Alexa out.
Luna in.

Read the settings and the wake scores

EchoLocal keeps its settings in /data/misc/echolocal/state.json on the Dot, and the same controls appear on the device's page in Home Assistant. Ours showed the problem straight away: microphone mixing on Center mic, 20 dB of gain and a wake cutoff of 0.85.

The Dot also logs every wake attempt with the score its wake-word model gave the phrase, from 0 to 1, and the cutoff it had to beat. That one line tells you whether the fault is the microphones, the cutoff or the model:

> adb shell "logcat -d | grep echolocal | grep -i wake"
wake detected slot=2 id=hey_luna peak=0.988 crossing=0.870 cutoff=0.85

Reading the logs and the file needs adb: over USB, or over Wi-Fi once EchoLocal's remote-adb switch is on.

Listen with more than one microphone

Center mic means one microphone out of the Dot's seven. We switched mixing to Beamformer, which runs Amazon's own far-field filters, the same ones a stock Alexa relies on: eight listening directions built from four of the seven microphones, aimed at whichever direction is loudest. Home Assistant offers Center mic, Delay and sum (EchoLocal's own mix, which uses all seven microphones in six directions) and Beamformer.

Played back on recordings of Kev's voice, the two aimed mixes came out close: delay and sum slightly ahead up close, the beamformer holding on to more of the far attempts. So the beamformer stayed.

A little more gain, in steps

20 dB went to 24, then 28, with a test from across the room after each. More gain lifts a quiet, distant voice, but it lifts the room's hiss and the TV too, so we stopped at 28; the control goes to 59. Levelling and echo cancellation stayed on, and the separate noise-reduction switch stayed off.

Past that point, don't expect more volume to rescue a distant voice. The wake-word model scores how far your voice stands above the room, so a boost that lifts the noise with it changes almost nothing: on the bench, a synthesised "hey luna" about 24 dB above the room's noise woke the Dot and one about 18 dB above didn't, at every level of amplification we tried. What helps is whatever raises your voice over the room: aiming the microphones, turning the TV down, facing the Dot.

Set the cutoff from the far-away scores

With the beamformer listening, "hey luna" still scored 0.99 up close but only 0.59 and 0.62 from further back, against a cutoff of 0.85: the Dot was hearing the phrase and throwing it away. At a cutoff of 0.55 every attempt woke it, but the two furthest only just (they crossed at 0.553 and 0.554), so it went down to 0.50 for some margin, and later to 0.40.

That last step was for misses nobody could see. The Dot only logs a near miss once the score reaches 0.50, so with the cutoff at 0.50 a failed attempt left no trace at all. At 0.40 a quieter attempt still wakes it, and ordinary speech is nowhere near: on the bench, "what time is it" scored 0.06 on the same model.

Two traps. Despite the name wake word sensitivity, a lower number is more sensitive, because it is the score the phrase has to beat. And Home Assistant's slider stops at 0.50, so anything lower has to go in state.json.

Choose a well-trained wake word

Every wake word is its own small model, and some are trained far better than others. "Okay computer" barely registered from a distance however it was tuned, and it twice woke on Luna's own reply ("7:46 PM"). Each time it cut her off and left the Dot listening for fifteen seconds, deaf to a real "hey luna". It was switched off.

In its place went "okay nabu", Home Assistant's own wake word, trained by the author of microWakeWord and already on the Dot. In Kev's first test it woke three times out of three, including from across the room, scoring 0.88 to 0.995 against a cutoff of 0.80. It's now the one to use from a distance; "hey luna" stays in the first slot.

Keep the word you use in slot 1. In EchoLocal 0.0.7, if slot 1 is empty, a word in slot 2 is heard but the turn fails with a red ring: the Dot looks it up in the wrong place.

Make it stick, then prove it

echod reads state.json only when it starts, so every edit was followed by a reboot and two checks: the new value is still in the file (echod quietly resets an invalid one, such as a misspelt mixing mode, to center), and the uptime is a few seconds, proving the Dot really restarted. Then test again from the far spot and read the scores.

"microphone": {
  "gain": 28, "mixing": "beamformer",
  "leveling": true, "cancel": true,
  "denoise": false },
"wake": { "words": [
  { "id": "hey_luna",  "threshold": 0.4 },
  { "id": "okay_nabu", "threshold": 0.8 } ] }
> adb reboot
> adb shell cat /proc/uptime
18.00 46.08

Then prove it in the room, not only on a recording. We built a bench on the PC that replays raw recordings from all seven microphones through the Dot's own mixes and wake-word models. On it, a better-trained community "hey luna" woke on 11 of Kev's 11 recorded attempts, against 4 for the one on the Dot. Live, it wouldn't wake from across the room and once woke by itself, so the old model went back. Nothing changes on the Dot on the bench's word alone.