Introducing NLPearl Speech-to-Speech 2.5: Natural, Real-Time Conversational Voice AI

Our new voice model, trained for natural, real-time conversation, inside a complete agent you can plug into any product.

Voice AI has learned to talk. Now it can hold a real conversation: help a shopper choose, play a character, answer a support call, and sound like a person doing it.

Today we're launching NLPearl Speech-to-Speech 2.5, our new voice model, trained and optimized specifically for natural, real-time conversation. It sounds more human than ever, answers faster, takes turns more like a person and handles harder requests. It powers a complete voice AI agent, not just a voice: it knows your business, and it gets things done. With the Realtime API, you can plug it into anything you build: your mobile app, your e-commerce site, your game, your virtual assistant, your support line, your devices. It's available today to every NLPearl account.

What's New in Speech-to-Speech 2.5

Speech-to-Speech 2.5 comes out of production. Every day our agents take real calls and answer hard customer questions, and those calls showed us what to improve. Four things changed.

It sounds more human. The voice has more life in it: the pitch moves, the pace changes, and it sounds less like reading.

It answers faster. The pause before it speaks is shorter than in v2, so the conversation keeps its pace.

It takes turns like a person. It is better at telling when you have finished and when you are only pausing. Talk over it, and it stops and listens.

It handles harder requests. It follows longer, messier questions, the kind real customers ask, and it stays on the task your business gave it.

None of this is a lab result. Speech-to-Speech 2.5 is already answering real calls in production.

A Speech-to-Speech Model Trained for Natural, Real-Time Conversation

Speech-to-Speech 2.5 is our own voice model, the successor to Speech-to-Speech v2. Most voice models are built to read text aloud. Ours was trained and optimized for one thing: holding a natural conversation in real time.

It was pre-trained on speech, fine-tuned on real conversations, then aligned to behave like a strong human agent. It goes audio to audio, with no transcript in between, so it hears how something is said, not just the words. Here's what your customers will notice.

It listens while it talks. Talk over it, and it stops mid-sentence, listens, and answers what you just said. It was tuned on the crosstalk and interruptions of real-world calls.

It hears how you feel. It picks up frustration and softens. It hears hesitation and slows down.

It speaks your customers' language. More than 20 languages, each with voices and accents to choose from, so it sounds local wherever your customers are.

It sounds clear. In HD, with 24 kHz audio: three times what a phone line carries.

It keeps up. It answers at the pace of a real conversation, without the pause that gives a machine away. It starts speaking 550 ms after you stop.

Don't take our word for it. Try Pearl now: start a conversation below and interrupt it.

A Voice AI Agent That Knows Your Business

A human voice is where most voice APIs stop: you get the voice, and your engineers build the agent. With NLPearl, you get the whole agent.

It has the answers. Give it what it needs to know, from FAQs to policies, and it uses them in every conversation.

It remembers your customers. With memory on, it picks up where the last conversation ended. With contact groups, that memory follows them across the phone, chat and your app.

It gets things done. It looks the customer up before it says hello, checks orders and books appointments during the conversation, and sends the follow-up after.

It knows when to hand off. When a person should take over, it passes on everything it collected, and your app or contact center decides what happens next.

You build it in PearlVibe, our AI agent engineer: describe what it should do, and PearlVibe builds it for you.

Conversational AI Use Cases: Support, Sales, Booking and Beyond

Speech-to-Speech 2.5 already does real work on real calls. These are the jobs companies give it most.

Customer support

It answers inbound calls, replaces the phone menu, and covers after hours and peak times. When a caller needs a person, it transfers the call with everything it collected.

Sales and lead qualification

It calls new leads back, asks your qualifying questions, and passes the ready ones to your sales team with the answers attached.

Booking and scheduling

It books, moves and cancels appointments during the call, checks availability as it speaks, and sends the confirmation after.

Front desk and assistant

It picks up every call, answers the common questions, takes messages and routes each caller to the right person.

Wherever your customers are

The same agent is not tied to one channel. With the Realtime API, you can put it anywhere people can talk.

