Back to Feed
article 3m read

TV dialogue kept waking the assistant, so now it checks what the TV just said

The TV was the main source of false wakes, so the agent gave the assistant a copy of what the TV is playing and a rule: if it can't line the two up, it says unsure, never TV. All in shadow mode until real playback proves it.

Article DepthPractice

Flow

Building a TV check, in order

Nothing acts on a verdict until real playback has bounded the rate.

  1. 1Copy

    Read-only copy of the TV sound, plus the assistant's own replies.

  2. 2Probe

    Prove the copy is live while something plays.

  3. 3Align

    Find the delay with a confidence, or answer unsure.

  4. 4Check

    Wake model on the copy first, energy share as backup.

  5. 5Shadow

    Count verdicts beside every wake during real TV time.

  6. 6Decide

    Let it act only once the counts say it's safe.

Most false wakes in my living room come from the TV. The agent's answer was not to cancel the TV's sound but to check it: when the assistant hears its name, it looks at a copy of what the TV just played and asks if the TV said it. If you have a speaker you control, you can do the same.

Copy the TV's sound, and don't touch it

Take a read-only copy of the audio where it's decoded, before it reaches the speakers. Keep a few seconds in memory. Don't mute, reroute, or change the volume. A check that breaks playback will get turned off. Put your assistant's own spoken replies in the same copy, since they play through the same speakers.

Line it up, or say you can't

The microphone hears the TV late, and a mic on Wi-Fi hears it much later. Use cross-correlation to find the delay, and keep a confidence with it. If the peak isn't clear, use a recent good delay. If you have none, answer unsure.

Ask your own detector first

Run the same wake model on the TV copy at that moment. If it fires there, the TV said it. In the agent's tests this did most of the work. Then, as a backup, measure how much of the mic's sound the TV copy explains. A person talking over the TV adds sound the copy can't explain.

Set the cutoff with your own voice

Record yourself saying the wake phrase over TV audio at a few loudness levels. Pick the strictest cutoff that almost never calls you the TV. Mine sits high, because lower values began rejecting some of my tries when I was about as loud as the TV.

Run it in shadow first

Record the verdict next to each wake and don't act on it. Before trusting the numbers, check these:

  • a probe shows the copy is live while something plays;
  • unsure is its own count, not lumped in with room;
  • a check that crashes or never answers ends up unsure;
  • you have enough real playback hours to bound the rate.

The agent's offline scenes showed the gate catching most TV wakes without blaming my voice. They can't tell me how often it happens at home. Only real playback can.

Keep reading

Pick up a connected idea or branch into a nearby one.

5 paths forward