Multilingual voicebot: one knowledge base, one team, many languages
Holding a conversation in a foreign language is the least of an AI agent's problems. The difficulties start when your knowledge base is in Polish and the customer asks in German, and when a case needs a human agent who does not speak the customer's language. We show how X-TALK solves both problems, in voice and in chat.
In short
- The X-TALK agent holds voice and chat conversations in more than 30 languages (over 40 regional variants), including German, English, Czech, Ukrainian and Spanish, and chat supports practically every widely used language.
- The agent greets the customer in the default language, recognises the caller's language and switches to it; you set the list of allowed languages and the default language per agent, number or widget.
- One knowledge base is enough, e.g. in Polish: when a customer asks in another language, the query is translated into the knowledge base language before searching.
- After a handover to a human, X-TALK translates the conversation live in both directions, in voice and in chat, so a Polish-speaking agent can serve a customer from Germany or Ukraine.
- The interpreter line is a separate number on which AI interprets the call between the caller and a member of your staff, with a glossary of terms.
- You check quality with test scenarios in every language and with the share of cases resolved by AI, measured separately for each language.
Does the voicebot speak Ukrainian, German and Czech?
Yes, the X-TALK agent holds voice conversations in Ukrainian, German, Czech, English and Spanish, among others, and supports more than 30 languages in total. The list also includes Slovak, Hungarian, Romanian, Lithuanian, Latvian, Estonian, the Scandinavian languages, Greek, Turkish, Arabic, Japanese, Korean and Chinese. Some languages have regional variants: German for Germany, Austria and Switzerland, Portuguese for Portugal and Brazil, Spanish for Spain, Mexico and Argentina.
Each language offers natural male and female voices, chosen so that they sound like a conversation rather than an announcement. Where needed, separate rules apply, e.g. Serbian in Cyrillic or Latin script, and greetings are prepared for every language. Chat supports practically every widely used language; the list of voice languages refers to full support with refined pronunciation.
Bear in mind that generating an answer in a foreign language is easy for today's models. The quality of multilingual service depends on two other things: whether the agent finds the right information in the knowledge base, and what happens when the case reaches a human. Those two topics take up most of this article.
How does the agent recognise the customer's language?
The agent recognises the language from the customer's first few utterances and switches to it without asking. This is how it works:
- the agent has a default language (e.g. Polish) in which it greets the caller or the person writing,
- when the customer starts speaking or writing in another language, the agent switches to that language,
- you can restrict the agent to selected languages, e.g. only Polish and English, if your company does not support others,
- separate phone numbers and widgets can have different default languages, e.g. a German-language number for customers in the DACH region.
In practice, that last setting matters a lot. A customer from Austria who calls the number printed on a German invoice hears a German greeting from the very first word. They do not have to wade through a Polish script, and the agent does not waste its first turn "guessing" the language.
In the voice channel there is also speech recognition to consider. X-TALK lets you provide transcription hints for each language, i.e. words the agent must hear correctly: product names, place names, industry terms. This matters especially for proper names, which are easy to mangle in a foreign language. We cover voice settings, silence and background noise in more detail in the playbook chapter Voicebot: voice conversation and noise.
Do you need to translate the knowledge base into every language?
No, a single knowledge base is enough, e.g. in Polish, because X-TALK translates the customer's query into the knowledge base language before searching. This solves a problem that is rarely discussed in multilingual rollouts.
Semantic search compares the meaning of the question and of the documents. It works best when both are in the same language. The question "Wie kann ich mein Passwort zurücksetzen?" and a Polish document titled "Reset hasła" (password reset) mean the same thing, but the match between them is weaker than between two texts in Polish. The result: the agent may not find the right procedure and gives a vague answer.
X-TALK does this automatically in four steps:
- The customer's question: "Wie kann ich mein Passwort zurücksetzen?"
- The query is translated into the language of the knowledge collection: "Jak zresetować hasło?" ("How do I reset my password?")
- The search runs in the Polish knowledge base, with full accuracy.
- The answer is given in German, based on the Polish document.
If the conversation language is the same as the knowledge base language, nothing is translated. If your company has documents in several languages, the agent picks the most relevant source. The same mechanism works in the voicebot and the chatbot, because both channels share the same search logic.
In business terms, this means updating your offer once. You change the price list or the terms and conditions once, in Polish, and the agent uses the new version straight away in every language. There is no risk of the German version of a document lagging two months behind.
