Prisma 2: The model for voice assistants that understands Catalan

What is Prisma 2?

It's a voice agent that picks up the phone for you. It handles the call, understands what the person on the other end is saying, responds naturally, and only passes the call to a human when it needs to. Built for local businesses: bilingual, with real customers, not for a demo at a conference.

Up to here it could sound like anything else you see these days. I'd bet you've already seen three similar ads this week. Let me tell you what's underneath.

The Catalan and Spanish mess

This is the heart of Prisma 2. And the reason models built for large markets crash the moment you bring them here.

A Catalan customer doesn't speak «in Catalan» or «in Spanish». They jump from one to the other within the same sentence, without warning and without noticing. It's a phenomenon with a name, code-switching, and it's well studied in linguistics. The problem is that almost no commercial voice system is ready for it. And speech-to-text systems, the first piece of any voice agent, have a nasty trap: a system tuned for Catalan transcribes perfect Catalan even when you're speaking Spanish.

It doesn't get the word wrong. It gets the language wrong. And with total confidence, too 😅

It's the worst kind of error there is, because it looks correct. If the transcript says something coherent, everything downstream (the response, the tone, the actions) assumes it's fine. And suddenly your agent replies in Catalan to someone speaking Spanish, and nobody noticed where the chain broke.

What we learned, after listening to many real calls (which is the only way to learn these things, by the way), is that no single system fixes this on its own. You have to cross several signals in real time and arbitrate between them to decide, at each moment, which language the person is actually speaking.

I'm not going to tell you how many signals we use or how we combine them; that's months of work and it's what makes Prisma 2 Prisma 2. But I will tell you the principle: in a bilingual market, language isn't data you receive. It's a problem you solve.

Ghost audio

When you speak, the agent has to shut up. Immediately. Sounds obvious, right? It's not obvious at all.

Here's why. When you tell the system to stop talking, the audio already on its way to the phone keeps playing for a few more seconds. We had what we internally called ghost audio: the «interrupted» agent kept talking because there were chunks in flight that the stop command couldn't catch.

The feeling on the other end is awful. You're trying to say something, the agent keeps going, and you hang up thinking it's a badly made robot. Which is exactly what it is, if you don't fix it.

We fixed it by changing how we send audio to the phone: in very small chunks instead of large blocks. So when the stop command arrives, what's still in flight is tiny and cuts off almost instantly instead of going on for seconds.

It's one of those details nobody notices in a demo. On a real call, it makes the difference between something that feels human and something that feels like a broken answering machine.

Why we're telling you like this

I know it's odd to publish a product article telling you about the bugs we've had. The usual thing would be a polished video, a big metric in the centre, and a «request a demo» button.

But the difference between a system that works and one that only looks like it does is right in these details nobody shows. And if you're going to trust us with your business calls, we think you deserve to know.