title: Character and audio description: Understand AIRI cards, display-model imports, rendering requirements, transcription, and speech as separate systems.

Character and audio

Character identity, visible artwork, and audio are separate resources. A successful configuration in one system does not prove the other systems work.

See workflows for procedures and providers-and-models for inference requirements.

An AIRI card is not model weights

The AIRI card store combines character information with module selections. Its prompt assembly includes a system prompt, description, personality, scenario, and an optional artistry instruction.

That source also excludes position-sensitive CCv3 fields until message assembly owns their ordering and role semantics. Importing a card does not establish complete CCv3 behavior.

The card store references Consciousness, Speech, Vision, Artistry, and display settings. A character can therefore select several resources without containing those resources itself.

Display assets and renderers

The display-model store distinguishes file-backed and URL-backed assets. Its format identifiers include Live2D, VRM, Spine, Tachie, and MMD-related imports.

The renderer packages have different responsibilities:

Asset family	Renderer package	Main requirement
Live2D	`packages/stage-ui-live2d`	Compatible artwork and the Live2D runtime
VRM	`packages/stage-ui-three`	A compatible VRM asset and graphics support
MMD	`packages/stage-ui-mmd`	Compatible PMX or PMD assets, graphics support, and WASM for physics
Spine	`packages/stage-ui-spine`	Compatible Spine assets and the corresponding runtime
Tachie	`packages/stage-ui-tachie`	Complete static emotion images, including a required neutral image

The format list is a source inventory, not a promise of feature parity across applications, devices, or arbitrary asset files.

Tachie is not a layered-sprite or rigged-animation format. Its guide describes one complete image per emotion, with matching image dimensions.

The shared model settings route mounts ModelSettings. Display-model imports use localforage for local persistence and can include preview images.

MMD preview generation is best-effort. The import path can retain a model after preview generation fails. A file entry or thumbnail is not proof of a working stage renderer.

flowchart TD
    Card[AIRI card] --> Prompt[Character context]
    Prompt --> Generation[Generation provider]
    Asset[Display-model file] --> Import[Import and local storage]
    Import --> Renderer[Selected renderer]
    Renderer --> Visible[Visible character]
    Mic[Microphone] --> Hearing[Transcription provider]
    Hearing --> Text[Text input]
    Text --> Generation
    Generation --> Reply[Text response]
    Reply --> Speech[Speech provider]
    Speech --> Playback[Audio output]

Hearing and speech are independent

Hearing owns transcription selection and audio-input behavior. Its source accepts different provider capabilities rather than one universal transcription path.

Speech owns the speech provider, model, voice, pitch, rate, and supported markup behavior. The initial provider is speech-noop.

packages/pipelines-audio owns reusable playback, text-chunking, and transcript-buffering policies. Raw audio utilities belong in packages/audio. Application lifecycle and credentials remain outside the pipeline package.

Browser speech recognition depends on browser capability. Apple Speech depends on a macOS Electron runtime. Kokoro offers WASM and WebGPU variants, with requirements specific to each model.

A headless page load does not prove microphone permission, native speech availability, audible output, or acceptable rendering performance.

Preserve assets and respect permissions

Keep original character files outside browser storage. Local persistence is not a backup or a cloud-sync guarantee.

Use assets whose licenses permit the intended use. The repository’s code license does not establish permission to redistribute every imported model, voice, or third-party SDK.

The Spine package guide requires a Spine Editor license for developers who integrate its runtime. It distinguishes developers from end users.

For a real acceptance check, inspect the selected asset on the intended application surface. Then inspect transcription and audible playback separately.

Source context

See provenance for the upstream pin and evidence labels.

packages/stage-ui/src/stores/modules/airi-card.ts, lines 14–42: character dependencies and prompt fields.
packages/stage-ui/src/stores/display-models.ts, lines 8–48: file, URL, and format identities.
packages/stage-ui/src/stores/display-models.ts, lines 117–166: imports, previews, and persistence.
packages/stage-pages/src/pages/settings/models/index.vue, lines 1–37: shared model settings surface.
packages/stage-ui-mmd/README.md, lines 29–73: renderer integration and WebGL/WASM boundary.
packages/stage-ui-tachie/README.md, lines 3–38: complete emotion images and archive requirements.
packages/stage-ui-spine/README.md, lines 5–11: developer license prerequisite.
packages/pipelines-audio/README.md, lines 1–15: reusable audio policies.

Navigation: index · modules · providers-and-models

Project AIRI

Explorer