Reference audio
Voice cloning quality begins with the reference. Prefer a clean recording that contains one speaker, consistent delivery, and minimal room echo.
- Aim for roughly 6–30 seconds of useful speech. References below 3 seconds are rejected.
- Avoid music, overlapping voices, reverb, and aggressive noise reduction.
- Natural variation is useful; long silence is not.
- WAV input is the most predictable choice.
Add a voice
Register the reference under a memorable local name:
llmvoice voice add friday reference.wav
Mono audio is convenient but not mandatory. LLMVoice can inspect stereo input and normalize the working format where needed. A clean stereo file is better than a damaged mono conversion.
Manage voices
Inspect the registry before changing defaults or deleting files:
llmvoice voice list
llmvoice voice info friday
llmvoice voice remove friday
Voice names should be short, stable, and easy to type. Use configuration to select the one used when --voice is omitted.
Privacy
Reference audio and derived voice data remain local to your LLMVoice installation. LLMVoice does not upload voices or provide a cloud sync layer.