AI server for care homes: voice documentation with health data kept in-house

What voice documentation does in a care home

AI care documentation by voice lets carers speak their care notes instead of typing them: a speech recognition model turns the recording into text, a small language model sorts it into the fields of the care record, and the carer reads, corrects and signs the note in the care software. On a server in the home or in the provider's own data room, recordings and notes never leave the organisation, and by our estimate one small server with two to four NVIDIA L4 cards carries several homes.

This guide is for managing directors, heads of care and home managers at providers with several homes and at home-care services who want to cut the time spent on writing without sending residents' data to an outside service. It explains how the process works, what the GDPR asks when health data is involved, whether one server for all homes or one per home makes more sense, and what the hardware needs in the building. A short section at the end is for your IT partner.

Two terms come up throughout. A server here is a computer that stands in a locked room and serves many users over the network. A GPU, or graphics card, is the part inside it that does the heavy computing for speech and language models; the NVIDIA L4 is a small GPU of this kind that draws 72 W.

Dictating a care note: from spoken note to signed entry

In our example a dictation takes about a minute. A carer finishes the morning round with a resident, opens that resident's record on the ward phone or tablet, presses record and says what matters: what the resident ate and drank, how the wound on the left heel looked, that the resident felt dizzy when getting up. The rest happens in five steps.

1. Recording. The app sends the recording over the home's wireless network to the server.
2. Speech to text. A speech recognition model writes out what was said, with punctuation, within seconds.
3. Sorting. A language model, the same kind of software as a chatbot but much smaller, places each statement into the matching part of the record, such as nutrition, wound, mobility or mood, and marks anything it is unsure about.
4. Checking. The carer reads the draft on the screen, corrects names of drugs or anything misheard, and signs the entry in the care software. Nothing goes into the record unsigned.
5. Handover. At the end of the shift, the model drafts a short handover for each unit from the signed notes, and the shift lead checks it before the handover meeting.

Each step keeps the carer in charge. The model writes drafts; the professional decides what goes into the record. Many care teams work in several languages, and the models can help there too: NVIDIA's Parakeet TDT 0.6B v3 detects the spoken language by itself, and Qwen3-8B, the language model in our sizing, supports more than 100 languages and translation according to its model card. Whether the result is good enough for your terms, abbreviations and accents is something to test with your own recordings before any decision.

The connection to your care documentation software depends on that software. Some products accept text through an interface, others only through copy and paste; ask your software provider which way it offers.

Care home AI and the GDPR: health data under Article 9

Care notes are personal data concerning health. Article 4(15) of the GDPR defines such data as "personal data related to the physical or mental health of a natural person", and Article 9(1) puts it among the special categories whose processing is prohibited unless an exception applies. Article 9(2)(h) covers processing necessary for "the provision of health or social care or treatment", under the conditions of Article 9(3), which ties it to professionals subject to professional secrecy. A voice recording of a care note contains health data just as the typed note does, and so are the transcripts, the drafts and the server's logs.

Regulation (EU) 2016/679, consolidated text on EUR-Lex, read 25 October 2026; the right-hand column is our summary of typical measures, not legal advice.

Article 35(3)(b) names "processing on a large scale of special categories of data referred to in Article 9(1)" among the cases that need a data protection impact assessment (DPIA), a written assessment of the risks and the measures against them. Our guide to a DPIA for an internal LLM shows how one is built. Running the models on your own server removes one question from the DPIA, because no outside AI service receives the data, as our article on local AI and the GDPR explains. It does not answer the others: the legal basis, how long recordings are kept, who may read which notes. Those are decisions for your data protection officer and, where your national care rules set requirements for documentation, for the people who know them. Our recommendation is to delete the audio as soon as the carer has signed the note, since the signed text is the record.

Documentation tools stay outside the Medical Device Regulation as long as they only record and summarise. Our guide to a private LLM for hospitals quotes the regulation's recital that "software for general purposes, even when used in a healthcare setting," is not a medical device, and explains when software with a medical purpose, such as one that suggests diagnoses, becomes one. Doctors' consultation notes are the subject of our guide to an on-premise AI scribe for medical practices.

We supply the server with L4 cards and, under our Private AI/ML service, set up the speech and language models; tell us how many homes, carers per shift and languages you have, and our team replies with a configuration.

One server for all homes, or one per home

Most providers with several homes do best with one central server in the main office or the provider's data room. The homes reach it over their existing connection to the office. One server is cheaper to run, gets its updates in one place and holds one access list. The point to plan is the line between the homes and the server: if it fails, carers cannot dictate until it is back, so either the app keeps recordings on the device until the line returns, or the home falls back on typing. Ask your software provider how its app behaves.

