30 voice minutes and 30 chat conversations of starting credit, no credit card needed - view pricing

Multilingual AI Voice Agent 2026: Serving International Customers in More Than 30 Languages by Phone and Chat

EchoCall Team avatar

EchoCall Team

Multilingual AI Voice Agent 2026: Serving International Customers in More Than 30 Languages by Phone and Chat

In short: A multilingual AI voice agent recognises the caller's language on the phone and carries on the conversation in that language. EchoCall supports more than 30 languages, and you switch automatic language detection for voice agents on or off. Widely spoken languages work better than rare ones, and names, addresses and numbers are the weak spot in every language. So only offer languages you have tested with native speakers first.

We rewrote this post in October 2026. The earlier version promised 85 languages, switching within milliseconds and processing on servers in the EU. That was wrong. EchoCall supports more than 30 languages, the EchoHub runs in data centres in Frankfurt and Karlsruhe, and we do not promise that the AI models process data only in the EU. We also removed a Bitkom figure on small and mid-sized firms and the example of a Frankfurt dental practice, because we had no source for either.

When a multilingual AI agent pays off

Most people in Germany speak German at home, but not all. According to the microcensus, around 77 % of the population in private households spoke only German at home in 2024, 17 % also used other languages, and 6 % did not speak German at home. 15.5 million people communicated at home mainly or only in another language, most often Turkish (14 %), Russian (12 %) and Arabic (9 %). [1]

The Federal Statistical Office stresses that the microcensus says nothing about how well someone speaks a language. [1] So the figure shows a habit, not a gap. Many people prefer to explain a problem in their home language, especially a bill, water damage or an appointment with a specialist.

In 2025 around 21.8 million people in Germany had an immigration background, 26.3 % of the population in private households. The most common countries of birth among immigrants were Poland and Turkey with 1.5 million each, Ukraine with 1.3 million, and Russia and Syria with 1.0 million each. [2]

Of 282.1 million overnight stays in German accommodation from January to July 2026, 47.1 million were by guests from abroad, roughly one in six. [3] English as a fallback does not reach everyone. In 2022 a quarter (25.3 %) of 25 to 64-year-olds in the EU said they knew no foreign language. [4]

How to tell whether it pays off for your business

Statistics describe a country, not your business. Your own figures take two weeks:

  1. Your team notes every conversation that failed because of language, with the language and the request.
  2. Your website analytics show which browser languages your visitors use.
  3. Bookings, orders and customer records show countries and languages, where you already hold that data.

If one language comes up regularly, test it. With ten languages at one call each, start with English and the most common other one.

BusinessIn the customer's languageWhat stays with people
Medical practice, home careappointment requests, opening hours, prescriptionsmedical questions, results
Hotel, holiday letarrival, check-in, parking, availabilityspecial rates, complaints
Online shoporder status, returns, shipping costscredit notes, goodwill
Property managementdamage reports with address and callback numberservice charge disputes, law
Trades, removalstaking enquiries, booking a site visitquote, fixed price

How language detection works on the phone

At every step a voice agent turns speech into text, has a language model write the reply and speaks it with a synthetic voice. Automatic language detection decides which language that happens in, and with EchoCall you switch it on or off for each voice agent. With detection on, a call goes like this:

  1. The agent greets the caller in your main language. The greeting comes before the caller says anything, so it can only be in one fixed language.
  2. The caller replies, for example in Polish.
  3. The agent recognises the language and continues in Polish. If the caller switches again later, the agent can follow. Only your own testing shows how reliably that works.

The greeting matters. Someone who does not understand German may hang up after a German-only greeting without saying a word. A second short sentence helps: "Guten Tag, hier ist der KI-Assistent der Praxis Weber. You can also speak English with me." Two sentences should be the limit.

When to switch detection off. If callbacks and appointments on site only happen in German anyway, an agent that speaks Turkish makes a promise your business cannot keep. Staying in German and saying so politely is more honest.

Mixed language and accents

People who speak two languages often mix them in one sentence, and some German words have no common equivalent: Krankenkassenkarte (health insurance card), Nebenkostenabrechnung (service charge statement), Aufenthaltstitel (residence permit). They come in German, in the middle of a Turkish or Arabic sentence. A single English "okay", on the other hand, should not trigger a switch. Whether the agent gets this right depends on the language, the line and the speaker. Test with sentences your customers actually say, spoken by people who sound like them.

And in chat?

In chat the visitor types, and the agent can reply in the same language. Set in the prompt that it answers in the visitor's language and which languages it offers. The EchoCall chat widget is available in several languages, and since October 2026 visitors can also talk to the agent by microphone in the widget. Those minutes come from the voice pool.