When should you add a glossary or documents in another language?
A glossary is worth adding wherever precise naming matters: products, legal terms, names of medicines, parts or procedures. Query translation copes well with everyday language, but a proper name or an industry abbreviation can be translated literally and miss the document.
Three practical steps:
- Multilingual keywords in document metadata (the
keywordsandsymptomsfields) or a separate glossary: if a Czech customer asks about a specific device model, the document should "know" that name in the form the customer will use, too. - Documents for high-volume markets: if, say, 40% of your traffic is in German, prepare a collection with the most important documents in that language. The agent will pick the most relevant source by itself.
- Words to recognise in voice: add the same names to the transcription hints so that the agent hears them correctly before it even starts searching.
How does live phone call translation work?
Live translation works in both directions: the customer speaks German, the agent hears them in Polish, replies in Polish, and the customer hears German. This removes the hard limit of multilingual voicebots, which worked brilliantly right up until the case was handed to a human, because at that point you needed an agent who spoke the customer's language.
A typical call goes like this:
- The customer calls and talks to the AI agent in their own language. The agent collects the details, e.g. the order number and a description of the problem.
- The case needs a human, so before transferring, the agent tells the customer in their language that the conversation will be translated by AI.
- What the customer says is transcribed in the original language in real time and translated into the human agent's language.
- The human agent hears the conversation in Polish and sees both versions in the agent console: the original and the translation, along with a case summary in Polish and the details collected by the AI agent.
- The human agent's reply travels the same route in the opposite direction.
You set the human agent's language per queue or per transfer number. You also decide what should happen when a given language pair cannot be translated: connect without translation, or do not transfer the call. A queue might be called, for example, "German-speaking customers" or "Complaints DE" and have its own translation settings. Routing, queues and agent statuses are covered in the chapter Handover to agents and queues.
That leaves response time. In a voicebot, silence longer than 2–3 seconds sounds like a fault, and every extra translation or repeated search adds hundreds of milliseconds. That is why the agent only reaches for knowledge when it needs it, and in the meantime can naturally say "let me check". The better prepared your knowledge base, the fewer extra steps in each turn.
How does translation work in chat, WhatsApp and on the interpreter line?
In chat, translation works just as it does in voice, only on text: the human agent reads and writes in Polish, the customer in their own language, and the original message is always visible. This applies to the website widget, the mobile app and WhatsApp.
| Channel | How translation works | What your staff see |
|---|---|---|
| Conversation with the AI agent (voice and chat) | The agent replies in the customer's language; knowledge base queries are translated into the knowledge base language | A conversation transcript that can be translated |
| Voice call after handover | Live two-way translation, after a notice to the customer | The conversation in Polish, the original and the translation in the panel, a summary in Polish |
| Chat after handover (widget, app, WhatsApp) | Two-way message translation | The translation alongside the original for every message |
| Interpreter line | A separate number with the caller's and the employee's language set, with a glossary of terms | The conversation in their own language, on the phone or in the panel |
| Post-call transcripts | Translation of any transcript | The conversation record in their own language, e.g. for quality review or complaints |
The interpreter line is a separate scenario in which the AI agent does not run the conversation but only interprets. The caller dials a designated number and speaks in their language, and a member of your staff (on the phone or in the panel) hears them in theirs. Each number has the caller's language and the employee's language set. You can add a glossary to the number, e.g. names of medicines, parts or procedures, so that the translation sticks to your company's terminology.
Where do a multilingual voicebot and translation make the biggest difference?
The biggest impact is where customers or employees speak many languages and the service team speaks one. A single team of agents can then serve several countries without hiring people for every language.
| Industry | Typical situation | What X-TALK delivers |
|---|---|---|
| Logistics | Drivers from different countries call about loading, documents and delivery slot booking | The agent replies in the driver's language; the dispatcher speaks Polish with translation |
| Clinics and medical facilities | Foreign patients book appointments and ask how to prepare | Booking in the patient's language, an interpreter line with a glossary of medicine and procedure names |
| Public offices | Foreign residents ask about procedures and documents | Answers from a single Polish procedures knowledge base in the resident's language, without hiring interpreters |
| Cross-border e-commerce | Customers from Germany, Czechia or Spain ask about orders and complaints | The agent handles status enquiries, and complaints go to the Polish team with translation |
| HR | Employees from Ukraine or Asia ask about rotas, leave and documents | An assistant in the employee's language, conversations with HR with live translation |
| Service and repairs | Foreign customers report faults | Fault report details collected in the customer's language; the technician speaks Polish |
An e-commerce example: a customer from Germany calls because the table top from their order arrived scratched. The agent collects the order number and a description of the damage in German, explains that the call will be translated and transfers it to the "Complaints DE" queue. A Polish agent hears the customer in Polish, sees the case summary and offers to replace the table top, and the customer hears her reply in German. A company with a team like this can enter another market without separate support for each language; you will find more examples on the Multi-country support page.
