Listening & Wake Word

The Intern listens continuously by default: whenever the voice-activity detector (VAD) hears speech, it opens a turn, streams it to STT, and closes the turn after a short silence. Everything on this page tunes that loop.

Most knobs here are HAL environment variables in /opt/hal/.env. They're read once at start:

sudo nano /opt/hal/.env
sudo systemctl restart hal
sudo journalctl -u hal -f      # watch the effect

caution

A device-profile OTA update replaces /opt/hal/.env with the shipped copy. Keep your changes in a file and re-apply them after updates, or disable the bootstrap service while you tune.

Wake word

Off by default — the Intern responds to any speech it hears. Turn it on to make it wait for a wake phrase.

Web UI: http://<device-ip>/setting#general → General → Require an attention trigger before handling speech. The panel also lists the current wake phrases.

API:

curl -s -X PUT $DEVICE/api/device/config \
  -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{"wakeword":true}'

Saving restarts HAL.

Wake phrases are generated from the agent's name, not typed in: hello, hey, hi, alo, okay, ok, wake up + the agent name (e.g. "hey Nova"). The same prefixes + "autonomous" and + the device type always work too. To change the phrases, rename the agent: edit the - **Name:** … line in /root/.openclaw/workspace/IDENTITY.md (or rename the device from the Autonomous app). os-server re-reads the file every 5 seconds and pushes the new phrases to HAL — no restart needed.

After a wake phrase, follow-up questions don't need the phrase again for a while:

HAL env	Default	Effect
HAL_WAKEWORD_FOLLOWUP_TIMEOUT_S	20	Seconds after a reply during which you can keep talking without the phrase. 0 = phrase required every time.

A single press of the button also wakes it (see Button).

Sensitivity and turn-taking

HAL env	Code default	Intern v2 ships	What it does
HAL_VAD_THRESHOLD	3500	500	RMS energy that counts as "someone is speaking". Lower = more sensitive.
HAL_SPEECH_HOLDOFF	0.2	0.05	Seconds speech must last before a turn opens (filters clicks and bumps).
HAL_PRE_ROLL_FRAMES	8	—	64 ms frames kept from before the trigger, so the first syllable isn't clipped.
HAL_SILENCE_TIMEOUT	2.5	1.2	Seconds of silence that end a turn when STT hasn't produced a final transcript (fallback clock).
HAL_ENDPOINT_SILENCE_S	0.8	—	Seconds of silence that end a turn once STT has a final transcript. Raise it if the Intern cuts you off mid-sentence; 0 disables.
HAL_MAX_SESSION_DURATION_S	30	—	Hard cap on one listening session.
HAL_SESSION_COOLDOWN_S	0.3	—	Pause before a new session can open.
HAL_STT_KEEPALIVE	false	true	Keep the STT connection warm — lower latency, more idle STT usage.

Noise filters

HAL env	Default	Intern v2 ships	What it does
HAL_WEBRTCVAD_ENABLED	false	true	Fast speech/non-speech pre-filter
HAL_WEBRTCVAD_AGGRESSIVENESS	2	—	0–3; higher rejects more noise. 0 lets room noise open sessions.
HAL_SILERO_ENABLED	false	false	Neural VAD that rejects TV, music and background chatter
HAL_SILERO_THRESHOLD	0.3	—	Silero speech probability needed (presets use 0.15–0.18)
HAL_SILENCE_VAD_ENABLED	true	—	Silero also confirms silence at end of turn — without it, a noisy room can keep a turn open forever

Echo handling (so it doesn't hear itself)

HAL env	Default	What it does
HAL_AEC_ENABLED	true	WebRTC acoustic echo cancellation
HAL_AEC_DELAY_MS	205	Speaker→mic delay the AEC assumes
HAL_AEC_NS	true	Noise suppression in the AEC
HAL_ECHO_RMS_FLOOR	200	Minimum level that counts as speech right after TTS
HAL_ECHO_GATE_MAX_WAIT_S	1.5	Max wait for the speaker tail to clear
HAL_ECHO_SIMILARITY_THRESHOLD	0.55	Transcripts this similar to what it just said are dropped as echo
HAL_WARM_MIC	true	Keep the mic open between turns for faster pick-up

Recipes

Noisy room / it wakes up on the TV:

HAL_VAD_THRESHOLD=650
HAL_SILERO_ENABLED=true
HAL_SILERO_THRESHOLD=0.18
HAL_WEBRTCVAD_ENABLED=true
HAL_WEBRTCVAD_AGGRESSIVENESS=2
HAL_SILENCE_VAD_ENABLED=true
HAL_STT_KEEPALIVE=false

…and consider turning on the wake word.

It cuts me off while I'm thinking:

HAL_ENDPOINT_SILENCE_S=1.5
HAL_SILENCE_TIMEOUT=2.5

It doesn't hear me from across the room: lower HAL_VAD_THRESHOLD (e.g. 400) and HAL_SPEECH_HOLDOFF; keep Silero on to stay filtered.

Ready-made presets

The repo ships three tested presets in hal/env-presets/. Copy the HAL_* lines into /opt/hal/.env (keep your audio device and key lines) and restart HAL. Try them in order:

Preset	Use it when
01-balanced-cost-safe.env	First choice. VAD 650, Silero 0.18, WebRTC VAD 2, STT keep-alive off, echo floor 300, realtime cost guards on.
02-sensitive-filtered.env	01 misses quiet speech. VAD 500, holdoff 0.05, silence 3.0 s, Silero 0.15.
03-max-sensitive-debug.env	Short debugging only — Silero and WebRTC VAD off, keep-alive on. Picks up everything, including noise.

Button

The top button is a physical control for the voice loop:

Gesture	Action
Single press	Stop speech and music, cancel the current reply, unmute the mic and start listening (also works as the wake trigger)
Triple press	Reboot
Hold 2 s	Sleep
Hold 5 s	Shut down
Hold 10 s	Factory reset

There is no voice barge-in in the default turn-based mode — press the button (or POST /tts/stop) to interrupt. Experimental voice interruption is available in Realtime Voice → Live mode.

Mic mute

The slide switch on top mutes the mic: HAL stops feeding audio to VAD/STT. To mute from code, see HAL API → Mic.

Audio devices and volume

HAL env	Intern v2 ships	What it does
HAL_AUDIO_INPUT_ALSA	plug:device_micro2	Mic ALSA device
HAL_AUDIO_OUTPUT_ALSA	plug:device_speaker	Speaker ALSA device
HAL_AUDIO_SENSING_DEVICE	—	Optional separate mic for ambient sound sensing

List devices with arecord -l / aplay -l over SSH.

Volume persists across reboots — the last value you set wins; a fresh device starts at 100 %. To read or set it from code, see HAL API → Speaker.