Quality by language: what you can realistically expect

No system understands every language equally well. A 2022 study of the Whisper speech recognition model shows how large the gaps can be: the more hours of training audio a language had, the fewer words the model got wrong, with a squared correlation coefficient of 0.83. By the researchers' estimate, the word error rate halves for every sixteenfold increase in training data. Languages with their own script that are only distantly related to the Indo-European languages did worse than expected, such as Hebrew, Telugu, Chinese and Korean. [5]

These figures describe a 2022 model, not EchoCall or today's systems. The principle holds: there are far more recordings of English, Spanish or German than of Albanian or Tigrinya.

As a rule of thumb: very widely spoken languages such as English, Spanish or German are usually solid, even with an accent. Widely spoken languages such as Turkish, Polish or Russian are often good but depend more on the line and speaking style, so test them thoroughly with native speakers. Offer rarely spoken languages only if the test convinces you.

Arabic is a special case: spoken Arabic varies a great deal between countries, so a caller from Syria sounds different from one from Morocco. Test with people from your customers' countries.

Names, addresses and numbers: the weak spot in every language

Even in German, voice agents find proper names and digits the hardest. In another language a second layer is added: a Turkish name in a German conversation, a German street in a Polish one.

WhatTypical problemWhat helps
Namesunfamiliar spelling, e.g. Öztürk, Wiśniewski, Nguyenask for spelling and repeat it back
Streets and townsGerman street names inside a sentence in another languagepostcode first, then repeat the street
Phone numbers97 in French is "quatre-vingt-dix-sept"confirm the digits one by one
Dates10/06 means 6 October in the US and 10 June in the UKconfirm with the weekday and the month name
Email addressesdot, hyphen and underscore have different names everywherespell it out or have it typed in chat

One rule belongs in every multilingual prompt: names, phone numbers and appointments are always repeated back for confirmation, digit by digit. If Gmail or SMTP is connected, the agent can also send a written confirmation.

One knowledge base, answers in many languages

You do not have to translate your knowledge base, because language models can reproduce content from German documents in another language. With EchoCall you fill it with PDF, TXT, Markdown and CSV files (voice agents also JSON), single web pages that are re-read regularly, a crawl of your whole website or free text.

The risk lies in translating on the fly. In a German medical practice, an "Überweisung" is a referral from the GP. If it becomes "bank transfer" instead of "referral" in English, the answer is wrong even though the knowledge base is right. Five rules help:

  1. Hard facts in one short document. Prices, opening hours, address and emergency numbers as a table. The agent only has to read out numbers.
  2. A small glossary. German terms your customers need to know, with a short explanation per language.
  3. Do not translate proper names. Company, product and street names stay as they are. Say so in the prompt.
  4. Only summarise legal texts. For the binding wording of terms and conditions, cancellation policy and privacy notice, the agent points to the original.
  5. Use existing language versions. If your website has an English or Turkish version, add it to the knowledge base.

Voice, pronunciation and forms of address

A voice that sounds natural in German may have an accent in another language. Listen to each language with a native speaker before you offer it, especially your company name and address. If a name sounds wrong, write it in the prompt the way it should be spoken.

The formal German "Sie" has its own form in each language: "vous" in French, "usted" in Spanish, "siz" in Turkish, "вы" in Russian. Set in the prompt that the agent always uses the polite form, short sentences and one question at a time. Callers using a second language will appreciate it.

The AI disclosure in the caller's language

Article 50 of the EU AI Act has applied since 2 August 2026: anyone interacting with an AI system must be told so clearly, at the latest at the first interaction. [6] By its wording the duty is addressed to the providers of such systems. Our post on the AI disclosure duty, linked below, covers what it means for businesses deploying an agent. For multilingual agents one point matters most: a notice the caller does not understand is hardly clear.

  • On the phone you write the notice into the agent's first message yourself. The editor shows an info box for this, and the help centre has example sentences. Because the greeting is in your main language, add to the prompt: "If you switch language, say in your first sentence in the new language that you are an AI assistant."
  • In chat the widget shows an automatic AI notice in German, English and French. For other languages, the same prompt instruction makes sure the first reply contains the notice.

Under our terms, the customer is responsible for giving the notice. These sentences are a starting point. Have native speakers check them:

LanguageExample sentence
GermanSie sprechen mit dem KI-Assistenten von [Firma].
EnglishYou are speaking with the AI assistant of [company].
FrenchVous parlez avec l'assistant IA de [entreprise].
SpanishEstá hablando con el asistente de IA de [empresa].
Turkish[Firma] şirketinin yapay zekâ asistanıyla konuşuyorsunuz.
RussianВы разговариваете с ИИ-ассистентом компании [название].

