conversation back and forth but we can get our voice agent to actually do something for us so that might be in the form of booking appointments or sending information to some kind of a crm system okay so Codex is almost complete at building up the voice agent.

The last thing that we do need to do is give Codex access to our Telnyx account. So once again, Telnyx is just used for our phone numbers. You will just need to click the login button at the top right and create an account if you don't have one already.

Once you've created that account, you should land on a dashboard that looks similar to this. All that we need to do is go to left -hand side and click on the voice suite. Inside of this section, underneath the numbers tab on the left -hand side, we can then click into phone numbers.

Now this is where all of our existing phone numbers are going to be stored you can see one of my australian phone numbers in which is on my dashboard and so all you need to do is click on buy number in the top right corner and then you're going to be able to buy any phone number that you want through telnyx and it will just store itself on our account and then we're just going to connect codex to our telnyx account and it's going to set up everything for us so if you are looking for a us phone number you can just click on search numbers and just make sure that the us is selected and you'll see that there are hundreds of phone numbers that we can get especially the us ones there is a lot of them and they're about a dollar a month to use once you do have your phone number there isn't anything else you need to do just make sure to follow through with any instructions that it gives you to set up that number but otherwise just go to the top right corner in this sort of profile icon you will of course need to add some balance into your account it does cost money to make these calls of course and so you could add in five ten fifteen dollars just to make sure that you can do some demo calls but for reference it will cost a couple of cents for every single call that you make so this will last quite a while

And also in this section, just make sure to click into API keys. And then we just need to create a brand new API key. And once again, Codex should be giving us these step -by -step instructions on where to place this API key in the files it's given us.

And once we've done that, that is literally all that we need to do. Codex will now have full access to our Telnyx account. And it's going to be able to get that phone number and fully connected to LiveKit without us having to do any of the technical work ourselves.

Okay, so this is now done. We can see built and deployed Harbor. Plumbing's AI receptionist to your LiveKit cloud account powered by GPT Live 1.

So this is all ready to go. And it's even telling us that this is the phone number in which is the one that we just saw on my Telenix account. And it is supposedly all set up and ready to go.

Hi there, this is Mia, Harbor Plumbing's AI demo receptionist. This call may be recorded. How can I help with your plumbing issue today?

Hey Mia, we've got actually a leaking shower tap, leaking shower head. I was wondering if that's something that you guys do any services for and if you could provide us with a rough price estimate for something like that, that would be good as well.

Yep, a leaking shower head is something we help with. We cover leaking taps, shower heads and pipe leaks. Pricing can vary depending on the job and what's found on site though.

So I can't promise a quote. What's happening with the leak? Like, is it, say, dripping or more of a constant flow?

Yeah, it's more of a constant flow. There is sort of a flood that's building up at the bottom, so it is a bit concerning.

I see. Thanks for telling me. Is there any active flooding spreading beyond the shower area or any sewage coming up?

Yeah, it's sort of flowing out of the bathroom now and almost down the stairs, so we do need to get this checked on pretty quickly.

Thanks for letting me know. I'm just checking on the next step. Please move everyone away from the water and stairs and contact local emergency services on 000 if there's any immediate danger.

Don't enter unsafe areas. Once you're safe, I can take down the details for this demo inquiry. Is anyone in immediate danger right now?

No, everyone's safe. We just wanted to get somebody to come out and help us. Okay, so there we have it.

Obviously, it now works and it's swapped over to that plumbing type receptionist system. Of course, there will always be refinements when it comes to the actual conversational aspect in terms of what the agent is or isn't capable of doing. So you may have heard that it was telling me that it was a demo agent and it wasn't able to do certain things.

So CurrentX has obviously prompted it to be a demo agent. What I could do is just tell it to remove the demo agent part of it, obviously. But hopefully you get the idea.

In just very little work, we've been able to build that full -on voice system. And if we spend some actual time into this, giving it some proper requirements for a legitimate business's... series of call reasons, we're going to be able to build this system pretty good to be able to handle a wide range of calls.

But of course, these agents are useless if we are just talking back and forth with them and we're not doing any actual work. And so this is where tool calls come into place. So a tool call might be booking an appointment or sending some information into a CRM system.

This could also be pulling information out of a CRM system into the voice agent as well. And so I'll quickly showcase exactly how we can set up tool calls very, very easily through this as well. It is as simple as me telling Codex to set up a function call within this voice agent.

We don't have to use any secondary or third party platforms to build any automation workflows anymore. We just need to tell Codex to literally build out the entire workflow. It's going to be able to connect into any external apps that have APIs that are readily available and build everything out for us.

Hey Codex, I would now like you to build a function call into this voice agent in which is able to look up the live price of gold. So if you could, if somebody is asking for the live price of gold, This should trigger the function call to hit a public API endpoint for the live price of gold, pull that price back into the conversation and read it out to the user so that they are getting the real up -to -date price and it's not some outdated price.

And now I'm going to send that off and that's all I need to do. Okay, Codex is now telling us that this is now complete. The tool call has been added to our voice agent.

I'm going to give it a call and see if this works. Hi there, you're through to Mia, Harbor Plumbing's AI demo receptionist. This call may be recorded.

How can I help today? Hey, Bia, I was just wondering if you'd be able to give me the live price of gold at the moment?

Sure thing. I'll check the latest gold price for you.

Okay, the latest spot quote from GoldAPI is $4 ,314 .50 per troy ounce, updated about 30 seconds ago. Is there anything else I can help with today? Perfect, that's it.

Okay, and there we go. You can see that it was able to obviously do that function call to get the live price of gold. It actually said the name of the API on there.

Obviously, you'd fine tune that. This is not really a use case for a plumbing business. But just to prove the point, we can do tool calls on this speech -to -speech model as well.

I hope you've enjoyed the video. Let me know down below in the comments what you think of this new voice model. And if you want to push GPT -6 Astra even further, I did a full -on video building out websites with this new model.

And honestly, the results have blown my mind as to how good it is at design now. So I'd highly recommend checking this video

The Hook

The bait, then the rug-pull.

The video opens mid-call: a synthetic voice is already negotiating an accounting quote before anyone explains what's happening. That's the pitch. OpenAI's new GPT Live-1 sounds close enough to human that the demo needs no introduction before it makes its own case.

Frameworks

Named ideas worth stealing.

01:21list

The old voice-agent stack (3 stitched models)

Speech to text
AI text model
Text to speech

The legacy pattern for voice agents before speech-to-speech: transcribe the caller's audio, generate a text reply, then synthesize speech from that text. Each hop adds latency and strips out tone, which is why these agents sound robotic.

Steal forexplaining why a demo voice agent sounds stilted, or scoping a build decision between speech-to-speech and a stitched pipeline

CTA Breakdown