Skip to main content
Metadata

Overview

The Voice extension defines provider-neutral signals used while an Aion Distribution is conducting a live voice session. The current version defines one signal: a short backchannel utterance that may be spoken while the canonical response is still being produced. A backchannel does not replace the canonical response, complete the task, or change the task status. Key aliases used below:
  • VoiceURI: https://docs.aion.to/a2a/extensions/aion/distribution/voice/1.0.0
  • BackchannelArtifactId: aion:voice-backchannel

Agent Card Declaration

Agents that may emit Voice extension artifacts declare the extension in their Agent Card:
The declaration advertises possible output. It does not mean that every response contains a backchannel artifact.

Artifacts

Backchannel

A backchannel is a brief, non-canonical utterance such as “One moment while I check that.” It is delivered as a TaskArtifactUpdateEvent so it can be synthesized without being mistaken for the agent’s final answer.

Payloads

BackchannelPayload

Schema URI: https://docs.aion.to/a2a/extensions/aion/distribution/voice/1.0.0#BackchannelPayload The payload is attached at artifact.metadata[VoiceURI]. The event must provide speakable text either in text or in one or more non-empty artifact TextPart values. When text is absent or blank, consumers join trimmed, non-empty text parts in their original order using newline boundaries.

Example

This example shows the complete streaming envelope emitted by the current Aion backchannel producer:
The current producer emits each backchannel as a complete, non-appending artifact update with append set to false and lastChunk set to true.

Processing Rules

Producers

Producers must include both the extension URI marker and its typed metadata on the artifact. The update may duplicate the payload in update-level metadata for envelope-level routing, as Aion’s current producer does. The payload in artifact.metadata[VoiceURI] remains canonical.

Consumers

Consumers recognize a backchannel only when the reserved artifact ID and a valid payload are both present. They ignore Voice metadata attached to any other artifact ID. A Voice distribution handles a recognized backchannel independently from canonical response text:
  1. Resolve the text from the payload or artifact text parts.
  2. Submit the backchannel for immediate speech synthesis.
  3. Preserve its position relative to canonical utterances already observed, without merging it with canonical response text.
  4. Continue processing the canonical A2A response normally.
Backchannel artifacts do not alter the task lifecycle. Producers must still return the canonical response through the ordinary A2A response stream.