Multilingual Voice Agents: One Phone Line

Why voice is harder than chat

In a text channel the first message arrives already written, so language detection has something to work with before anyone has been kept waiting. On a call the agent has to speak first, in some language, before it knows which one the caller uses.

That single asymmetry produces every difficulty in this article. Greet in the wrong language and you have started badly. Ask which language they prefer and you have rebuilt the phone menu you were trying to remove. Wait to find out and you have left silence on a live line, which callers interpret as a broken connection within about two seconds.

There are three workable answers — a separate number per language, a short neutral greeting that invites the caller to speak, and detection from the caller's first utterance — and most deployments need a combination rather than one of them.

Choosing the language without a menu

1. 1A different number per language, where you canThe cleanest solution by a wide margin: the caller's choice is made before the call connects, the greeting is correct immediately, and no detection is needed. If several DID numbers can sit behind the same system, this should be the first option considered.
2. 2A short neutral greetingWhere one number serves everyone, open with a brief greeting containing the company name and little else, then let the caller speak. The company name is usually language-neutral, and one sentence from the caller is enough to detect from.
3. 3Detect from the first utterance, then commitDetection on a short spoken phrase is less reliable than on written text, so treat the first detection as provisional and allow it to be corrected by the second utterance rather than locking it for the call.
4. 4Offer a switch in plain words, once'I can continue in Russian if that is easier' offered once, in the language you think is right, is better than a menu and better than guessing silently. Offering it repeatedly is worse than both.
5. 5Decide the default deliberatelyWhen detection is uncertain, the correct behaviour is to continue in the number's default language rather than to switch back and forth, which is the most disorienting possible experience.

Pronunciation is a product decision

This is where multilingual voice projects lose credibility, and it is almost never discussed before launch. A correct answer pronounced wrongly sounds like a machine reading a foreign word, and callers judge competence on it.

• Proper nouns are the main problem: your company name, product names, street names, branch names and staff names. These must be checked in every language rather than assumed.
• Numbers, dates and currency are read differently across languages, and a phone number read in the wrong grouping is unusable even when every digit is correct.
• Loanwords and technical terms are often kept in English inside another language, and forcing a translation makes them harder to understand rather than easier.
• Mixed sentences — a local sentence containing an English product name — are routine in multilingual markets and are where pronunciation most often breaks.
• Voice choice matters per language. A voice that sounds natural in one may sound noticeably foreign in another, and it is worth listening rather than assuming.

The practical test is cheap: write down the twenty proper nouns your agent will have to say, have it say each one in every language, and listen. Most teams find two or three that need to be respelled phonetically in the material, and that is the whole fix.

Localised knowledge, not translated knowledge

As with text, some material is genuinely the same in every language and some is not — but on a call the consequence is sharper, because a caller cannot scan a page to find the part that applies to them.

• Shared: how a product works, technical specifications, general policy.
• Different by language and market: prices in local currency, delivery times, which branches exist, opening hours, applicable regulations and consumer rights wording.
• Different in length: an answer that is comfortable to read in one language may be too long to listen to in another. Spoken answers should be shorter than written ones in every language.
• Different in the questions themselves: the distribution of what callers ask shifts by language, and it shows in the transcripts within a month.

Handover in the right language

This is the operational question that decides whether multilingual voice support is real or theatrical, and it is the one most often left until after launch.

• Know which languages your operators cover, and in which hours. An agent that handles a call beautifully in one language and transfers it to a queue where nobody speaks it has produced a worse outcome than a polite refusal.
• Route by language before routing by topic wherever coverage differs, or calls arrive where they cannot be handled.
• Record the detected language on the customer record, so a callback is made by someone who can hold the conversation.
• State honestly when a language is only supported by automation. Callers accept that; they do not accept discovering it after being transferred.
• Write the out-of-hours message in every language. An untranslated closing message is the most visible failure a multilingual line can produce.

What to measure per language

1. 1Everything, split by languageVolume, containment, transfer rate, call duration. Aggregate figures hide the language that is performing badly, and it is almost always the smallest by volume — which is exactly why nobody notices.
2. 2Language detection accuracySample calls and check whether the agent's language matched the caller's, and how many turns it took to settle. This is the metric unique to voice.
3. 3Early hang-up rate by languageCallers who drop in the first few seconds. A high figure in one language usually means the greeting is wrong or the voice is poor rather than anything about the content.
4. 4Transfer outcomes by languageWhether transferred calls actually reached someone who could answer. This is where a multilingual line looks fine on a dashboard and fails in reality.
5. 5Knowledge coverage per languageThe same audit run three times. It is almost never equal, and the gap is in the language the material was not originally written in.

How Vexvon handles multilingual calls

Several DID numbers can sit behind the same system, which is the cleanest of the three language-selection strategies above: a different number can run a different language, so the caller's choice is made before the call connects and the greeting is right from the first word.

Scenarios define the language of a call alongside its direction, the voice and the fields to extract, and eight agent voices are available — so voice can be chosen per language rather than one voice being stretched across all of them. On the messaging side language is detected per message rather than fixed per conversation, and a company can restrict the list of languages it answers in.

The knowledge base is shared rather than duplicated: entries are typed into thirteen categories and targeted at the chatbot, the call agent or both, so material that genuinely differs by market can be held as its own entry while everything shared stays in one place. Answer transparency records which entry produced a reply, which makes a per-language coverage audit concrete rather than impressionistic.

Handover carries the language with it. Every channel writes to one customer record, so a caller who wrote in Russian on Instagram and calls in Azerbaijani is one record, and routing is built from time policies, destinations and priority-ordered rules that can be simulated before going live — which is how a language-aware routing rule gets tested without waiting for a real caller.

Frequently asked questions

1. 1How should a voice agent choose the language?A separate number per language is the cleanest option, because the choice is made before the call connects. Where one number serves everyone, use a short neutral greeting and detect from the caller's first utterance, treating that detection as provisional.
2. 2Should the agent ask which language the caller wants?Not as a menu. That reintroduces the friction the agent exists to remove. Offering a switch once, in plain words, is acceptable; offering it repeatedly is worse than guessing.
3. 3Why does pronunciation matter so much?Because callers judge competence on it, and a correct answer pronounced wrongly sounds like a machine reading a foreign word. Proper nouns, numbers and loanwords are where it breaks.
4. 4What should be localised rather than translated?Prices in local currency, delivery times, branches, opening hours and regulatory wording. Product mechanics and technical specifications can usually be shared. Spoken answers should also be shorter than written ones.
5. 5What is the most common operational failure?Transferring a call to a queue where nobody speaks the language. Route by language before topic wherever coverage differs, and record the language so callbacks are made by someone who can hold the conversation.
6. 6What should be measured?Everything split by language, plus detection accuracy and early hang-up rate — the two metrics unique to voice. Aggregates hide the language that is failing, and it is usually the smallest one.

Listen to twenty calls in your smallest language

The fastest diagnostic is to listen to twenty real calls in whichever language has the lowest volume. Detection errors, mispronounced proper nouns and untranslated closing messages concentrate there, and it is the language nobody reviews precisely because it is small.