WebRTC + AI for Accessibility: Live Captions and Sign-Language Overlays in 2026

Accessibility is not a feature flag. In 2026 the W3C RTC Accessibility User Requirements (RAUR) make it explicit: real-time captions belong next to the speaker, sign-language interpretation belongs on a parallel video track, and live transcripts must survive the call. WebRTC plus on-device AI is finally good enough to ship all three by default.

Why this matters

Every regulated vertical now has an accessibility surface. Healthcare intake calls, real-estate showings, behavioral-health check-ins, and legal consultations all run through WebRTC, and all of them are subject to ADA, Section 508, and the EU Accessibility Act (which became enforceable June 2025). The W3C RAUR draft updated in 2026 lists ten distinct accessibility needs for real-time communication — and only two of them (echo cancellation, audio routing) are solved by classic WebRTC. The remaining eight — live captions, alternate-format transmission for sign language, real-time text (RTT), speaker-position cues, status polling, transcript export, captioner-dial-in, and configurable caption rendering — need an AI layer.

The legal stakes also rose: 11,452 ADA web accessibility lawsuits were filed against US businesses in 2025, and a 2026 DOJ ruling clarified that real-time-communication products fall under the ADA Title III "place of public accommodation" definition when they front a regulated service. Translation: if your AI voice agent does not caption, you are exposed.

Architecture

```mermaid flowchart LR Browser[Caller Browser] -- DTLS-SRTP audio --> Gateway[Pion Go gateway 1.23] Gateway -- NATS --> Caption[Realtime ASR Service] Caption -- WebSocket --> CaptionUI[Caption Overlay next to speaker] Browser -- WebRTC video track 2 --> SignLang[Sign Language Interpreter Track] SignLang -- SFU forward --> Viewer[Viewer Browser] Caption -- transcript --> Audit[(115+ table audit)] ```

CallSphere implementation

CallSphere ships accessible WebRTC across all six verticals — real estate, healthcare, behavioral health, legal, salon, and insurance — using the same gateway-and-pod pattern that powers the rest of the platform:

• Real Estate (OneRoof) runs WebRTC with a Pion Go gateway 1.23, bridging to NATS and a 6-container agent pod (CRM, MLS, calendar, SMS, audit, transcript). The caption overlay is a separate React component that subscribes to the transcript NATS subject and renders next to the active speaker. See /industries/real-estate.
• /demo browser path uses the same WebRTC pipeline directly in Safari/Chrome — no native SDK, no install. The accessible captions render in the same overlay used in production. Try it at /demo.
• HIPAA + SOC 2 controls mean the live transcript is signed, hashed, and dropped into one of 115+ database tables that survive the call. Captions and transcripts are first-class compliance artifacts, not log spam.

CallSphere's stack — 37 agents, 90+ tools, 6 verticals — uses a single accessibility layer across them all. Pricing remains $149/$499/$1499 with a 14-day /trial; affiliates earn 22% — see /affiliate.

Build steps with code

```typescript // 1. Add a parallel data channel for caption events const pc = new RTCPeerConnection({ iceServers }); const captions = pc.createDataChannel("captions", { ordered: true });

// 2. Subscribe to ASR over WebSocket; forward to data channel + DOM const asr = new WebSocket("wss://asr.callsphere.ai/stream"); asr.onmessage = (ev) => { const { speakerId, text, isFinal } = JSON.parse(ev.data); captions.send(JSON.stringify({ speakerId, text, isFinal })); renderCaption(speakerId, text, isFinal); };

// 3. Render captions ANCHORED to the active speaker tile (RAUR requirement) function renderCaption(speakerId: string, text: string, isFinal: boolean) { const tile = document.querySelector([data-speaker="${speakerId}"]); if (!tile) return; let overlay = tile.querySelector(".caption-overlay"); if (!overlay) { overlay = document.createElement("div"); overlay.className = "caption-overlay"; overlay.setAttribute("aria-live", "polite"); overlay.setAttribute("role", "status"); tile.appendChild(overlay); } overlay.textContent = text; overlay.classList.toggle("interim", !isFinal); }

// 4. Add a second video transceiver for the sign-language track pc.addTransceiver("video", { direction: "recvonly", streams: [], }); // SFU forwards interpreter video as a parallel media stream ```

Pitfalls

• Burning captions into the video stream — destroys configurability (font, size, color, position) and violates RAUR 7.2. Always render captions as DOM overlays.
• Routing sign-language video through a "shared screen" — degrades to 5 fps and is not a substitute for a real video track. Use a parallel transceiver or a second peer connection.
• Letting the transcript die with the call — ADA and most consent decrees require persistent transcripts. Stream to a durable store (CallSphere uses one of its 115+ audit tables) before the call ends.
• Forgetting RTT — Real-Time Text (RFC 4103 / T.140) is required by the FCC for some call types and is the only accessible path for many DeafBlind users. WebRTC data channels handle this natively if you enforce ordered + reliable.
• Captioning interim only — provide both interim and final captions; users with cognitive disabilities rely on the stability of finals.

FAQ

Do I need a human captioner? For broadcast and legal contexts, yes (CART). For routine business calls, AI captions at 95%+ WER on clean audio meet ADA reasonable-accommodation standards.

How do I caption multiple languages? Run language ID on the first 2 seconds, then route to the matching ASR. CallSphere's pipeline switches per-utterance with ~120 ms overhead.

Where do I source sign-language interpreters? For scheduled calls, integrate with VRI providers (Sorenson, ZP, etc.) over SIP or WebRTC; for ad hoc, queue against an interpreter pool.

Is on-device captioning possible? Yes — Chrome's Live Caption API and on-device whisper.cpp can both run for privacy-sensitive calls; quality is lower than server-side Whisper-large-v3.

Do captions count as PHI under HIPAA? Yes when the underlying call is PHI. Captions inherit the classification of the audio they describe.

Sources

• https://w3c.github.io/raur/
• https://www.w3.org/WAI/APA/wiki/Accessible_RTC_Use_Cases
• https://www.w3.org/WAI/media/av/captions/
• https://swarmify.com/blog/video-accessibility-captions-wcag/
• https://verbit.ai/captioning/communication-access-real-time-translation-considerations-requirements/

Try the accessible WebRTC pipeline live at /demo, browse plans at /pricing, or start a /trial.