How it works

Your browser passes the microphone to YAMNet, a sound classifier made by Google and trained on AudioSet, a large set of labelled sound clips. Twice a second the counter gives it the last second of audio and gets back a score for each of 521 kinds of sound, from speech and music to doorbells, cars and dogs.

A bark has to sound like a dog and like a bark at the same time. So the counter multiplies the Dog score by the strongest of Bark, Bow-wow and Yip and takes the square root, the same rule the DogQuiet app uses. In our tests on recorded barking, speech and TV, most barking scored above 0.7 and speech and TV stayed below 0.05.

  • Sensitivity sets the line a score has to cross: 0.70 on Low, 0.50 on Medium, 0.35 on High.
  • An episode starts with the first half-second over the line and ends after 5 seconds without one. A dog barking steadily for a minute is one episode, so this counts bouts of barking, not single woofs.
  • Time barking adds up the half-seconds over the line, which is closer to how much barking there was than the episode lengths are.

Getting a useful count

  • Put the device where you hear it. A windowsill facing the garden, or near the wall you share.
  • Plug it in and keep this tab in front. The page asks the browser to keep the screen awake, but some devices refuse in battery saver mode.
  • Turn off the TV and radio, or expect a dog on screen to be counted.
  • Check the setting first. Watch the bark score for a few minutes on Medium. Move to Low if it logs things that are not barking, or to High if it misses barking you can hear.

What it can’t do

  • Listen in the background. It works only while this tab is open and the device is awake. Phones usually pause web pages when they lock.
  • Be perfect. Accuracy varies with distance, walls, wind, other dogs and anything playing on a speaker. Faint barking can be missed.
  • Tell dogs apart. It hears barking, not whose dog it is.
  • Keep the sound. It never stores audio, so there is nothing to play back afterwards. The DogQuiet app saves a 10-second clip of each bark if you need that.
  • Replace your own record. A count from a website is a helpful start, not proof on its own.

From count to bark diary

Download the log as a CSV file (it opens in Excel, Numbers or Google Sheets) or print it. Then copy the times into a printable bark diary and add what the counter can’t know: where you were, what you heard and how it affected you. How to keep a dog barking diary covers what makes a diary credible to a landlord, an HOA or a council.

Privacy

The model runs in your browser with TensorFlow.js. Microphone audio is analysed in memory, about a second at a time, and then discarded. It is never recorded, saved or uploaded, and the page has no analytics or ads. The only thing it downloads is the sound model itself, about 8 MB from dogquiet.com, which your browser keeps for next time.

Questions

Is the online bark counter free?

Yes. There is no sign-up and nothing to install. It runs in any recent Chrome, Edge, Firefox or Safari that can use a microphone.

Does it record or upload what it hears?

No. Audio is analysed in memory about a second at a time and then discarded. Nothing is recorded, saved or sent anywhere; the log of times stays in the tab until you download it, print it or close the page.

Does it work on a phone?

Yes, while the page is open and the screen is on. Locking the phone or switching apps usually pauses web pages, so for a whole night a laptop that stays awake, or the DogQuiet app, works better.

Can it count a neighbour’s dog through a wall?

Often, if you can hear the barking clearly yourself. Walls muffle sound, so try the High sensitivity and expect faint or distant barking to be missed.

How accurate is it?

It is good at telling barking from speech and TV dialogue, but no detector is perfect. Distance, walls, wind and other sounds change the result, and a dog barking on TV counts as barking. Use it as a guide alongside your own notes.

Credits

The sound model is YAMNet by Google, used under the Apache License 2.0. We store its weights at half precision to halve the download; scores change by less than 0.01. It runs on TensorFlow.js, also Apache 2.0.