One voice turn, traced

In: microphone audio. Out: a plain string. That is the entire STT contract — nothing downstream knows or cares that this came from a voice rather than a keyboard.

// Web Speech: e.results[i][0].transcript
"direct flight to Lisbon next Friday, back Sunday night, under 300 euros"

Worth noticing: the demo's text box produces the identical string. Every stage after this is modality-blind, which is why the whole thing can be tested without a microphone.

In: the transcript plus conversation history. Out: one request. No tool loop, no framework — the interesting part is output_config.format.

{
  "model": "claude-opus-5",
  "max_tokens": 4000,
  "system": [{ "type": "text", "text": "<3617 chars — the show-or-say rules>",
              "cache_control": { "type": "ephemeral" } }],
  "messages": [{ "role": "user", "content": "direct flight to Lisbon…" }],
  "output_config": {
    "effort": "low",
    "format": { "type": "json_schema", "schema": /* compiled from the catalog, 5421 bytes */ }
  }
}

This is the enforcement. Each catalog component becomes an anyOf variant discriminated by a const. A component that isn't in the catalog isn't rejected after generation — it is ungeneratable.

{
  "type": "object",
  "description": "Rows the user compares and picks from…",
  "properties": {
    "id":         { "type": "string" },
    "component":  { "const": "OptionList" },      ← the discriminator
    "itemsPath":  { "type": "string" },
    "itemTitle":  { "type": "string" },
    "itemDetail": { "type": "string" },
    "actionName": { "type": "string" }
  },
  "required": ["id", "component", "itemsPath", "itemTitle", "itemDetail", "actionName"],
  "additionalProperties": false
}

Seven variants compiled from seven catalog components: Column, Heading, OptionList, SeatMap, PaymentConfirm, SensitiveList, Status.

One JSON object, streamed. say is first in the schema and therefore generated first — that ordering is what makes early speech possible.

{
  "say": "Four direct round-trips come in under three hundred — they're on the screen.",
  "components": [
    { "id": "root", "component": "Column", "children": ["h1", "list"] },
    { "id": "h1",   "component": "Heading", "text": "Lisbon, Fri–Sun" },
    { "id": "list", "component": "OptionList",
      "itemsPath": "/options", "itemTitle": "title",
      "itemDetail": "detail", "actionName": "pick_flight" }
  ],
  "data": {
    "options": [
      { "title": "TAP Portugal",
        "detail": "two hundred forty-six euros, out Friday seven twenty in the morning,
                   back Sunday nine forty at night" },
      { "title": "Ryanair",
        "detail": "one hundred eighty-nine euros, out Friday six fifteen in the evening,
                   back Sunday ten thirty at night" },
      { "title": "easyJet",  "detail": "two hundred twelve euros, …" },
      { "title": "British Airways", "detail": "two hundred eighty-four euros, …" }
    ],
    "sensitive": [],
    "note": ""
  }
}

Read the details again. “two hundred forty-six euros, out Friday seven twenty in the morning” — not €246 · 07:20. The prompt asks for values that read well aloud, because a renderer with no display speaks these strings verbatim. Nothing downstream reformats them.

The model's flat form is normalised into real A2UI. itemsPath becomes a binding; actionName becomes a full action with a context template. This is what both renderers actually receive.

{
  "version": "v1.0",
  "createSurface": {
    "surfaceId": "surface_1",
    "catalogId": "https://a2ui-voice.dev/catalogs/travel/v0.1/catalog.json",
    "dataModel": { "options": [ /* the four rows, verbatim from stage 3 */ ],
                   "sensitive": [], "note": "" },
    "components": [
      { "id": "root", "component": "Column", "children": ["h1", "list"] },
      { "id": "h1", "component": "Heading", "text": "Lisbon, Fri–Sun",
        "accessibility": { "label": "Lisbon, Fri–Sun" } },
      { "id": "list", "component": "OptionList",
        "items": { "path": "/options" },        ← itemsPath became a binding
        "itemTitle": "title", "itemDetail": "detail",
        "action": { "event": { "name": "pick_flight",
                              "context": { "index": "$index", "title": "$item.title" } } } }
    ]
  }
}

In: the message above. Out: not speech — a plan. Which gates are active, and the referent set: the things the user can now point at with words.

{
  "blockers": [],              // nothing requiresVisual → the turn can be spoken
  "surfaceStakes": "none",      // no readback needed to commit
  "silentComponents": [],      // nothing spokenSensitive
  "referents": [
    { "n": 1, "title": "TAP Portugal",    "action": {"event":{"name":"pick_flight","context":{"index":0,…}} },
    { "n": 2, "title": "Ryanair",         "action": {"event":{"name":"pick_flight","context":{"index":1,…}} },
    { "n": 3, "title": "easyJet",         "action": {"event":{"name":"pick_flight","context":{"index":2,…}} },
    { "n": 4, "title": "British Airways", "action": {"event":{"name":"pick_flight","context":{"index":3,…}} }
  ]
}

Each referent carries the action a tap would have fired. That is the mechanism behind the whole claim — the voice path doesn't construct its own event, it reuses the one already attached to the row.

In: a string. Out: audio. The renderer did not write this sentence — the agent did, at stage 3. In the ordinary case nothing is “converted to speech” at all.

// TTS input, verbatim
"Four direct round-trips come in under three hundred — they're on the screen."

// fired at t=2278ms, while the components were still generating

The renderer speaks only the mechanical extras: announcing a value that changed after the agent stopped (accessibility.live), reading back before a commit, confirming a pick. Everything else is the agent's own words.

In: a two-word transcript. Out: the same event a tap produces.

// STT output
"the second one"

// voice renderer resolution → status: "dispatched"
{
  "event": {
    "name": "pick_flight",
    "context": { "index": 1, "title": "Ryanair" }
  }
}

Byte-identical to a tap on row 2. The agent receives this and cannot tell which happened — which is what lets voice be added to an existing A2UI app without the agent changing at all.