• Mobile apps. A signed-in user taps Talk, and your agent already has their account in hand. Onboarding, account help and bookings happen by voice, and nobody spells their email twice.
• E-commerce websites. A shopper asks which size fits, whether it ships by Friday, or how two products compare, and gets an answer in a voice, not a contact form.
• Contact centers. The agent plugs into the phone systems your contact center already runs, with 8 kHz μ-law audio for platforms such as Twilio, and works with your dialer on outbound calls.
• Virtual assistants. An assistant, tutor or coach inside your product that users simply talk to. It remembers them and picks up where the last conversation ended.
• Games. Characters, companions and game masters answer what the player actually says, at the pace of play.
• Kiosks and devices. A kiosk in your store, a screen in your lobby, a device you make: if it can record and play sound, your agent can talk through it.

Everything that makes a conversation work is handled for you: the listening, the voice, the timing. Your server starts the conversation, your product opens it, and your customer starts talking. Your API key never leaves your server.

Every conversation shows up in your call history with its recording, and its results flow to your other tools. Change your agent, publish once, and every channel runs the new version.

Realtime Voice API for Developers

Here is the whole integration, starting on your server:

Your app opens the WebSocket with the token:

That is the whole handshake.

A few things we decided so you would not have to:

• The key stays home. Your app gets a single-use token that opens one call, once, and expires in 60 seconds. It is never a key to your account.
• Your server steps aside. Audio flows straight between your app and NLPearl. Your backend is not in the audio path.
• Raw audio, no wrapping. Binary 20 ms frames both ways: 24 kHz PCM for HD on the web, or 8 kHz μ-law for phone systems such as Twilio. No base64, no JSON envelopes.
• Interruptions need no code. When someone talks over the agent, its next frames are already silence. No buffer to flush, no event to catch.
• Three events, that's all. ready, transfer and ended. Your app can send hangup.

Speech-to-Speech 2.5 FAQ

What is Speech-to-Speech 2.5?

Speech-to-Speech 2.5 is NLPearl's own voice model, trained and optimized for natural, real-time conversation. It was pre-trained on speech, fine-tuned on real conversations and aligned to behave like a strong human agent. It goes audio to audio, with no transcript in between.

What is new in Speech-to-Speech 2.5?

It sounds more human than v2, answers faster, takes turns more like a person and handles harder customer requests. The improvements come from real calls in production.

Can people interrupt the AI agent?

Yes. The agent listens while it talks: interrupt it, and it stops mid-sentence and answers what was just said. Developers write no code for this, because the agent's next audio frames are already silence.

What languages does it support?

More than 20 languages, each with voices and accents to choose from.

What is the NLPearl Realtime API?

The NLPearl Realtime API is a realtime voice API that connects a complete voice AI agent to your mobile app, website, game, devices or contact center. Your server creates the call, and your app streams the audio over a WebSocket.

What can I build with the Realtime API?

Any product where people talk: voice assistants in mobile apps, shopping assistants on e-commerce sites, characters in games, virtual assistants, tutors and coaches, customer support on the web or the phone, and interactive kiosks and devices.

How is it different from other realtime voice APIs?

Most realtime voice APIs give you a voice model and leave the agent to your engineers. NLPearl gives you the whole agent: it knows your business, remembers your customers, takes actions during the call and hands off to a person when it should.

Can I use the Realtime API in my call center?

Yes. It streams 8 kHz μ-law audio, the format of phone systems such as Twilio, so it connects to the telephony your contact center already runs. It answers inbound calls, works with your dialer on outbound calls, and transfers to a live agent with full context.

How much does the Realtime API cost?

Realtime API calls use the same credits as every other NLPearl call, HD audio included.

Does it work with my existing Pearl?

Yes. Every existing Pearl works with the Realtime API today, with nothing to rebuild.

Start Building With Speech-to-Speech 2.5 Today

Realtime API calls use the same credits as every other call, HD included.

• See it. Book a demo with our team.
• Build it. Create your agent in PearlVibe, then connect it with the Realtime API Quickstart or the Browser App guide. The Create Stream Token and Stream WebSocket references have every detail.
• Already have a Pearl? It works with the Realtime API today. Nothing to rebuild.

Your users are already in your app, on your site and on your phone lines. Now they can talk to you wherever they are.

Build it. Plug it in. Start talking.

Ofek Darhi