voize

Proprietary domain specific speech recognition and an on device large language model, both built in house by founders with systems engineering backgrounds. The technical choice to run inference locally rather than calling a hosted model is itself evidence that model work is the company rather than a wrapper around someone else's API.

The review gate is stated on the vendor's United States site: caregivers review, edit and approve each charting entry on the device before sign off, every data entry must be approved or edited by the caregiver, and the system cleans up grammar and maps speech to fixed data fields without adding or assuming information. What is missing is any confidence threshold, low confidence behavior, or edit and rejection rate.

The gate matters more here than for an office visit scribe because of what this product writes into. Medication administration records and vital signs are legally significant clinical documents where an error is a patient safety event rather than a wording problem. Fluid intake feeds clinical decisions in a frail population.

There are two things to confirm rather than one. The vendor presents recognizing residents and assigning documentation to the right person as a feature. In a facility with many residents, a misassigned medication entry affects two people at once: the resident whose record gains an entry that did not happen and the resident whose record loses one that did. Nothing published says what the caregiver sees about attribution before approving, per documentation category.

The speech input compounds it. The system handles colloquial and non native speech and converts it into formal documentation, which is a real capability and also means the committed entry may differ in wording from what was said.

The architecture is specific. voize states what the model is, an on device large language model, where it runs, locally on smartphones, and what it was built to handle, namely medical shorthand, regional accents and multilingual speech patterns, and it says its models are continuously learning from actual usage. That is a description that can be tested rather than a set of adjectives.

No base model or version is named, nothing says how on device models are updated when they learn, and the headline accuracy claim of 99.2 percent is published without any methodology, denominator, reference standard or breakdown.

The architecture answers the central question structurally. The model is described as an on device large language model running locally on a smartphone and working without a constant internet connection, so speech is processed where it is spoken and no third party model provider sits in the inference path. A buyer can test that claim by using the application offline. The vendor names AWS as its underlying cloud provider for United States backend services such as device and user management, under an executed business associate agreement.

Two gaps remain. The model is described by type and location but not by identity, so no base model, provider or version is named, and an on device model is very often a third party open weight model adapted for the task. And the chain does not end at the device. Structured documentation synchronizes into the care record system, and the vendor says its models are continuously learning from actual usage, while its privacy policy allows de identified health data to be combined and provided to third party collaborators and partners. Whatever carries usage data back for learning, and whoever receives de identified data, are parties in the chain that are not named.

An academic institution was involved. The vendor reports a 27 percent reduction in documentation time from a study conducted with the Charité in Berlin, and names customer outcomes alongside it: 33 minutes saved per caregiver per day at Caritas Stuttgart, an 82 percent reduction in overtime at one operator, and a 15 percent reduction in staff turnover at another. Deployment scale is stated as more than 180,000 nurses and caregivers across more than 2,000 facilities. The figures are the vendor's report of the Charité study rather than the study itself: no peer reviewed publication, design or cohort is published.

Local processing is a real privacy property: the model runs on the caregiver's device and works without a constant internet connection, so speech is processed where it is spoken rather than sent to a vendor for inference, and a buyer can verify that by testing the app offline. The vendor also commits in its privacy policy to using protected health information only as its business associate agreements and HIPAA permit, and not for its own commercial purposes.

None of retention, training use or de identification is stated specifically enough to be contradicted. No retention period is stated for audio, for anything held on the device, or for data on the backend. On training, the vendor says its models are continuously learning from actual usage, without saying what usage data leaves the device, whether it is de identified first, or whether a customer can opt out. And the privacy policy allows de identified health data to be combined with other data and provided to third party collaborators and partners, with no de identification method described.

On safety engineering, nothing is published beyond the caregiver's approval of each entry and the statement that the system maps speech to fixed fields without adding or assuming information. No guardrail, hallucination mitigation or safety event process is described.

Its boundary matters: inference is local, but structured documentation still synchronizes into the care record system, so the protection covers the audio and the processing rather than the finished record.

Business associate status is stated first party. The vendor's United States site says voize acts as a business associate when serving United States healthcare customers and signs a business associate agreement with every covered entity before processing any protected health information, and its privacy policy, updated June 2026, says it uses and discloses protected health information only as the agreement, the services agreement and HIPAA permit, and not for its own commercial purposes. It describes an internal HIPAA program covering the privacy, security and breach notification rules, and names AWS as its underlying cloud provider with an executed business associate agreement.

The architecture changes the shape of the relationship. In the standard United States deployment, speech recognition and documentation run on the caregiver's device and structured documentation goes directly into the customer's record system, with backend services such as device and user management running in regions the vendor describes as aligned with HIPAA expectations. The vendor holds less than a cloud processing competitor, but the application, the synchronization path, the backend and any learning from usage remain its responsibility.

The agreement itself is not published, and no subprocessor list beyond the cloud provider is public, though the trust center may list more.

