Hands-Free and wake
This guide describes the current build.
Hands-Free is a persistent combined dictation-and-command session. The wake listener is an activation layer that normally opens one short command window, or sends your words to a named app. They are not the same setting.
Hands-Free sessions
- Open Settings → Dictation and turn on Session button (command_mode.enabled, off by default).
- Open Settings → Shortcuts → Session button. Choose the button (default Right Ctrl; Mouse 4 and 5, other modifier keys and F13 to F24 are available) and set the behaviour to toggle. The default, hold, only listens while the key is held; toggle opens a session that stays open until you end it.
- Press the button to start. A toggled session starts in the Dictate lane.
- Say “dictate mode” for the Dictate lane, “command mode” for the command lane, or “ava mode” to send every utterance to Ava. “Hey ava”, “so ava” and “oracle” also switch to Ava; bare “ava” deliberately does not.
- In Dictate, say “finish” alone to paste the staged thought and stay in the session. Say “go to sleep”, “samsara sleep” or “sleep now” alone to leave it; the staged draft is kept and comes back next session.
In the Dictate lane, speech is staged across pauses and exact command phrases still run without switching lane. Switch words only count when they are the whole utterance, so they cannot fire from the middle of a sentence.
The words that end things
| You say | What happens |
|---|---|
| “finish” | Pastes the staged thought and keeps the session open. Default commit word; config-only. |
| “go to sleep”, “samsara sleep”, “sleep now”, “stop listening”, “exit hands free”, “exit command mode” | Leaves the session. Your draft is kept. Add your own under Exit phrases. |
| “halt”, “cease” | Emergency stop: cancels speech, queues and any pending question, keeps your draft and the session. Replace them under Stop words. |
| “scratch that”, “scratch everything” | Removes the last staged chunk, or all of it after a confirmation. “Bring back my draft” restores it for five minutes. |
| “cancel”, “cancel dictation”, “abort” | Abort in a wake session; the draft is kept for recovery. |
A session also ends after 300 seconds of silence, or after five unrecognised requests in a row.
Wake listener
Wake listening is off by default. Turn it on in Settings → Dictation → Wake activation, or with the Wake toggle shortcut, Ctrl+Alt+W by default. The default wake phrase is “Jarvis”; the choices are jarvis, hey jarvis, computer, hey computer, samsa and hey samsa, and only the Jarvis phrases have a dedicated detector model.
When wake listening is enabled, say the phrase and then a command. By default Samsara waits up to five seconds for the command after the phrase. A wake command does not change the active recording mode.
wake_word_config.opens_session setting can make a wake phrase open Hands-Free, but it requires the toggle behaviour and has no control in Settings.Wake profiles: “okay claude”
newer than v0.30.0-beta.2A wake profile pairs a phrase with an app. Saying the phrase focuses that app’s window and starts dictating into it. Edit them in Settings → Dictation → Advanced → Wake profiles (wake_profiles).
| Profile | Phrase | App | Send word | Default |
|---|---|---|---|---|
| claude | “okay claude” | claude.exe | “over” | on |
| hermes | “activate hermes” | Hermes.exe | “send” | on |
| warp | “activate warp” | warp.exe | “over” | off |
The send word is the last word of an utterance. For the claude and warp profiles it presses Enter in the target app after Samsara checks the window; the hermes profile stages the text and never presses Enter. Each session accepts only its own profile’s send word. If the target app is not running you hear the error sound. After Claude answers, the reply can be read aloud, and “read the rest”, “repeat that” and “stop claude” follow up; see Claude desktop.
The Warp profile is switched off by default. Warp’s window exposes no readable reply text, so unlike Claude there is no reply reader for it.
A phrase of fewer than three syllables is disabled unless it has a model that passed the false-accept test. No custom wake models ship in the build, so profile phrases rely on the speech recogniser plus the Jarvis detector gate. A custom wake-model sample recorder exists for developers (tools/train_wake_model.py); no custom model has passed acceptance yet. Treat profile wake as new and still being tuned.
Samsara does not hear itself
newer than v0.30.0-beta.2When Ava is speaking, a command or hold capture that lies wholly inside her speech is dropped before transcription, so her own voice is not run as a command. You see the chip ignored – Ava was speaking. Speech that interrupts her is not dropped. The grace after she stops is 400 ms.
Music inside a session
These work inside Hands-Free and wake dictation without disturbing your draft, when they are the whole utterance: “pause music”, “unpause music”, “resume music”, “play music”, “stop music”, “next track”, “previous track”, “mute”. Bare “pause” and “resume” are different: they pause and resume the dictation buffer.
Timing and pace
Speech pace (standard, relaxed or unhurried) stretches every silence timing at once, which is the easiest way to stop being cut off. Individual timings are under Settings → Dictation → Advanced: Command pause, Dictation pause, Timeout after wake and Quick silence timeout.