What Fluent conforms to, and what it does not.

Why this says self-assessment on every page

A conformance report that claims Supports on every line, with no named gap, is not credible, and a procurement office reads dozens of them. The value of this document is the part where it says what is missing. It is a self-assessment in the VPAT 2.5 structure, not a third-party audit, and it is labelled that way throughout rather than in a footnote.

Fluent is itself an assistive technology, which creates an obligation most software does not carry: it has to be usable by people who cannot use a keyboard, cannot use a mouse, cannot see the screen, or cannot speak clearly. An accessibility defect here is not a compliance checkbox - it is the product failing at its purpose.

The gaps, up front

Voice recognition accuracy has not been formally evaluated for dysarthric or accented speech. That is the most significant gap for exactly the audience this product exists for, and it is the one gap engineering cannot close on its own - it needs a diverse speech corpus and real measurement, not code.

The interface chrome is English only. Foreign-language content inside a response is marked so a screen reader pronounces it correctly, but translating Fluent's own words has not started. Neither EN 301 549 nor any WCAG criterion requires it, so this is a reach gap rather than a conformance one, and it is listed because it still affects people.

There is no completed third-party accessibility audit and no SOC 2 Type II report. Both are open items and are stated here rather than left for an evaluator to discover.

The accessibility properties that are enforced in code

Several accessibility behaviours are treated as invariants and guarded by an automated check that fails the build, because each is a one-line deletion that would typecheck, lint clean, and read in a diff as a simplification.

Status announcements queue rather than overwrite, because a live region is only spoken when a screen reader notices it change - so a second message written before the first was observed loses the first outright. The automatic logoff announces itself before it fires, and the extension control refuses unless a warning is actually pending, so it cannot be used as a keep-alive. Failures reach the user as a sentence and a suggested next step rather than an error code, because relaying a code to somebody else assumes an ability the user may not have. The overlay stops being topmost when it has nothing to show, so it cannot hide the control somebody just tabbed to. A change of writing system inside an answer is marked so a screen reader does not read it with the wrong letter-to-sound rules.

• WCAG 4.1.3 - announcements queue; urgent messages jump the queue and are never dropped.
• WCAG 2.2.1 - the security timeout stays the administrator's, but it warns before it fires.
• WCAG 3.3.3 and Revised 508 302.9 - errors in words, with a next step.
• WCAG 2.4.11 - the always-on-top overlay yields focus visibility when idle.
• WCAG 3.1.2 - language changes marked by writing system, never guessed by a model.

How it is tested, and how often

Evaluation methods are manual keyboard testing, NVDA and Windows Narrator, automated axe-core scans, contrast measurement, and testing through the product's own voice control path. The axe-core and keyboard suite runs against the real application as a required check on every pull request, because a conformance report backed by a suite nobody runs is a claim rather than a control.

A parity test also runs one request through tool selection, risk assessment, the confirmation gate and speech in twenty languages, comparing each against English. A rule written only in English does not degrade outside English - it stops firing, silently, which removes capability from the user and hands an attacker a bypass costing one translation.

For a procurement or accessibility office

The full report is available on request in its complete VPAT 2.5 structure, criterion by criterion, along with the control matrix, the data map, and a CycloneDX software bill of materials published with each release. Requests, corrections and disagreements go to the accessibility address below, and a correction does not require a backlink or positive coverage.

If your evaluation requires a third-party audit rather than a self-assessment, Fluent does not currently meet that bar, and the right answer is to say so at the start of a procurement rather than at the end of one.

Questions

Keep reading

• Can voice control be used on a workstation that displays PHI?The interesting question is not what the software can do. It is what it refuses to send, and whether you can prove it afterwards.
• What does Fluent collect, and what leaves my computer?Fluent listens to your microphone and reads your screen. That makes this policy unusually important, so it is written to be read rather than to be defensible.
• Is voice control suitable for someone with ALS or MND?The honest version: Fluent is not an access system for ALS. It can be one layer of one, for as long as speech holds.

Sources

• WCAG 2.2 — The W3C recommendation this report is assessed against.
• Revised Section 508 standards — US Access Board, 36 CFR 1194 Appendix C.
• EN 301 549 v3.2.1 — The European accessibility requirements for ICT.
• Request the full report — accessibility@fluentforall.com - the complete VPAT 2.5 structure, criterion by criterion.
• Fluent clinical deployment guide — On-device inference, egress control and the audit chain.