Simultaneous Basque translation with AI: serve your customers in their language and comply with Spain's Law 10/2025

Simultaneous Basque translation with AI lets an agent serve customers in Spanish while the customer speaks their own language. In a pilot we measured a 3.0 s wait and 4.7 out of 5 accuracy. Spain's Ley 10/2025 requires many companies to serve customers in the co-official languages.

What simultaneous Basque translation with AI is

Simultaneous Basque translation with AI is a call in which the customer speaks Basque and the agent hears it in Spanish, almost in real time. The agent replies in Spanish and the customer hears the reply in Basque. It is what many companies look for when they search for a simultaneous translator or Basque interpreter. A person stays in charge.

It helps to tell it apart from three other services. Text translation does not cover voice or data dictated live. A human interpreter does the same job but is paid by the minute. An automated voice agent handles the call alone, whereas here the human agent sets the language and can cut in or ask for something to be repeated.

It fits Ley 10/2025 (Spain's customer service law): the person stays on the other end and the AI covers the language. In one real case, a company's help desk only had content in Spanish. Covering the co-official languages with in-house staff requires entire teams.

What Ley 10/2025 requires on co-official languages

Ley 10/2025, de 26 de diciembre, de servicios de atención a la clientela (Spain's customer service law) was published in the BOE (Spain's Official State Gazette) on 27 December 2025 and came into force the following day. Its single transitional provision allows twelve months to adapt, until 28 December 2026.

Article 2.1 applies it, with no size threshold, to basic services of general interest: water, gas, electricity, passenger transport, postal services, telephony and financial services. Article 2.2 extends it to other companies that sell to consumers and have at least 250 employees. It also covers those with turnover above 50 million euros or a balance sheet above 43 million. Meeting any one of the three thresholds is enough. In financial services the law applies as a supplementary rule (art. 2.5).

Article 23 punishes non-compliance as a consumer infringement. The regime of the consolidated text of the Ley General para la Defensa de los Consumidores y Usuarios (General Law for the Defence of Consumers and Users) and regional rules apply. Under its article 49.1, minor fines range from 150 to 10,000 euros and serious ones from 10,001 to 100,000. Very serious fines range from 100,001 to 1,000,000 euros. They can exceed those amounts, up to between two and eight times the unlawful profit obtained. For very serious infringements penalised under Regulation (EU) 2017/2394, the maximum reaches 4% of annual turnover (art. 49.5).

How an AI simultaneous interpreter works

The AI simultaneous interpreter chains speech recognition, context-aware translation and speech synthesis. At Lagunbot we offer it as a video call in the browser, with nothing to install.

1. The customer speaksIn Basque, Catalan or Galician. The agent sets the language or presses "Detect language".
2. Speech recognitionThe audio becomes text. For Basque we use an open model from HiTZ, a research centre of the UPV/EHU.
3. Context-aware translationThe model sees the history and the company glossary.
4. VoiceIt sounds sentence by sentence, from the first natural pause.
5. Parallel verificationData containing figures is transcribed twice and the code compares the versions.
6. The agent hears SpanishThe customer hears the agent in their own language.

Two engines you can swap mid-call

The standard engine waits for the sentence to finish and translates it with full context. It works with Basque and has a median wait of 3.0 s. The real-time engine interprets before the sentence ends. Its median wait is 1.8 s and it costs roughly twice as much. In the pilot it had no voice into Basque, so we measured it in Catalan.

Data verification with double transcription

Every sentence containing an ID number, IBAN, phone number, postcode or amount is transcribed twice. If the two versions do not match, the agent sees a warning, for example "postcode: 08014 or 08114?". They press "Ask to repeat" and the interpreter asks the customer in their language. The check runs in parallel. Each call leaves a report with wait, cost and reliability, exportable to CSV.

What we measured in the pilot

We ran a three-day pilot in November 2026, not yet in production.

The wait runs from the moment the speaker's voice stops until the interpretation is heard. A virtual customer simulated the 36 calls and an automated judge, a separate AI model, scored each sentence without knowing which configuration was used. Wrongly recognised words were measured with synthetic voice: with real people and over the phone, results will be somewhat worse.

In Basque we assessed 9 calls, with an accuracy of 4.74 out of 5 and a wait of 3.1 s. In Catalan, accuracy was 4.72 with the standard engine and 4.91 in real time; the latter had no serious errors. In total there were 84 sessions and 476 sentences.

The serious errors we found

Almost all were of the same kind: the translation corrected a data item using context instead of repeating what was said. Contract 34712 came out as 34792. In Basque, "zero zortzi zero bat lau" (08014) came out as 08004. The answer is a strict fidelity rule and an automatic check of every data item. It already exists for figures and amounts in Basque. It is still missing for names, spelled-out letters and addresses.

A real test between two countries

On 6 November we held a 24-minute call between two people in different countries. We interpreted into English, Catalan and Galician and switched engines mid-call. The standard engine had a median of 2.6 s. The real-time engine ranged between 0.5 and 1.9 s, apart from one long sentence of 7.6 s.

Basque: why it is the hard language and how we solve it

Basque is the hard language because the commercial real-time recognition model we tested barely understands it. On five real clips from the HiTZ corpus we measured the words wrongly recognised.

