Serving customers in several languages
Connect answers in the language the conversation is being held in, grounded in the same Knowledge whichever language the question arrived in. On voice a regional speaking-style layer shapes delivery, and a human voice profile shapes character. What this documentation does not carry is a certified list of supported languages, and this page does not invent one.
How Connect decides which one to answer in#
There is no selector for this on a conversation. It is a property of what arrived: the words a caller speaks, the mail a customer wrote, the WhatsApp message they sent. Connect answers in kind.
Two layers sit on top of that on the phone. voice_region carries a regional speaking-style layer, ranked on evidence, which changes how a reply is delivered — pace, phrasing, the shape of a greeting — rather than what it says. voice_profiles carries the human voice profiles, which set character. Both are configuration a workspace chooses in Voice Lab, not something the caller triggers.
Grounding a question your Knowledge was not written for#
Answers are grounded in Knowledge whatever language the question arrives in, and an answer Knowledge cannot support is refused rather than improvised. That rule does not soften because the question was asked in another one — which is exactly what you want, and also what makes thin Knowledge look like a translation problem when it is not.
- Where a phrase has to be exact — a product name, a legal formula, a term of art your customers use — put that phrasing into Knowledge in the language it must appear in, rather than relying on the model to render it.
- Knowledge is budgeted per turn on the voice path: roughly 2,000 characters and four facts reach the model on a call. A Knowledge base that says the same thing four times in four languages spends that budget on repetition.
- Memory tiers work the same way in any language. A direction recorded against one contact outranks a channel-wide one, so 'always write to this account in Tamil' belongs at the contact tier.
What is not claimed here#
- A supported-language list
- The capability registry records the regional speaking-style layer and the human voice profiles; it does not record a list of languages, and neither does this corpus. Test the languages your customers actually use rather than trusting a list.
- Certified or legally valid translation
- Nothing in Connect is a translation service and no page here claims translation accuracy of any grade. A document whose wording carries legal weight is a document for a person and a translator.
- Script and transliteration behaviour
- Not documented in these sources. Whether a reply comes back in a native script or in Latin transliteration is a property of the model and the prompt, so it is something to check on a real conversation before promising it to anybody.
- Per-language delivery reporting
- On WhatsApp, templates, session windows, media handling and delivery reporting are provider-dependent. Connect exposes what the configured provider does and does not simulate the rest.
The India case, specifically#
A multilingual business in India runs into a rule that has nothing to do with language models. Commercial SMS requires DLT registration with the telecom operators — entity, header (sender ID) and template registration — and a template is registered per language variant. That is a process with the operators rather than a setting in Connect, and dlt.py holds the registration data once it exists.
It is also moot on the live account, because that carrier carries no outbound SMS at all. A multilingual business that expected to reach customers by text is looking at WhatsApp or email instead. The Phone screen states the position through readiness.messaging rather than offering a thread that cannot send.
Where the extra cost sits#
Language does not change the meter — audio tokens on voice, the daily allowance on mail — but it changes prompt size, and prompt size is measurable on a live call. Instructions of roughly 7,500 characters produced a first token in 1.2–1.8 s; at 9,600 characters that became 2.3–3.4 s. Writing one careful instruction rather than one per language is worth about a second of silence at the start of every reply.
The measured floor does not move for a second language either. The best measured call had a 3.3 s median reply, and a caller in any language hears that same pause.
Questions#
Which languages does Connect speak?
This documentation does not carry a supported-language list, and writing one from an assumption would be worse than saying so. What is recorded is the regional speaking-style layer and the human voice profiles on the voice path. The dependable answer for a specific language is a test call and a test thread on your own workspace.
Do I need to keep my Knowledge in every language?
Not as a rule, because answers are grounded rather than looked up verbatim. The exception matters though: anything that must be word-perfect — a product name, a regulated phrase, a formula your customers quote back at you — belongs in Knowledge in that language, because the alternative is trusting a rendering nobody reviewed.
Can Connect send a text message in a local language?
Not on the live carrier, which carries no outbound SMS. Where a provider that does carry SMS is configured, Indian commercial messaging additionally needs DLT entity, header and template registration with the operators, per template and per language variant. WhatsApp and email are the channels that work today.