Why almost no voice agent really speaks Catalan
When a voice platform says it supports "70 languages", that is usually both true and misleading. Language support in a real-time speaking agent is spread across three separate pieces, and one failing is enough to make the result sound wrong. In Catalan, that piece usually fails.
A voice agent is three things, not one
For a call to work, three layers have to line up: the one transcribing what the caller says, the one deciding what to reply, and the one turning that into speech. Each has its own list of supported languages, and the lists do not match.
What the marketing site publishes tends to be the union of the three, not the intersection. The intersection is what talks to you on the phone.
| Layer | What it does | Where Catalan breaks |
|---|---|---|
| Transcription | Turns the caller's speech into text | Usually fine; struggles with proper nouns and place names |
| Language model | Decides what to answer | Large models write good Catalan; the problem is not here |
| Real-time speech synthesis | Says the answer out loud | This is where it falls down: fast models often have no Catalan |
The fast model and the Catalan-speaking model are usually different models
Voice platforms have models optimised for latency, designed precisely for phone calls, and fuller but slower ones. Catalan often exists only in the second group. We checked this against our provider's API directly: of the real-time models available, only one accepts Catalan.
That has a practical consequence worth knowing before signing anything: if a supplier tells you they will build your Catalan agent on their fastest model, either they have something nobody else has, or they have not tried it.
The accent is not set by the language setting
This is the trap that is hardest to spot. You can set the agent's language to Catalan and have it still sound foreign. The reason is that the accent comes from the chosen voice, and voices are built from original audio: a voice recorded in American English will say Catalan words with the mouth of an American English speaker.
The result is unmistakable to any Catalan speaker and it is worse than neutral: it does not sound like a professional assistant, it sounds like an unfinished product. The test takes ten seconds and should always be run before buying: ask to hear the voice saying a sentence with street and town names from your area.
| Question | Good answer | Answer that should worry you |
|---|---|---|
| Which voice model do you use for real time? | The specific name, and that it supports Catalan | "We support 70 languages" |
| Where does the voice come from? | Generated or recorded in Catalan | "It is a multilingual voice" |
| Can I hear it saying local names? | An audio file, right now | A Spanish demo "which is the same thing" |
| What if the caller switches to Spanish? | It switches and stays there | "The model handles that" |
| What if it is a clone of a real person's voice? | Written consent from that person | Silence, or "no need" |
The caller who speaks two languages
In Catalonia it is common to start in Catalan and continue in Spanish, or to mix both in one conversation. An agent has to detect the switch, follow it and stay there, without mixing within a single sentence — which is exactly what makes it sound artificial.
It is a design decision that has to be written explicitly into the agent's instructions. It does not emerge on its own from ticking "multilingual".
What about cloned voices?
Cloning a real person's voice to answer your phone is technically easy and legally thick: an identifiable person's voice is personal data, and cloning it requires their explicit written consent, not a nod in a corridor.
We decided against it: we use generated voices that correspond to no real person. That is a deliberate choice, not a technical limitation, and it is explained in how we handle data.
Frequently asked questions
Can you tell it is an AI speaking Catalan?
You can tell it is an AI because it says so in its first sentence, as the rules have required since August 2026. What should not be noticeable is the accent: if it sounds foreign, that is a defect of the chosen voice, not a limitation of the language.
Can the same agent handle Catalan and Spanish?
Yes. One agent detects the caller's language and stays in it. What we do not recommend is separate numbers per language: it complicates forwarding and confuses the caller.
Does the fastest model give worse Catalan?
The problem is not that it is worse: often it simply is not there. That is why the choice of model is not a technical detail but a product decision, and you should ask for it by name.
What about town and street names?
That is where transcription fails most, which is why the agent must repeat names and phone numbers out loud and ask for confirmation. Good capture compensates for imperfect transcription.
Can I use the voice of someone on my team?
Technically yes, but you need their explicit written consent, and you have to plan for what happens to that voice if the person leaves. That is why we use generated voices.
Want to hear it in Catalan before deciding anything?
We will play you the agent saying your business name and your street. Ten seconds, and you judge.
Free demo