Multilingual shop-floor training with local AI: instructions, briefings, spoken questions

Training workers in their own language with local AI

Multilingual shop-floor training with local AI means three things on a server the company runs: approved work instructions and safety briefings translated and simplified into each worker's language, spoken questions at the workstation answered from that approved content only, and short quizzes that document the instruction. The content is approved by the people responsible for safety and production; the AI translates, finds and reads aloud. A 2U server with four NVIDIA RTX PRO 4500 Blackwell Server Edition cards, or two such cards in a server you already run, carries the models for a plant.

We wrote this for the production manager and the health and safety lead of a plant where the workforce speaks many languages, temporary staff change often, and a new colleague needs to understand a work instruction on the first shift. The obligations come first, then the method, then the hardware for whoever runs the IT.

What the employer has to provide under EU law

Council Directive 89/391/EEC of 12 June 1989, the framework directive on safety and health at work, sets the obligation. Article 12(1) requires the employer to ensure that each worker receives "adequate safety and health training", "in particular in the form of information and instructions specific to his workstation or job". It has to be given "on recruitment", "in the event of a transfer or a change of job", "in the event of the introduction of new work equipment or a change in equipment" and "in the event of the introduction of any new technology". The training must be "adapted to take account of new or changed risks" and "repeated periodically if necessary". The same article says it may not be at the workers' expense and takes place during working hours. Article 10 adds the information workers must receive on the risks and on first aid, fire-fighting and evacuation.

The directive does not say in which language the training is given. Member States transpose it into national law, and the national rules and your safety specialist decide how you show that a worker has understood. This section is general information as of October 2026, not legal advice; your legal counsel and safety specialist confirm what applies in each plant.

For the plant that has the obligation, the practical problem is volume: every work instruction, every change to a machine and every new colleague, in every language spoken on the floor. That is where a local AI system helps.

Answers only from approved content

The system never invents safety content. Every answer comes from a document that a responsible person has approved, and the system says so when no approved document covers a question.

1. Source documents. Work instructions, safety briefings, operating instructions for machines and hazardous substances, written and approved in the plant's working language as today.
2. Simplified and translated versions. The language model produces a version in plain language and translations into the languages spoken on the floor. A bilingual colleague or a translator checks each translation of safety content before it is released.
3. Questions at the station. The assistant searches only the approved versions, answers in the worker's language, shows the passage it used and names the document and its version.
4. Questions it cannot answer. If no approved passage fits, or a question concerns an immediate danger, the answer is to stop and call the supervisor, and the question goes to a list for the safety specialist.

This is retrieval over your own documents, as our guide to RAG on company data explains, with the index limited to approved versions. Our guide on how good local LLM answers are compared with ChatGPT covers what to test before use.

Translating work instructions and safety briefings with AI

Translation is the part with the largest volume and the clearest review step. A language model translates a work instruction with its context, keeps numbers, units and part names, and can produce a plain-language version first, with short sentences and one action per step. The translated text is reviewed once and then reused until the source changes.

Hugging Face model cards Qwen/Qwen3-14B-FP8 (with its file list) and utter-project/EuroLLM-9B-Instruct, read 25 October 2026.

The Qwen3-14B model card claims "Support of 100+ languages and dialects", and its FP8 files add up to 16.3 GB. EuroLLM-9B-Instruct names all official EU languages among its 35, but its model card warns that it "has not been aligned to human preferences, so the model may generate problematic outputs", which makes the review step even more important. For dedicated translation models and their licences, our guide to on-premise machine translation compares NMT models and translation LLMs, and our comparison of open LLMs for European languages covers language quality. Test the languages of your own workforce with two or three real instructions before choosing.

Spoken questions at the workstation

At the station, a worker often has gloves on and no time to type. A tablet or a terminal with a headset takes the spoken question, the server turns it into text, the assistant answers from approved content, and the answer is shown on screen and read aloud.

Hugging Face model cards openai/whisper-large-v3, nvidia/parakeet-tdt-0.6b-v3, Qwen/Qwen3-14B-FP8 and ResembleAI/chatterbox, read 25 October 2026.

Speech output is the narrowest stage. Chatterbox's multilingual model lists 23 languages, among them Arabic, German, Greek, Polish, Russian and Turkish, but not Romanian, Czech, Hungarian, Bulgarian or Croatian, and its model card states that every audio file it generates carries Resemble AI's Perth watermark. For languages without a voice, the answer is shown as text in that language. Parakeet covers Romanian, Czech, Hungarian, Bulgarian and Croatian among its 25, so the question can still be spoken.

Show the recognised question on the screen before the answer. Whisper's model card states that "the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination)", and a worker who sees a wrong transcript can repeat the question. Noise at a machine makes recognition harder, so a headset microphone helps; test at the loudest station. Our guide to a voice agent server on-premise covers live speech pipelines and their latency in depth.

