Skip to main content
An agent no longer has a single language. It has a primary language - the one every call starts in - and optionally a set of additional languages it is allowed to move into while the call is running. You set both in the Language pill, the first of the four pills at the top of the agent editor.
Language is the first choice for a reason. The Transcriber, Intelligence, and Voice pills are all filtered by it - a speech model that cannot hear Kannada is greyed out the moment Kannada is in the set. Choose the languages your callers actually speak, then let the rest of the row narrow itself.

​
Primary language

Primary language is the language the call starts in. It is also the language the greeting is spoken in, because the greeting happens before the caller has said anything - there is nothing to detect yet, so it can only be configured. Every agent has exactly one primary language.
There is no separate greeting language setting any more. The greeting always follows the primary language.

​
Changing it can change your voice

Some voice providers cannot be handed some languages at all. Cartesia raises an error on Kannada, Odia, Assamese, and Urdu; Sarvam AI speaks eleven languages, which exclude Urdu and Assamese; and Vocily AI’s catalogue is India-only. That is not a “the voice sounds wrong” problem - the call fails to connect, on every voice in that provider’s catalogue. So when you change the primary language to something your current voice or provider cannot speak, Vocily swaps it and tells you it did, right there in the pill:
Voice provider changed to Sarvam AI - the previous one cannot speak Kannada at all, so calls would not have connected.
The swap happens while you are looking at it, never silently mid-call. If you would rather choose the replacement yourself, open the Voice pill and pick one.

​
Also speaks

Add languages under Also speaks to let the agent switch into them during a call. Each one appears as a chip you can remove. The primary language can never also be an additional one - it is already the language the call starts in. An agent with no additional languages speaks one language only. It will not follow a caller who switches; it continues in its primary language and asks them to switch back.
More languages costs some accuracy on the primary one. With a single language the recognizer is tuned for exactly that language. The moment there are two, it has to listen for every selected language at once and cannot be re-tuned mid-call - narrowing it later would mean dropping the audio connection in the middle of a sentence.Add the languages your callers really use, not every language they might.

​
Switch when

Once there is more than one language, the Switch when setting (under the pill’s gear) decides what is allowed to move the conversation: The setting is disabled until you add a second language, because with one language there is nothing to switch to.
“Only when the caller asks” is a promise about stability. A bilingual caller who drops one English word into a Hindi sentence is not asking for anything. On this setting the agent keeps going in the language it was in, which is what you want for a scripted or compliance-sensitive call.

​
How automatic switching behaves

Automatic switching is not “the agent heard Tamil, so it speaks Tamil”. A recognizer’s per-turn guess is noisy, and an agent that flips on every uncertain turn is worse than one that never switches at all. Vocily applies four rules:

A confident turn switches immediately

When the recognizer is sure about the language, the agent moves on that turn.

An unsure turn has to repeat

A weak signal must show up on consecutive turns before it counts. One uncertain turn never moves the voice.

Numbers never count

“4 2 8 1 9 7” - an OTP or an order ID read out digit by digit - looks like English to any recognizer, whatever language the call is in. Digits are ignored as evidence entirely.

Unlisted languages are held

An agent switches only into languages you configured. A Telugu caller on a Hindi + English agent does not move the voice.
Very short utterances (“haan”, “ok”) are also ignored - there is not enough in them to tell one language from another.

​
The first thing the caller says is special

On an inbound call the caller usually speaks first, and their opening words are the strongest evidence that will ever exist about which language they want. There is also nothing to flap away from yet, because the conversation has not committed to a language. So the first caller turn can switch on its own, without waiting for a repeat. Later turns need the full anti-flap treatment.

​
What the agent replies in

The languages you select also bound what the agent is allowed to answer in. Vocily writes that constraint into the agent’s instructions automatically, generated from the same setting the detector reads, so the two can never disagree. You do not write it and you cannot edit it. It changes on its own whenever the language set or the switch trigger changes.
Do not write language rules into Response Style. Instructions like “reply in Hindi if they speak Hindi” compete with the platform’s own switching, and which one wins is not predictable. Response Style is for tone, length, and how the agent handles uncertainty - see Agent Core.Keeping names, emails, URLs, and product codes untranslated is already handled for you.

​
If a provider fails mid-call

If the speech or voice provider stops responding and the agent falls back to its backup, the conversation returns to the primary language and the agent says so:
Sorry - I’m having a technical problem. Let me continue in Hindi.
A backup only has to cover the primary language, not the whole set. That is deliberate: nothing else on the roster covers every Indic language, so requiring full parity would leave Indic agents with no backup at all. Announcing the reset is what makes it honest - a caller who had switched to Tamil is told the agent is going back to Hindi, instead of quietly finding it no longer understands them. See Transcriber for choosing the backup.

​
Supported languages

Not every speech model covers every language, and no model covers all of them. The language list here is the full set the platform offers; the Transcriber pill shows which models can serve the particular combination you picked, and greys out the ones that cannot.
Urdu and Assamese can be heard but not spoken. The speech models transcribe them, and no voice provider on the platform can voice them. Keep them under Also speaks so the agent understands a caller who uses them; an agent with either as its primary language will not be able to speak at all.
A new agent starts on Hindi with English alongside it, with switching set to “the caller asks, or we detect it”.

​
If you used Vocily before

The old Conversation Language dropdown, with its seventeen fixed modes, is gone. So are Greeting Language and Auto Language Rules. Existing agents were migrated to the equivalent language set. The thing worth revisiting is Also speaks: an agent that was on Indian Multilingual now needs its languages named explicitly, because “any Indian language” is not a set a recognizer can be held to.

​
Checklist

Before you rely on a multilingual agent, place a test call and check that:
  • The greeting comes out in the primary language.
  • Speaking an additional language moves the agent within a turn or two.
  • Reading out a number does not move it.
  • Speaking a language you did not configure leaves the agent where it is.
  • On “Only when the caller asks”, using another language does nothing until you ask.

​
What’s next

  • Transcriber - which speech model hears these languages, and what it costs.
  • Voice - choosing a voice that can actually speak them.
  • Agent Core - the greeting and the prompt sections.