Wake-word listener

Always-on "Hey Saga" voice activation — the first working hands-free client, as opposed to the push-to-talk run_demon.py client.

How to use it

Say "Hey Saga" to the listener device's microphone, wait for the acknowledgment beep, then speak your request naturally.

How it works

• Script: python/services/saga_wakeword_listener.py, run as a systemd user service (needs a lingering session to survive reboots without an interactive login — see the script's own docstring for the exact units).
• Pipeline: mic (via PipeWire) → openWakeWord (SAGA_WW_MODEL, default hey_saga.onnx) detects "Hey Saga" → acknowledgment beep → energy-based endpointing records the utterance → POST to the local saga-voice-gateway's /converse → play the reply WAV back through the same output device via pw-play.
• Config, all via env (see the script's docstring for the full list): SAGA_WW_THRESHOLD (detection score, default 0.5), SAGA_WW_COOLDOWN (re-trigger debounce, default 3s), SAGA_WW_INPUT / SAGA_WW_SINK (audio device selection), SAGA_GATEWAY_URL / SAGA_GATEWAY_TOKEN.
• Why PipeWire mediates both directions: capture and playback go through one audio server so the mic and the speaker never fight over the same device mid-conversation.

Status & known limits

• One listener. Wake-word capture currently runs on a single always-on device — other clients are push-to-talk, not wake-word-activated.
• False-trigger tuning is live but unexercised: confirmation beep + threshold + cooldown + detection-paused-during-playback are all in place, but as of this writing no false triggers have been observed in real use to tune against.
• Which physical device and host this actually runs on is deployment state, not documented here — see nthmost-systems/saga/DEPLOYMENTS.md.