-
Notifications
You must be signed in to change notification settings - Fork 0
Voice Mode
Voice mode, the Talk entry in the sidebar, is a spoken conversation rather than dictation. It listens continuously, works out when you have finished a sentence, answers out loud, and lets you cut in mid-reply.
It is marked beta in the interface, and that is honest. It works well with a decent GPU and a headset. Over laptop speakers, echo cancellation has more to do and the assistant will occasionally answer itself.
Press Start talking. The first time, this does three things: downloads the talk model if you do not have it, downloads the speech recogniser, and loads the voice. After that the same button takes a few seconds.
There is a cancel button throughout. Leaving the screen mid-download stops the download and releases the microphone.
One flat circle, and it is the whole status display.
- Growing and shrinking with your voice: listening
- Shrunk to a dot with a rotating arc: thinking, or searching the web
- Pulsing steadily: speaking
- An outline rather than a filled circle: the microphone is muted
Under it, one line of text says the same thing in words, and after each reply a small number shows the milliseconds between you falling silent and the first word coming back. That number is the thing worth watching if you are tuning performance.
Just talk. The assistant stops as soon as it is sure you meant it, which takes about a quarter of a second of sustained speech, long enough that a cough or a chair scrape will not cut it off.
If what you said turns out to be "mhm" or "right", it picks up where it left off rather than starting the sentence again. Backchannels are recognised in every interface language.
While a conversation is live:
| Key | Action |
|---|---|
Esc |
End the conversation |
M |
Mute or unmute the microphone |
C |
Show or hide the transcript |
Space |
Skip the rest of the current reply |
System uses the voices your operating system already has. They are instant, cost nothing, and sound like your operating system. Which ones you get depends on the platform; Windows ships few by default and more can be added in Windows settings.
Natural is a neural voice that runs on your machine, about 90 MB the first time. It sounds considerably better and it is English only. If it fails to load for any reason, Draggy falls back to a system voice rather than losing the conversation.
Speed is adjustable from 0.9x to 1.25x. Anything faster stops being comprehensible for most people, which is why the range stops there.
All three settings, system or natural, which voice, and speed, live in
Settings → Personalization, not on the Talk screen itself. Talk is deliberately just the
orb and the Start talking button.
If web search is enabled, the talk model can ask for it. When it does, the conversation says something like "let me look that up" out loud, runs the search, and answers from the results. Silence for several seconds reads as a broken assistant, so it says something rather than nothing.
It is allowed one search per question. A small model handed search results will sometimes ask for more of them instead of answering, and a conversation that loops through "let me look that up" is worse than one that answers from what it has.
Turn web search off in Settings → Tools and the mechanism is removed from the
prompt entirely, rather than merely discouraged.
Nothing leaves the machine except web searches. Speech recognition runs locally with Whisper, on WebGPU if your machine supports it and on WebAssembly if not. The neural voice runs locally too. Audio is never uploaded, never stored, and never written to disk.
It interrupts itself. The assistant's own voice is coming back through the microphone. Use headphones, or lower the output volume.
It answers before you have finished. You pause mid-thought more than average. Draggy waits longer after a trailing "um" or a conjunction than after a full stop, but it is reading your transcript, not your mind.
It takes a long time to answer. Look at the millisecond counter. If it is
consistently over about 1500 ms, the talk model is too big for your card. Pick
a smaller one in Settings → Personalization.
The status line says "basic detection". The neural voice-activity detector did not load and a simpler energy-based one is in use. It works, but it is more easily fooled by background noise.