A server in each home makes sense for a large home with a weak connection, or for a home that sits apart from the others. It needs a lockable, ventilated room and someone on site who can restart it. Without any room of your own, our Private AI/ML service runs the models on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU; the DPIA then names that location and the service provider.

How much computing several care homes need

Speech recognition needs little computing, and the busiest moment of the day decides the size. Dictations bunch up in the last hour before a handover, when carers catch up on their notes. The worked example below uses example values that you replace with your own figures.

Our estimate with example values. Speed from E2E Networks' L4 test as quoted in our speech-to-text guide (Whisper large-v3-turbo 41.1 times faster than real time, Parakeet TDT 0.6B v3 238.9 times, short test clips, an indication only); dictations of about one minute and 20 seconds of model time per note (about 300 tokens of output) are our assumptions.

By this estimate one L4 handles speech recognition for all homes in the example with room to spare, and Parakeet TDT 0.6B v3, about six times faster in the same test, leaves even more. The language model needs its own card. By our users-per-card rule, which takes 90 per cent of the memory the card reports and subtracts 3 GiB for the software and the size of the model, an L4 keeps about 17.2 GiB for model and conversations with memory error correction (ECC) on. Qwen3-8B, stored in FP8, a compact number format of one byte per value, takes 8.8 GiB of that and leaves room for 14 conversations of 8,192 tokens, three times the four notes in progress of the example. A token is a piece of a word; 8,192 tokens are several pages of text, far more than one care note. A speed limit applies: the card's 300 GB/s of memory bandwidth caps one answer from this model at about 32 tokens per second by our arithmetic, so a long handover summary takes tens of seconds, which is fine for drafts that a person reads afterwards.

The hardware in one table

These options cover a single home up to a large provider. The NVIDIA L4 has 24 GB of memory, needs no extra power cable and is cooled by the server's fans. For larger providers, or for a larger language model, the NVIDIA RTX PRO 4500 Blackwell Server Edition has 32 GB at 165 W.

Our estimates. L4 72 W, 24 GB (NVIDIA product page as quoted in our L4 buying guide); four L4 with two processors about 0.8 kW and four RTX PRO 4500 Server Edition about 1.4 kW at the wall (cards, processors and 15 per cent for the rest), confirmed by a measurement on the delivered system; carers per shift from the worked example above; RTX PRO 4500 Server Edition 25.7 GiB per card by our rule, Qwen3-8B FP8 (25.7 − 8.8) / 0.5625 = 30.

The four-card L4 server is the option we would propose to most providers with several homes: two cards for speech and two for the language model mean that the service keeps running if one card fails. It runs from one ordinary 230 V 16 A feed, which carries 2.94 kW when planned at 80 per cent. Servers are loud and give off their power as heat, so they belong in a closed, ventilated room, not in the nurses' office; our guide to a GPU server in an office covers noise and heat. Our L4 buying guide lists the chassis that take four or eight L4 cards.

Notes for your IT partner

1. Models. Parakeet TDT 0.6B v3 is under CC BY 4.0 and lists Ampere, Hopper, Blackwell and Volta as supported architectures; the L4, an Ada card, appears in its test hardware list. Its card gives a German word error rate of 5.04 per cent on FLEURS. Whisper large-v3-turbo has 809 million parameters, 99 languages and an MIT licence tag. Qwen3-8B has 8.2 billion parameters and an Apache 2.0 licence; its FP8 weight files total 9.43 GB. Our speech-to-text server guide compares the speech models in detail.
2. FP8 on the L4. Qwen's FP8 release scales its weights in blocks of 128 by 128, and the L4 is an Ada card; confirm that the serving engine you use runs this checkpoint on the L4, as our L4 buying guide notes. The RTX PRO 4500 Server Edition is a Blackwell card.
3. Network. The server sits in the internal network; homes reach it over the existing site-to-site connection. No inbound access from the internet.
4. Access. Sign-in through the directory the care software already uses, where possible; carers see only the drafts they recorded.
5. Retention. Audio deleted after signing, drafts deleted after a short period, logs with health data restricted; the data protection officer sets the periods.
6. Delivery. On request the server arrives with the operating system, the NVIDIA driver, CUDA and a container runtime installed, assembled and tested before delivery.

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 NVIDIA L4 and the RTX PRO 4500 Server Edition as cards for servers you already run, after we confirm that the server's maker lists the card for that slot, or in AI servers built to order, delivered anywhere in 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 the speech and language models, access control and the logs. We do not supply care documentation software; the models connect to the product you use through the interface it offers.

FAQ