Multilingual
One number, several languages, one assistant for each.
In a hotel, a private clinic or an exporting company, the calls do not all arrive in one language. And the usual answer — one person who speaks English and is never at their desk when somebody rings — is not a solution, it is a lucky coincidence.
You run one assistant per language on the same public number, and the call reaches the right one after the first few seconds of conversation. The language is pinned explicitly on each assistant, because transcription does not guess language well, least of all across accents.
Where it matters in practice
In Romania the real cases are not exotic: English for tourists and export clients, Hungarian in Transylvania, German and Italian for companies with partners there.
Hotel reception
Accommodation enquiries and guest questions, in the language the person rang in.
Clinics with foreign patients
Administrative bookings, where a misunderstanding costs a wasted visit.
Companies with export clients
Support and orders, where the delay while somebody who speaks the language is found is a direct loss.
Institutions in mixed areas
Where part of the public prefers another language and is entitled to use it.
Shifts with nobody available
At night, when the person who speaks that language is not working.
How the call reaches its language
The mechanism that works is not one assistant speaking every language, but several assistants each with its language pinned, and routing between them.
- 01 The first assistant greets briefly, in two languages One sentence in Romanian and one in English. That is enough for the caller to answer in their own.
- 02 Route on what it hears The call passes to the assistant with the right language, as a handoff between assistants inside the same call, with nothing ringing again.
- 03 The context goes across What was said up to that point is carried in the configured form, so the caller does not start over.
- 04 Each assistant has its language pinned The transcriber, the model and the voice are chosen per language rather than switched mid-call.
- 05 The human transfer respects the language The route leads to the queue or person who speaks it, otherwise the whole routing achieved nothing.
What the record shows
After a month you know exactly which languages people call you in, which is usually different from what everybody in the company believed.
| The call language | On every call, through the assistant that handled it, so it can be counted. |
|---|---|
| The handoff between assistants | The moment of routing and the context carried, as an event on the call. |
| The transcript | In the language of the conversation, untranslated, because machine-translating a transcript adds a second layer of errors. |
| The result | Structured output with the same fields whatever the language, so it stays comparable. |
What a multilingual reception does not do
It is easy here to promise an assistant speaking thirty languages. What can be built honestly is less, and more reliable.
- It does not switch language mid-sentence Language is a property of the assistant. A caller mixing two languages in one sentence will be understood badly, whichever vendor is underneath.
- It does not reliably detect language by itself Automatic detection fails on exactly the accents that matter, which is why the recommended pattern is a short bilingual greeting and routing on the reply.
- It does not translate It is not an interpreter. It does not translate between caller and operator, and it does not translate the transcript.
- Not every language sounds equally good Quality depends on the transcription and voice vendor for that language. Exactly one speech synthesis vendor is connected end to end in this build.
- It does not solve the human half If the transfer lands on an operator who does not speak the language, the whole chain breaks at the last step.
What businesses ask us
Why not one assistant that speaks every language?
Because language is pinned on the assistant at the transcriber, model and voice level, and an assistant with no pinned language transcribes worse in every language than one specialised in a single one. Several assistants with routing between them give a better result and are far easier to debug.
How many languages can be added?
As many assistants as you are willing to maintain. The practical limit is not technical but editorial: each language means a prompt written by somebody who genuinely speaks it, plus translated documents. Two well-made assistants beat five rushed ones.
Does Romanian work as well as English?
Romanian works, with the language pinned explicitly on the assistant. The hardest part remains proper names and digits, which is why the assistants in our examples always read names and numbers back before writing anything down.
What about a caller speaking a language you do not cover?
The assistant says, in English, that it cannot help in that language and transfers to a person or takes the contact details. That is better than a conversation in which both sides assume they understood each other.
Where to go next
Assistants
Transcriber, model, voice, tools, knowledge, and a way to prove it still works.
First assistant
Transcriber, model, voice, prompt, publish. And how to tell it worked.
Integrations
25 business connectors, 18 selectable providers, eight storage destinations.
Pricing
Components at their published price, the platform on top, and what we cannot price yet.
Notes
-
1
The tools named on this page are the ones in
App\Services\Integrations\ConnectorRegistrycarryingimplemented. Anything that is only in the catalogue, without that flag, is not named here as though it worked. - 2 Language is pinned on the assistant, together with the transcriber, the model and the voice. Exactly one speech synthesis vendor is connected end to end in this build, and the integrations page shows the split.