Publishing Voice Presets

中文 | English

A DiceFrame voice-pack is a voice preset, not a TTS engine installer. It can distribute voice_id values for an OpenAI-compatible service or small, authorized GPT-SoVITS reference files. Upstream projects should install Kokoro, GPT-SoVITS, CUDA, Python, and base models; do not bundle those large dependencies again.

Confirm redistribution rights first

The publisher must be able to document the right to distribute every item in the package:

For voice IDs and documentation only, record the model, service, version, source URL, and license. Do not imply that DiceFrame owns or trained the voice.
For preview or reference audio, confirm the recording, performer voice, spoken text, and redistribution rights. Attribution alone is not permission.
For a cloned real-person voice, obtain explicit permission and avoid material that could enable impersonation, fraud, or privacy abuse.
consent: true is the author's confirmation that the voice descriptor and every referenced audio asset may be distributed. It is not a placeholder.

When rights are unclear, publish setup instructions that let users enter a voice ID on their own device instead of publishing a public voice pack.

Recommended repository layout

The public store requires one plugin per public GitHub repository, with the plugin at the repository root:

text

my-voice-presets/
├─ plugin.json
├─ config.schema.json
├─ README_CN.md
├─ README_EN.md
├─ LICENSE
├─ NOTICE.md
└─ voices/
   ├─ narrator.json
   └─ companion.json

OpenAI-compatible presets need JSON only. GPT-SoVITS packages must list every authorized WAV or preview file under contributes.voice_assets. Every voice needs a unique ID, display name, engine, language, license, and authorization confirmation.

Manifest example

json

{
  "schema_version": 1,
  "id": "my-voice-presets",
  "name": "My Voice Presets",
  "version": "1.0.0",
  "description": "Small voice-ID presets for an existing local TTS service.",
  "plugin_type": "voice-pack",
  "repository_url": "https://github.com/username/my-voice-presets",
  "config_schema": "config.schema.json",
  "permissions": ["plugin.config", "voice.assets"],
  "contributes": {"voices": ["voices/*.json"]},
  "docs": "README_EN.md"
}

One OpenAI-compatible voice:

json

{
  "schema_version": 1,
  "id": "narrator-zh",
  "name": "Chinese Narrator",
  "engine": "openai-compatible",
  "voice_id": "service-native-voice-id",
  "language": "zh-CN",
  "license": "Apache-2.0",
  "consent": true
}

license describes the material actually distributed by the preset. README and NOTICE should separately identify upstream licenses for models, servers, and voices. A project-level license must not be treated as automatic permission for unrelated third-party recordings.

Local validation

Run this from the DiceFrame repository:

powershell

python scripts\package_plugin.py C:\path\to\my-voice-presets --overwrite

Install the resulting .dfplugin under Manage → Plugins → Local Install, then verify:

Install, enable, disable, and uninstall all work.
Voice names, languages, and engines appear correctly.
Save and test produces audio while the matching local TTS service is running.
A stopped service or missing voice ID produces a readable error without affecting the game.
The package contains no model weights, .env, token, cookie, absolute local path, log, or user recording.

The .dfplugin is for local or offline distribution. The public store does not require authors to upload it.

Publish to the plugin store

Put the plugin in its own public GitHub repository and verify the root layout and repository_url.
Commit the files and create a Git tag matching plugin.json.version, such as v1.0.0.
Publish a non-draft, non-prerelease GitHub Release. No ZIP upload or SHA-256 is required.
Open the `diceframe/diceframe-plugins` Add plugin issue with the plugin ID and repository URL.
Fix automated findings and reply /recheck; after validation, wait for maintainer approval.

Ordinary authors do not log in to the DiceFrame Hub administration panel. Hub login is only for maintainers performing review, synchronization, and moderation; public submission is based on the GitHub Release and submission issue.

Later releases

When the ID, repository, and permissions stay unchanged, bump the semantic version and publish a new tag and stable Release. The store resolves the new Release to a fixed commit and updates after user confirmation. Permission expansion, repository replacement, ownership transfer, or a switch to process execution requires another review.

Kokoro example

plugins/examples/kokoro-zh-voice-presets distributes four Kokoro voice IDs and documentation only. It contains no audio, model, or server code. Kokoro-82M and Kokoro-FastAPI use Apache-2.0, while users start the compatible service separately. The example is suitable for learning, packaging, and local installation tests; DiceFrame does not guarantee upstream voice quality, identity, or long-term compatibility.

DiceFrame setup and user documentation

Publish voice presets