No named or dated attestation, SOC 2 report or ISO 27001 certificate is published. A trust center hosted on a compliance platform is linked from the site footer, and it may hold the reports. The footer also links a security page for a bug bounty program, which invites outside testing of the product.

The United States FAQ describes an internal HIPAA program with access controls, audit logging, encryption, workforce training and incident response, and says facilities manage the handsets through mobile device management so they cannot be used for other purposes or taken off site. Those are described controls rather than an examination.

The architecture changes what an attestation would need to cover. Where speech recognition and documentation run on the caregiver's device, the vendor's estate is smaller than a cloud processing competitor's, but the mobile application, the synchronization path, the model distribution mechanism and the backend on AWS all remain the vendor's responsibility. The scale of the customer base and expansion into the United States raise the expectation.

No clearance claimed and none required, and no device pathway attaches in either market served.

One question is pressing, because the United States expansion has happened and the documentation scope is wider than a narrative note.

The product generates entries across medication logs, vital signs, fluid intake, wound documentation with photographs, care reports and emergency documentation. Several of those sit further from narrative documentation than an office visit note does. A medication administration record is a legally significant document in its own right, vital signs feed clinical judgment directly, and emergency documentation accompanies an action rather than describing a completed encounter. None of that makes the product a regulated device today. It does mean the classification question is open, particularly as scope widens, and the ambient scribe answer does not carry across to a product that writes structured clinical values.

What actually governs in the home market is long term care documentation requirement and inspection rather than device regulation, where documentation quality is assessed as part of the care quality regime and failures surface as findings against the facility rather than as device incidents. The facility remains accountable for its records whatever produced them.

For the United States market the equivalent frame is the conditions of participation and survey process applying to skilled nursing facilities, which likewise place the obligation on the operator. Nothing published says how the product supports evidencing that obligation, and the value of accurate contemporaneous documentation in that setting is partly defensive.

Treats accent and language variation as a design requirement rather than a disclaimer. voize states its model was built to understand regional accents and multilingual speech patterns alongside medical shorthand, and its United States site says it handles everything from Southern accents to non native English speakers, with grammar correction to support non native staff.

That matters for this user population: care work in Germany, Austria and United States skilled nursing is disproportionately performed by multilingual and migrant staff, so a speech model that degrades on non native accents would fail hardest on the workforce it is sold to support. No performance breakdown by accent, language or speaker group is published, so the design intent is stated rather than demonstrated.

The vendor's contribution here is architecture rather than a commitment.

The review step is stated: every entry must be approved or edited by the caregiver on the device before it is committed, and the vendor says the system cleans up grammar and maps speech to fixed fields without adding or assuming information. Local processing on the caregiver's device also removes a class of exposure structurally rather than by promise. What neither does is say anything about what happens when the documentation is wrong.

The published accuracy figure is the weak point. A stated 99.2 percent arrives with no methodology, denominator, reference standard or breakdown, and a number given to one decimal place invites precisely the question of what was measured. An unfalsifiable figure is not a commitment however precise it looks, because there is no way for a customer to establish it was not met. No warranty, indemnity or remediation obligation is published.

The statutory position is real. Operating under European data protection law, the person recorded holds enforceable statutory rights over their own personal data, including rectification of inaccurate records, with a supervisory authority behind them and obligations that reach processors as well as the customer. Those rights exist whether or not the vendor offers a route, and this vendor has published none of its own.

Integrates into existing care documentation systems through the voize Gateway, which pushes confirmed entries into the facility record and synchronizes client information back to the device, with Vivendi PD named as a supported German system. The United States site says voize integrates bidirectionally with all major record systems used in skilled nursing and long term care, without naming them. These are long term care documentation platforms rather than hospital electronic health records, so integration breadth should be judged against that segment. No United States integration is named.

The A rested on the answer to where the data goes being that it does not go. That is true of inference, which runs on the device in the caregiver's hand and keeps working offline, a practical advantage in care homes with poor connectivity. It is not true of the whole product. Structured documentation synchronizes into the customer's record system; in the United States, backend services such as device and user management run on AWS in regions the vendor describes only as aligned with HIPAA expectations; the privacy policy places voize servers in the United States; and the vendor says its models learn continuously from actual usage.

So residency and the deployment model are stated, and the processing path beyond the device is left open: no region is named, nothing says what usage data travels for learning or where, and tenant isolation is not described.

No rate card is published. Sold to care facility and hospital operators rather than to individual clinicians, so pricing is negotiated at the organization level. One market signal rather than a price: some care providers now cite voize in job listings as a recruitment advantage, which suggests operators are valuing it partly as a retention instrument rather than purely as a documentation cost line.

Covers the nursing and caregiving workforce rather than the physician office visit: documentation, shift handovers, wound documentation with imaging, vital signs, fluid intake, movement logs, medication administration, scheduling and emergency alerts, across care homes, skilled nursing and hospitals. Shift handover in particular is a high risk communication event with no equivalent in outpatient documentation. Geographic reach now spans Germany and Austria, the United States skilled nursing and long term care market, and localized sites for the United Kingdom and France.