ESPnet3 Demo

This section documents the developer-facing demo subsystem used by pack_demo and upload_demo.

Important

The current demo implementation lives under espnet3/publication/demo.

What a demo is

An ESPnet3 demo is a packed inference app built from:

• demo.yaml
• a packed model directory or remote model tag
• a recipe-local Gradio launcher such as egs3/TEMPLATE/asr/src/app.py
• optional extra files copied into the bundle

Demo config

demo.yaml tells pack_demo where to load the model from, which app.py to copy, and which Gradio inputs and outputs to expose.

demo.yamleditable

model: dir_or_tag: exp/training_asr/model_pack trust_user_code: true ui: app_script: src/app.py inputs: - type: audio key: speech label: "Input speech" outputs: - type: text key: transcription label: "Transcription" upload_demo: hf_repo: your-org/your-demo

Gradio previewlive

Gradio Demo

Input speech

🎤

Drop audio or record

Transcription

Output will appear here

Built-in types: only audio and text are registered by default in UIAssetRegistry. Any other value in type: will cause a KeyError at runtime — unless you register a custom UIAsset in your app.py first.

Keep the config small at first:

Key	Purpose
model.dir_or_tag	Local packed model directory or remote pretrained tag
model.trust_user_code	Allow bundled recipe code when the model requires it
ui.app_script	Recipe-local Gradio launcher copied into the demo bundle
ui.inputs	Gradio input components mapped to model input keys
ui.outputs	Gradio output components mapped from model output keys
upload_demo.hf_repo	Hugging Face Space or repository used by upload

The built-in UI registry currently supports audio and text. Register a custom UI asset in app.py before using another type.

Main runtime idea

The demo runtime does not build provider or runner objects directly. It loads an InferenceModel, maps Gradio inputs to model inputs, and maps model outputs back to Gradio outputs.

Main modules

Module	Role
espnet3/publication/demo/packing.py	Pack and upload demo directories
espnet3/publication/demo/session.py	Load demo config, model, and UI specs into a runtime session
espnet3/publication/demo/assets.py	Built-in UI asset registry and Gradio component builders
espnet3/publication/inference_model.py	Packed-model inference runtime used by demos
egs3/TEMPLATE/asr/src/app.py	Default recipe-local Gradio launcher

Typical flow

1. pack_demo resolves and rewrites demo.yaml
2. pack_demo copies app.py, description files, and pack.include assets
3. a packed demo is launched with python app.py
4. app.py loads a DemoSession
5. the session loads InferenceModel
6. UI inputs are mapped to model inputs
7. model outputs are mapped back to UI outputs

What to read next

Pack Pipeline

See how `pack_demo` writes the bundle.

Runtime

See how demo sessions load models and run inference.

UI Definition

See how built-in demo UI assets are defined.