[object Object]

Voice Interface Considerations

Speech recognition struggles with diverse accents, speech differences, and disability-related variations. Word error rates on Whisper-large-v3 still run 2–4x higher for users with dysarthria (ALS, cerebral palsy, post-stroke) than for unaffected speakers. Stuttering breaks turn-detection — the model interprets a block as end-of-utterance and cuts in. Deafblind users cannot use voice at all.

Mitigations: enable per-user acoustic adaptation (Project Relate, Apple Personal Voice, Whisper fine-tuning on customer audio with consent), expose a “give me more time to finish” affordance, and always provide a parallel text path. Voice-only flows fail WCAG 2.2 SC 1.4.5 for any user who cannot or will not speak.

Chat Interface Considerations

Screen readers must parse AI responses correctly. Streaming responses break NVDA and JAWS unless the live region is throttled — emit one aria-live="polite" update per sentence, not per token. Set aria-busy="true" while the model is generating, and false on completion. Visual confidence indicators (“85% sure”) need text equivalents read by the screen reader.

<div role="log" aria-live="polite" aria-atomic="false" aria-busy="true">
  <p>The agent is thinking...</p>
</div>

Dynamic UI changes — quick reply chips appearing, forms expanding mid-conversation — must announce themselves. A keyboard user must be able to reach every interactive element with Tab, and focus must move predictably when the agent inserts new content.

Multimodal Support

Users who can’t see images need alt text on AI-generated images and charts; the model that produced the image must produce the description in the same call. Users who can’t hear audio need real-time transcripts on voice responses, plus speaker labels for multi-party calls. AI systems must be designed from the start for at least three input modes (voice, text, switch/keyboard) and three output modes (audio, text, visual).

EU AI Act Article 16 requires accessibility for high-risk AI systems, cross-referencing the European Accessibility Act (EAA), enforceable June 28, 2025. US obligations: ADA Title III for customer-facing systems, Section 508 for federal contractors, and WCAG 2.2 AA as the operative standard. The DOJ’s April 2024 ADA web rule sets a hard April 2026 compliance deadline for state and local government public-facing AI tools.

Penalties stack. EAA fines reach €1M per violation in some member states. Class actions under the ADA cost mid-market US companies $50K–$500K per settlement, plus remediation.

Implementation Sequence

  1. Audit existing AI surfaces with axe-core and a screen reader (NVDA + Safari/VoiceOver minimum).
  2. Add ARIA live regions and focus management to streaming UIs.
  3. Bake alt-text generation into every image-producing prompt.
  4. Run live user testing with disabled users — automated tools catch ~30% of issues.
  5. Document accessibility decisions in your AI Act technical file.

Accessibility invested in during design is cheap; retrofitted is 5–10x the cost and rarely as good.

[object Object]
Share