Voice Forge

VoiceForge

Use VoiceForge as the owner of a person's private communication corpus and derived style profile. Other tools may consume its portable export; they should not read or copy the raw corpus unless the user explicitly requests it.

Discover the installed interface

Run forge --help and the relevant subcommand's --help before acting. The CLI is evolving, so prefer the installed interface over remembered flags.

Core workflow

1. Check the environment with forge doctor.
2. Ingest only material the user has authorized:
   • forge ingest <path> for paired audio and transcripts.
   • forge ingest-text <path> for writing.
   • forge ingest-code <path> for coding-style signals.
   • forge ingest-video <path> or forge ingest-photo <path> when relevant.
   • forge ingest-bulk ... only after confirming every source directory.
3. Inspect corpus state with forge stats or forge corpus --help.
4. Run forge analyze to derive a style profile, or forge refresh to update an existing profile when the corpus has materially changed.
5. Review with forge profile --brief or forge profile.
6. Export the interoperability contract when another tool needs style guidance: forge profile --output ./voice-profile.json.

The portable file is versioned and conforms to schemas/voice-profile.schema.json. Share that export instead of ~/.forge/profile/style.json or raw corpus entries.

Using the profile

• Treat extracted patterns as evidence, not permanent identity.
• Preserve the user's meaning; never manufacture opinions to match a style.
• Confirm sensitive or surprising conclusions before relying on them.
• Use forge write <topic> only when the user wants style-assisted drafting.
• Use forge speak <text> only when the user requests audio generation.
• Inspect forge backends before selecting a TTS backend.

Talk Once integration

Export a portable profile, then reference it from a Talk Once content package:

{
  "voiceProfile": {
    "provider": "voiceforge",
    "path": "./private/voice-profile.json",
    "version": "1.0"
  }
}

VoiceForge owns the corpus and profile. Talk Once owns interviews, canonical content, approvals, and derivatives.

Safety

• Do not ingest broad directories without reviewing their contents and scope.
• Keep recordings, transcripts, derived profiles, and cloned voices private by default.
• Never publish, upload, or send generated text or audio without approval.
• Do not clone another person's voice without their explicit authorization.
• Report the profile path, export path, and commands run without exposing corpus contents in logs or summaries.