The HiTZ model has an open licence and the audio never leaves our server.

What fails: non-Basque names ("Sants" came out as "zantz"), spelled-out letters ("C letra" was translated as "zeta") and numbers, which Basque counts in twenties. Letters and numbers are already corrected with fixed code rules; names will be handled with name lists for each call.

Voices and our own library

Today there is no native Basque voice in the commercial voice engines we use for the other languages. In the pilot, a multilingual voice reads Basque with 3% of words wrong when recognised again, but native speakers have yet to judge the accent. That is why the plan includes our own library of six components, built with native speakers.

1. Language identification in the first seconds.
2. Recognition adapted from the HiTZ model.
3. Normalisation of numbers, letters, ID numbers, IBANs, phone numbers and dates, with more than 500 test cases.
4. Data fidelity: it alerts the agent and never corrects on its own.
5. Glossary and style, from 104 official terms of the Basque Government.
6. A fixed test bench that is rerun after every change.

Basque remains experimental: it still needs validating with real speakers and with phone audio.

Galician and Catalan

Catalan is the most measured language and has native voices. Galician also has native recognition and voice. We tested it in the real call, but the judge did not score it and a test with native speakers is still needed.

Integrations: phone, web, help desk and data

Today the interpreter works as a video call in the browser. The rest is built during implementation.

Latency and quality over the phone have not been measured yet. We measure them in the first month. If you are looking for an agent that handles calls by voice without a person, see our agentes de voz (voice agents). For text and web, we have a chatbot en euskera (Basque chatbot).

What it costs compared with hiring or outsourcing

The technology cost of the interpreter is measured in cents per call minute. What matters is comparing it with the two alternatives a company has today.

• Human telephone interpreter. As a reference, around 3 euros per minute excluding VAT. Ten minutes would be around 30 euros.
• In-house staff. Covering office hours in two languages takes at least two people per language. Four people at 2,000 euros a month in company cost, a low figure, add up to 96,000 euros a year.

Privacy and GDPR

Lagunbot acts as data processor. The conversations belong to the client company.

• Server in the EU. The test version already runs in the European Union. Before going into production, data processing agreements are signed with the providers.
• Keys only on the server. The browser never sees them.
• Notice to the customer. At the start they hear that the interpretation is automatic and the voice synthetic.
• Consent to clone the voice. Voice is personal data. Express consent is required and each person can only clone their own.

For where the data lives, see our page on IA privada (private AI).

Three-month implementation plan

The plan lasts thirteen weeks.

1. Month 1, weeks 1 to 4: data and baseline. Consents, scripts of real calls and 8 to 10 hours of audio from the sector, including by phone.
2. Month 2, weeks 5 to 9: language library. The six components with their tests and the evaluation of voices by native speakers.
3. Month 3, weeks 9 to 13: integration and pilot. Connection to the help desk and the switchboard, supervised pilot and the decision to go into production.

The target for Basque is a maximum of 8% of words wrongly recognised with real speakers and phone audio.

Frequently asked questions about simultaneous Basque translation

Does it work over the phone?

Today it works as a video call in the browser, with nothing to install. The bridge to the switchboard, so the customer calls the usual number, is built during implementation. Quality over the phone has not been measured yet and we assess it in the first month with real audio.

How long does it take to translate?

With the standard engine the median wait is 3.0 s from the moment the speaker stops talking. In Basque we measured 3.1 s, and 3.6 s from Basque into Spanish. The real-time engine drops to 1.8 s, measured in Catalan.

Does it work for Galician and Catalan?

Yes, at different levels of maturity. Catalan is the most measured language, with an accuracy of 4.7 out of 5 on the standard engine and 4.9 in real time. Galician has native recognition and voice and we tested it in a real call. It still needs measuring with the automated judge and with native speakers.

Does Ley 10/2025 apply to my company?

It depends. It applies with no threshold to basic services of general interest. It reaches other companies that sell to consumers if they have at least 250 employees, more than 50 M€ of turnover or more than 43 M€ on the balance sheet. Consult your legal adviser.

Does it replace staff?

No. The law prohibits using automated means as the only channel and requires that customers can speak to a person. The interpreter helps the agent serve in languages they do not speak. The agent sets the language, sees the written conversation and can cut in or ask for something to be repeated.

What happens with personal data?

Lagunbot acts as data processor and the client company is the controller. The test version runs on a server in the European Union and Basque recognition runs on our own server. Data processing agreements with the providers are signed before going into production. Keys live only on the server and voice cloning requires each person's express consent.

Can it be tried out?

Yes. We prepare a demo with your scripts and your vocabulary. You will see the interpretation in Basque, Catalan and Galician, the data verification and the report for each call. It is a technical pilot, not a production service. Write to us from the contact page and we will arrange it.

Sources

• BOE: Ley 10/2025, de 26 de diciembre, de servicios de atención a la clientela
• BOE: Real Decreto Legislativo 1/2007, texto refundido de la Ley General para la Defensa de los Consumidores y Usuarios
• HiTZ, research centre in language technologies of the UPV/EHU