We supply the RTX PRO 4500 Server Edition as a card for servers you already run and build 2U servers with four of them; send us the languages on your shop floor, the number of stations and your documents, and our team replies with a configuration.

Quizzes and the record of instruction

A short quiz after a briefing shows that the content reached the worker and gives the plant a record. The language model can draft questions from the approved briefing in each language. The safety specialist approves the questions and the answer key, and the quiz is then scored against that key, in the same way as on paper.

Keep the quiz a record of instruction. Annex III point 4(b) of the AI Act lists as high-risk AI systems intended "to monitor and evaluate the performance and behaviour of persons" in work relationships, and under Article 113 as amended those rules apply from 2 December 2027. A fixed answer key and a record of who completed which briefing stay outside that description; an AI that rates individual workers' understanding or behaviour comes close to it. Your legal counsel confirms the line for your plant. Article 50(1) also obliges providers to design AI systems that interact directly with people so that the people are informed of it, so label the assistant as an AI system on the screen. Teaching staff about AI itself is a separate obligation, covered in our guide to AI literacy training under Article 4.

Questions and quiz results are personal data once they are linked to a person. Decide with the data-protection officer which records are kept, for how long, and who sees them; anonymous questions at a station need no link to a person at all.

Hardware: stations on the floor, one server in the server room

The tablets at the stations only capture audio and show text. The models run on one server, connected over the plant network.

Usable memory 25.7 GiB per RTX PRO 4500 Server Edition by our users-per-card rule, Qwen3-14B FP8 15.2 GiB with 0.625 GiB of FP8 cache per 8,192-token conversation; power by our estimate (cards at 165 W, two AMD EPYC 9355 at 280 W, plus 15 per cent), to be confirmed by a measurement on the delivered system.

The model on one card. Our users-per-card rule takes 90 per cent of the memory the driver reports, less 3 GiB per card: for the 32 GB RTX PRO 4500 Server Edition that is 25.7 GiB, as our guide to the four-card RTX PRO 4500 Server Edition server derives. Qwen3-14B's FP8 files of 16.3 GB are 15.2 GiB (FP8 stores each model weight in 8 bits, half the usual 16), which leaves 10.5 GiB. Each running conversation keeps a working memory, the cache, on the card; with an FP8 cache of 0.625 GiB per conversation of 8,192 tokens (tokens are the word pieces a model reads and writes) one card holds 16 conversations at the same moment. A spoken question with its retrieved passages is short compared with 8,192 tokens, and a pilot measures how many questions arrive at the same moment.

Speech on a split card. MIG (Multi-Instance GPU) is NVIDIA's way of dividing one card into isolated parts that behave like separate cards; the RTX PRO 4500 Server Edition splits into two halves of 16 GB. Whisper large needs about 10 GB by OpenAI's repository, which fits one half; Chatterbox's model card gives no memory figure, and at 0.5B parameters its weights are about 1 GB in 16-bit by our estimate, so the second half holds it with room to test. The card draws 165 W at most, passive and single-slot, for servers with server airflow.

Four cards in 2U. With two copies of the answer model, a card for speech and a card for translation runs and tests, the server draws about 1.4 kW by our estimate and fits one 230 V 16 A feed. Our buy page for four RTX PRO 4500 Server Edition lists the configuration.

One company server. If the company already plans an AI server for other departments, the training assistant can run on it. Our entry configuration with two RTX PRO 6000 Server Edition cards splits a card by MIG into four 24 GB instances, which hold the speech models and a copy of the answer model next to the office assistants. Several departments on one server are usually a better use of the hardware than a server per task.

Notes for the IT team

1. Network. The stations reach the server over the plant network or Wi-Fi; put the engines behind a gateway with keys and a query log, as our guide to an LLM gateway describes. vLLM's --api-key protects only paths under /v1, /v2, /inference and /cohere (vLLM 0.32.0 security guide), so keep the engines on an internal network.
2. When the server is unreachable. Keep the approved documents of each station on the tablet as files, so instructions stay readable if the network or the server fails; only questions and speech need the server.
3. Versions of content. The index holds only approved versions with a version number; a changed source document withdraws its translations until they are checked again.
4. Accounts. Station accounts are enough for questions; a personal login is needed only where a quiz result is recorded for a person.
5. Software on delivery. The server arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime; the Private AI/ML service sets up the models and retrieval; the station app and its rollout on the tablets stay with your team or your software provider.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the RTX PRO 4500 Server Edition as a card for servers you already run, after a compatibility check of platform, power and cooling, and build 2U servers with four cards or company AI servers with RTX PRO 6000 Server Edition cards to order, assembled and tested before delivery, delivered within the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up translation, retrieval and speech on it. Our guide to capturing expert knowledge before staff retire shows how new instructions can come from your experienced staff.

FAQ