Text-to-speech voices

Mastery HQ speaks with one of four engines, chosen in Settings → Voice → Text-to-speech. Turn spoken output on with Enable TTS; add Speak agent output aloud if you also want agent replies voiced, not just the assistant’s.

The four engines

Engine (exact option)	Voices	Needs	Notes
System voices — free, offline	A curated set of English voices (up to 5 female + 5 male) from your OS	Nothing	Works offline; quality depends on installed Windows voices
OpenAI neural — natural (uses your OpenAI key, ~$0.015/1K chars)	Alloy, Ash, Ballad, Coral, Echo, Nova, Sage, Shimmer, Verse	OpenAI API key (entered under the OpenAI engine section)	Streams MP3 audio; falls back to a system voice until a key is set
Grok (xAI) — expressive (uses the xAI key below, ~$0.0042/1K chars)	Eve, Ara, Rex, Sal, Leo	xAI key — paste it in Settings, or leave blank to use XAI_API_KEY from your .env	Audio is synthesized on the Rust side, so the key never enters the page
Qwen (Alibaba) — multilingual (DashScope key below, ~$0.013/1K chars, 1M free)	Cherry, Ethan, Jennifer, Ryan, Katerina, Elias, Dylan, Sunny	DashScope key from Alibaba Cloud Model Studio — paste it in Settings, or leave blank to use DASHSCOPE_API_KEY from your .env	16 languages on qwen3-tts-flash; new DashScope accounts include roughly a million free characters. Synthesized on the Rust side like Grok

Good to know

• Speech is a summary channel: spoken output never reads website addresses aloud (a link speaks as its title, bare addresses are skipped) and very long replies are trimmed at a natural sentence. The transcript always keeps the full text — ask for the detail if you want it spoken.
• Agent output is spoken once, not line by line. With Speak agent output aloud on, an agent used to be narrated as it typed, which meant a long reply arrived as a stream of interruptions. Now what an agent is doing is announced once per kind of activity (“reading files”, “running commands”) rather than once per step, and what it says is spoken as a single piece when its turn ends. Workers running inside a graph, crew or loop stay silent — only the agent you’re talking to speaks, so an orchestrated run doesn’t talk over itself.
• Cloud engines fail soft: a network or key problem re-speaks the message with the system engine instead of dropping it.
• While anything is being spoken, the reactive LED backlight around the workspace pulses with the speech — a quick visual cue of who’s talking.
• The same Settings section also picks the Assistant brain (what thinks up Hermes’ replies) — that’s a separate choice from what speaks them; see voice settings.