Stokoe Architecture

Stokoe is an accessibility architecture for real-time systems.

It defines how accessibility-relevant information — such as captions, speaker changes, and interpretive signals — is captured, structured, transported, and made observable in real time.

Captioning is the first concrete application of this architecture, not its boundary.

Architectural Intent

Most accessibility systems treat captions and related outputs as a user-interface feature or a side effect of model behavior. This makes accessibility difficult to inspect, difficult to audit, and difficult to reason about when failures occur.

Stokoe addresses this by treating accessibility as infrastructure.

Accessibility data is produced and transported explicitly, independent of any specific model, user interface, or presentation surface.

Core Properties

Stokoe is built around a small number of non-negotiable properties:

  • Accessibility is transport-layer data Captions and interpretive outputs are first-class events, not inferred UI state.

  • Append-only semantics Events are emitted, not rewritten. History is preserved.

  • Explicit timing Each event carries clear temporal meaning, enabling latency and ordering to be observed.

  • Model independence No accessibility guarantee depends on a specific speech or interpretation model.

  • Presentation decoupling Correctness is independent of how information is displayed.

System Overview

Stokoe is organized as a layered, event-driven system. Each layer has a narrowly defined responsibility and clear limits.

Layered Architecture

1. Media Ingress

Purpose: Capture raw media faithfully.

This layer handles live and recorded media sources and preserves signal fidelity and timing accuracy. It performs no semantic interpretation.

Responsibilities:

  • Media capture
  • Timing preservation
  • Continuous buffering

Non-responsibilities:

  • Speech detection
  • Interpretation
  • Captioning logic

2. Signal Analysis

Purpose: Extract observable facts from media.

This layer derives low-level signals such as energy, timing markers, overlap, and silence. These signals are descriptive, not interpretive.

Responsibilities:

  • Audio feature extraction
  • Temporal markers
  • Non-speech event detection

Non-responsibilities:

  • Text generation
  • Speaker identity
  • Finality decisions

3. Interpretation Engines

Purpose: Generate human-meaningful interpretations.

Examples include speech-to-text, diarization, translation, or audio description. Engines operate independently and emit structured events.

Key property: Interim and final outputs coexist. “Final” is metadata, not mutation.

4. Session Core and Event Transport

Purpose: Make accessibility data reliable and inspectable.

The session core orders events, buffers recent history, supports late joiners, and enables replay.

This layer establishes the accessibility timeline as the source of truth.

Responsibilities:

  • Event ordering
  • Buffering and replay
  • Session lifecycle management

5. Presentation Surfaces

Purpose: Deliver accessibility information without constraining form.

Because accessibility data is transported cleanly, it may be rendered as traditional captions, ambient displays, spatial overlays, or external data feeds.

Presentation does not affect correctness.

Observability and Failure

Stokoe is designed so that failures are visible, not hidden.

Latency, omissions, reordering, and dropped events are observable properties of the system. The architecture does not attempt to conceal or auto-correct these conditions.

This enables inspection, debugging, and audit without reliance on trust in model behavior.

Scope

Stokoe defines:

  • Accessibility signal transport
  • Event structure and semantics
  • Session and replay behavior
  • Architectural boundaries between media, models, and presentation

Stokoe does not define:

  • User interface design
  • Model training or quality metrics
  • Product-specific workflows
  • Policy or regulatory interpretation

Status

Stokoe is an evolving reference architecture.

It is intentionally conservative in design, emphasizing clarity, stability, and inspectability over novelty or optimization. Accessibility infrastructure must remain reliable as technologies change.