Workflows saved in ComfyUI are read again for each picture, so changes saved there apply at once. The workflow's own style text stays in place; the default is set in the Settings tab.
Audio is sent to this server for transcription (Whisper); nothing leaves your network. With “Talk to interrupt” off, listening pauses while it thinks and speaks. In push-to-talk mode Space talks unless a text field, checkbox or menu has the focus.
Saving under a new name makes a copy. The CLI uses these too: narrate --persona NAME.
Narrating:
Answering questions:
Saved on the server, so they apply in every browser and to the command line. A session keeps what it was started with; change that on the Live tab.
LLM model: the Live tab's choice for a session, else the persona's model, else this. Speech
steps: the header's Speech menu (this browser), else the voice's own default (Voices tab), else this, else the
server's --steps. Workflow: the Live tab's Picture settings, else this. Workflows saved in
ComfyUI are read again for every picture: save in ComfyUI and the next picture uses your changes.
One choice per category. Works strongly with no clone selected (voice design); on a cloned voice the effect is weak.
API: curl -X POST $HOST/api/speak -H 'content-type: application/json' -d '{"text":"hi","voice":"NAME","instruct":"female, low pitch"}' -o out.wav