How much does a minute of live-translated conversation cost?
A minute of live-translated conversation (interpreter line or handover with translation) costs 100 credits, i.e. from PLN 1.20 at the entry rate down to PLN 0.80 at large scale. That is the same rate as a minute of conversation with the voicebot. For comparison, a minute of conversation with a human agent after handover, without translation, is about 50 credits with recording and about 30 without.
The credit price falls with your plan and volume, and larger companies can pay for actual usage. You will find plan details and the credit converter in the price list. When working out the business case, compare the cost of translated minutes with the cost of hiring or outsourcing interpreters for every language you support today or plan to support.
How do you test a multilingual agent before launch?
You test a multilingual agent with Q&A scenarios run separately in each target language. An agent that answers correctly in Polish will not necessarily perform as well in Czech, because knowledge base matching behaves differently and proper names sound different.
In X-TALK, a test scenario describes a situation and the expected behaviour. During the test, AI plays the customer, the agent runs the conversation in chat or by voice, and an AI "judge" assesses whether the behaviour meets the criteria. Tools that write data are simulated in tests. A chat test costs about 1 grosz (PLN 0.01) per conversation, so scenarios in five languages cost next to nothing. The model does not answer identically every time, so run key scenarios several times (e.g. 3×) and in the voice channel too. For multilingual rollouts, the X-ONE team tests the agent in every target language before launch. The whole process is described in the playbook chapter Testing AI agents.
How do you measure voicebot performance in each language?
Measure performance separately for each language, above all by the share of cases resolved by AI without a handover to a human. An average across all traffic can hide a problem: a good result in Polish may mask a weaker one in Ukrainian if there are fewer Ukrainian conversations.
What to check regularly:
- the share of cases resolved by AI per language and how it changes after each knowledge base fix,
- language quality reports, which catch errors in a specific language, e.g. incorrectly inflected names or mixed languages,
- unresolved conversations: once a week, review about 10 of them; you can translate a foreign-language transcript into Polish with one click,
- a new scenario after every fix, so that the same error does not come back in another language.
Checklist: rolling out a multilingual agent
- Default language and list of allowed languages set for the agent.
- Separate numbers or widgets for markets that should have their own greeting language.
- Decision made: one knowledge base with query translation, or additional documents for the largest markets.
- Product names and industry terms added to the metadata, glossary and transcription hints.
- Queues have the agents' language and live translation rules set, including what happens with an unsupported language pair.
- The AI translation notice checked in every language.
- Test scenarios run in every target language, including in the voice channel.
- Share of cases resolved by AI measured separately for each language.
Frequently asked questions
Does the customer know the conversation is being translated by AI?
Yes. Before transferring to a human agent, the AI agent tells the customer in their language that the conversation will be translated by AI. The customer is not caught off guard by how the conversation sounds after the handover.
Can the human agent see exactly what the customer wrote?
Yes. In chat, every message shows the translation direction and the original text. In a voice call, the agent hears the translation and sees the transcript in the original language and in Polish in the panel.
What happens when a language pair cannot be translated?
You decide that yourself, per queue or transfer number. You can choose to connect without translation or not to transfer the call. It is worth deciding this before launch rather than when the first such case comes in.
Can the agent handle only Polish and English?
Yes. You can restrict the agent to selected languages and give different numbers and widgets different default languages. You switch on multilingual mode only where you need it.
Do I need documents in every language?
No. The agent uses a single knowledge base, e.g. in Polish, and replies in the customer's language, because the query is translated into the knowledge base language before searching. Documents in other languages are only worth adding for high-volume markets or where precise terminology matters.
Want to see a conversation in the language pair your customers use? Visit the Multi-country support page, build your own agent or book a PoC, in which we test the agent and live translation on your data.