A test plan for each language

Each language goes live only after its own test, ideally by a native speaker who sounds like your customers, such as a colleague, an acquaintance or a translator paid by the hour. These eight test cases are enough to start with:

No.Test casePassed if
1reply to the greeting straight away in the target languagethe agent switches and stays in that language
2the most common question plus a price or opening timecontent and numbers are exactly right
3book an appointment, spell the name, dictate a numbername, number, date and time are correct
4a German technical term in a sentence in another languagethe agent places it correctly
5switch language mid-conversationthe agent follows without losing the thread
6a question that is not in the knowledge basethe agent says so and offers a callback or transfer
7listen for the AI disclosureit comes in the caller's language
8call from a noisy place or with poor receptionthe agent asks again instead of guessing

The tester rates each call: did the agent understand, was the content right, did it sound natural? If a case fails, adjust the prompt or the knowledge base and test again. After launch, read the first 20 transcripts per language, then a monthly sample is enough.

Where a multilingual agent reaches its limits

  • Rare languages and strong regional variation. This is where recognition varies most. Better to offer one language fewer than one that does not hold up.
  • The handover. Transferring to a German-only team just moves the problem. Voice agents can transfer to a phone number, for example a colleague who speaks the language. Otherwise the agent records the full request and you call back with support.
  • Sensitive topics. Medical, legal and financial advice stays with people, all the more so in a foreign language.
  • Anything in writing. Invoices and contracts usually stay in German. Say so before the customer expects a confirmation in their language.
  • Call length. In a second language there is more asking back and spelling. Plan minutes for that.

What a multilingual AI agent costs with EchoCall

The plans are the same whatever language the agent speaks. [7] Voice agents cost €75 for 250 minutes, €150 for 500, €450 for 1,500 and €1,500 for 5,000+ minutes a month. That is around €0.30 per minute, with no setup fee. Chat agents cost €40 for 100 sessions, €120 for 300, €400 for 1,000 and €1,000 for 5,000+ sessions. Phone numbers and telephony costs are not included.

To try it out there are 30 voice minutes and 30 chat conversations for 14 days, and the credit card is verified with €1. That covers the test plan for a first foreign language. The live demo shows how an agent sounds without signing up: enter your website address, and after about 20 seconds you can call or write to it in your browser.

Frequently asked questions about multilingual AI agents

How many languages does an EchoCall AI voice agent speak?

More than 30. You decide which of them to offer. We recommend only languages you have tested with native speakers first.

Does the agent automatically recognise which language the caller speaks?

Yes, voice agents have automatic language detection, which you switch on or off. The greeting comes before the caller speaks, so it is in your main language. A second short sentence in another language helps callers who do not understand German.

Do I have to translate my knowledge base into every language?

No. The agent can reproduce content from a German knowledge base in other languages. Put hard facts in one short document, add a glossary and do not let legal texts be translated freely.

Does the agent understand accents and mixed language?

It depends on the language, the line and the speaker, and widely spoken languages work better than rare ones. Test with people who speak like your customers, using sentences with German technical terms.

Does the AI disclosure have to be in the caller's language?

Article 50 of the AI Act requires a clear notice at the latest at the first interaction, and a notice in a language the caller does not speak is hardly clear. On the phone you write it into the greeting and have the prompt repeat it after a language switch. In chat the widget shows it automatically in German, English and French.

What does a multilingual AI agent cost?

The same as a single-language one: voice agents from €75 a month for 250 minutes, chat agents from €40 for 100 sessions. Phone numbers and telephony costs come on top.

Sources

  1. Destatis (Federal Statistical Office), figure of the week, 17 February 2026, "77 % der Bevölkerung sprechen zu Hause ausschließlich Deutsch" (in German): destatis.de
  2. Destatis, press release no. 128, 13 April 2026, population with an immigration background 2025 (in German): destatis.de
  3. Destatis, press release no. 324, 11 September 2026, tourism in July 2026 (in German): destatis.de
  4. Eurostat, Statistics Explained, Foreign language skills statistics, Adult Education Survey 2022: ec.europa.eu
  5. Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision, 2022, section 3.4 and figure 3: arxiv.org
  6. Regulation (EU) 2024/1689 (AI Act), Article 50: eur-lex.europa.eu
  7. EchoCall, pricing: echocall.de/en/pricing

All links were checked on 6 October 2026. This post is not legal advice.

You might also like

Share this post
cta image
cta image

Ready to try EchoCall?

Start with 30 voice minutes and 30 chat conversations of starting credit at hub.echocall.de, valid for 14 days, no credit card required.