SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsvoicestudio

voicestudio

Agents & models

View repository

Updated 10-29 05:36

About this skill

When to use this skill

Use this skill for source-grounded VoiceStudio fit checks, offline repository preflight, approval-gated Electron or backend setup handoff, local voice cloning, text-described voice design, speech generation, transcription, and ASR → translation → TTS → mux dubbing workflows. Treat model weights, tokens, generated files, and remote providers as separate boundaries from the AGPL application.

Instructions

  1. Run the offline gate first. From the repository root run python3 .agent-skills/voicestudio/scripts/preflight.py --json; from an installed skill root run python3 "$SKILLS_ROOT/voicestudio/scripts/preflight.py" --json. It is offline-only and never launches VoiceStudio, invokes upstream tools, reads secrets, downloads weights, or calls a network.
  2. For a checkout, use --repo PATH --expect-commit 28bf2c7a83b8a64836297340c52a45d716684aed; report missing markers, tool availability, and the fact that runtime_verified remains false.
  3. Select the maintained Electron desktop path. Source setup is bun install, bun run setup:api, then bun run dev; the local FastAPI backend defaults to loopback port 3900, with web 3901 and Electron development 3902. Docker is a separate headless backend and binds published container traffic.
  4. Before any install, launch, server/port bind, credential entry, model or weight download, remote API call, analytics enablement, publishing, or destructive uninstall/data deletion, state the exact action, paths, network endpoints, model size/license implications, and obtain approval. Do not answer first-run analytics consent for the user.
  5. For voice cloning, require documented consent from the speaker for a real person's voice. Refuse impersonation, fraud, unauthorized identity capture, or deceptive disclosure; prefer a consent-verified profile or a designed voice.
  6. For voice design, use text attributes with an engine that supports design; OmniVoice uses one attribute per category, while some engines accept free-form descriptions. Do not claim that reference-only engines can design voices.
  7. For dubbing, preserve the local source and output paths, identify ASR/translation/TTS choices, and describe the ASR → translate → synthesize → mux stages. Verify output existence and playback only after an approved runtime run.
  8. Troubleshoot from /health, /openapi.json, /models, /models/install/status, backend logs, device/ffmpeg checks, and the pinned docs. Distinguish source evidence from an unverified live result; never call a paid or remote provider implicitly.

Examples

Use VoiceStudio for a local dub: run the offline preflight, report whether Python, Bun, Node, ffmpeg, and git are discoverable, then ask before installing the required Hugging Face models. Clone only a speaker with documented consent; otherwise use Voice Design. Keep the backend loopback-only and route timeline assembly to opencut.

Best practices

  • Pin and report the upstream commit and distinguish AGPL application code from model, tokenizer, voice, and generated-output terms.
  • Keep HF caches and VoiceStudio data paths explicit; default data locations are platform-specific under OmniVoice and generated audio is under outputs.
  • Treat 3900 as the default backend endpoint and restrict any non-loopback exposure to a trusted network with HTTPS/WSS and explicit authentication.
  • Label model download size, gated-access requirements, hardware mode, and license review before approval.
  • Never present a preflight or source inspection as proof that a model loaded, a server is healthy, or audio quality is acceptable.
  • Prefer loopback host defaults for local development and state explicitly when Docker changes the bind boundary.
  • Keep speaker consent records outside generated audio paths and avoid embedding personal identifiers in filenames.
  • Separate ASR transcript review from translation review; preserve subtitle timing before muxing.
  • Treat /models catalog metadata as a planning hint, not proof that weights are installed or usable on the host.
  • If a model is gated, explain token scope and cache location without printing or copying the token.
  • Route hosted TTS, timeline editing, and hosted video generation to their dedicated catalog skills rather than adding provider code here.
  • For failures, capture the first actionable error and the exact endpoint or engine selected before changing settings.
  • Keep generated samples attributable to the chosen engine and disclose synthetic voice use where the surrounding product requires it.

References