0078 — “Hey Conch”: a wake word that stays on your computer

• Status: accepted
• Date: 2026-10-04
• Goes with: ADR 0077 (voice notes, natural voices, barge-in), ADR 0027 (voice), ADR 0054 (the desktop app)

Context

Talk mode starts with a click. OpenClaw's apps answer to a spoken wake word; Conch's didn't. A wake word means a microphone that's always open, which is exactly what a person should be able to trust completely or not turn on at all. The bar:

• Local. Nothing a person says near their computer leaves it. No account, no key, no cloud.
• Off by default, and turned on only by the person, on the device where they want it.
• Always visible while it listens, with one press to stop, even with the window closed.
• No new heavy dependency for one feature, and nothing to train.

The engine

Choice	Why not (or why)
Picovoice Porcupine	Good and small, but proprietary: an access key from Picovoice's console, a custom phrase trained and licensed there. Not offline-free, not ours to ship.
openWakeWord	Open and light, but each phrase is a model trained with a synthetic-speech pipeline on a GPU. There is no “Hey Conch” model, and one per persona name is out of reach.
The browser's speech recognition	Electron has no speech service, and in Chrome it's Google's servers or a per-language on-device pack. Neither is Conch's to rely on.
sherpa-onnx keyword spotting	Open-vocabulary and streaming (any phrase, no training), the best fit long term, but a native addon of about 30 MB per platform in the desktop app, with Electron builds to keep in step. The upgrade path, behind the same check.
whisper.cpp, gated by voice activity	Chosen. Already here for private listening (the whisper need and its model), offline, any phrase in any language, nothing to train or ship. It runs only on short bursts of speech the window has already cut out, never on silence.

Tried on this computer: Piper saying the phrase, whisper.cpp's base model reading it. Given “Hey Conch.” as its prompt, whisper.cpp leaves the phrase out of what it writes (it takes it for what came before), so the prompt is “Talking to an assistant called Conch:”, which heard “Hey Conch”, “Hey Conch, what is the weather tomorrow?” and not “I found a conch shell on the beach”. A smaller audio context (-ac 320) was faster but heard “Hey coach”, so it isn't used. A check takes about two seconds on a laptop's processor, most of it loading the model; keeping it loaded (whisper-server) or sherpa-onnx would bring that down.

Decision

• Only in the desktop app. The gateway knows it runs there (theApp()); elsewhere the switch isn't shown and the routes refuse. Settings → Voice → Hey Conch: off until turned on, kept per device (localStorage, like every voice choice), so a backup can't turn it on anywhere. It needs private listening, and offers its setup when it isn't ready.
• The window listens, the computer decides (features/voice/WakeWord.tsx, wake.ts): the microphone with echo cancellation and noise suppression; a burst cutter (louder than the room, with 300 ms before it, ended by a 350 ms pause, 0.35–3.6 s of speech; talk that runs longer isn't a call, nor is the rest of it). Only a burst is sent, as a WAV, to POST /api/voice/wake on the same computer; one at a time, the rest dropped. whisper.cpp reads it with Conch's own prompt and heardWake looks for the phrase at the start. The recording is deleted the moment it's read; only “heard or not” goes back, with what came after the phrase.
• What happens when it hears it: the window comes to the front ({ type: 'show' } to the app), talk mode opens on the chat that's showing (or a new one), and anything said after the phrase is the first thing said (“Hey Conch, what's on tomorrow?”).
• It shows, the whole time. In the window's header, Listening for “Hey Conch” with Stop (Nacre ListeningIndicator, in words, never only a light). In the tray, shown even when the icon is turned off: “Listening for “Hey Conch”” and Stop listening, which turns it off in the window too (wake.stop). The window tells the gateway it's listening every 45 seconds; if it goes quiet (closed, crashed), the tray stops saying so within about a minute. The system's own microphone light shows too.
• While it listens with the window closed, the window keeps running unthrottled (setBackgroundThrottling(false)); otherwise it's throttled as before. It pauses while talk mode has the microphone.

Consequences

• Conch can be called by voice in the desktop app, offline, with nothing new installed beyond private listening.
• Nothing about it reaches a backup, another device, or the web.
• Not done: the phrase follows the persona's name; a wake word on phones (a page in a mobile browser can't listen in the background); a faster engine (above).