Alert bark vs lonely bark: what actually changes in the sound

Every dog owner runs the same classifier a dozen times a day without noticing.

Somebody is at the door. Somebody is at the door who should not be. The dog has been left alone and would like that to stop. A fox went past the fence. You get up for one of those and not the others, and you rarely stop to ask how you decided.

The interesting thing is that this is a measurable question, and it has been measured. When researchers put a large number of barks through acoustic analysis and asked human listeners to label them, three parameters kept doing the work.

Three numbers, and one of them dominates

Pongrácz and colleagues (2006) recorded barks from dogs in different situations, measured them, and then played them to people who had never met the animals. The parameters that predicted what listeners heard fell into clean bins.

Pitch. Barks associated with alert and aggressive contexts sat low, roughly in the 401 to 531 Hz band. Barks associated with fear and being left alone sat much higher, in a wide band from about 732 Hz up to 1833 Hz.

Harshness. This is the parameter most people have never heard of and can hear perfectly well. Acousticians call it tonality, or harmonic-to-noise ratio: how much of the sound is a clean musical tone and how much is noise sitting on top of it. Alert barks in that study were harsh, down at negative 2.1 to 4.6 decibels. Isolation barks were tonal and clean, 11.6 to 35.4 decibels. A harsh bark sounds like it has grit in it. A tonal one almost sings.

The gap. And here is the one that surprised us when we first read it. The single strongest cue was not pitch and not harshness. It was the interval between one bark and the next. Alert barking came in fast volleys with gaps of around a tenth of a second. Isolation barking came slowly, with gaps three to five times longer.

That matches something you already know without having words for it. The doorbell bark is a burst. The bark of a dog who has been shut in the kitchen is spaced out, almost patient, and it goes on.

What field recordings added

Yin (2002) took a different route: 4,672 barks recorded from dogs in real situations rather than in a lab, then analysed. The numbers landed in a similar place. Disturbance barks averaged 686 Hz and ran about 345 milliseconds each, lower and longer. Isolation barks averaged 860 Hz.

Two independent studies, two methods, and the same direction of travel. When that happens you can be reasonably confident the effect is real rather than an artefact of how one team recorded their dogs.

The pair nobody can separate

Now the part that does not appear in the marketing copy of any app in this category, including the parts of ours where it would be easier to leave it out.

In that same set of field recordings, play barks averaged 840 Hz. Isolation barks averaged 860 Hz. Those two distributions sit almost entirely on top of each other. A dog having a wonderful time and a dog who wants company make, acoustically, close to the same sound.

This is not a gap in the equipment. It is a fact about dogs. Both states are high-arousal, both produce high tonal barks, and the sound alone does not carry enough information to separate them. The thing that tells you which one you are hearing is the context you can see and the microphone cannot: whether there is a toy involved, whether you just came home, whether the dog has been alone for two hours.

Why there is no confident "happy" reading

The same problem gets worse when you ask for emotional labels rather than situations.

Pongrácz and colleagues (2006) and Jégh-Czinege and colleagues (2019) both looked at how pitch relates to a dog's emotional state, and they disagree about which direction the effect runs. Not about the size of it. About the sign. When two careful studies point opposite ways on the same parameter, the honest reading is that the underlying signal is weak, not that one team made an arithmetic error.

Machine classification tells the same story from the other end. Molnár and colleagues (2008) built a system to sort barks into six contexts automatically. It got fewer than half of them right, which was roughly what untrained human listeners managed on the same task. That is the published ceiling for this problem, from people who were trying hard.

So when an app tells you with confidence that your dog is happy, it is not reporting a measurement. There is no measurement to report.

What you can listen for

None of this means the sound is uninformative. It means the informative parts are narrower than the marketing suggests, and they are learnable.

The three that hold up across studies:

Count the gaps, not the barks. Rapid-fire with almost no space between is the alert pattern. Spaced and repetitive is the isolation pattern. This is the cue you can use from another room without seeing anything.

Listen for grit. A rough, noisy bark and a clean, almost tonal one sit at opposite ends of the range that separated alert from distress in the Pongrácz work.

Treat high and tonal as arousal, not as an emotion. High and clean tells you the dog is worked up. It does not tell you whether that is delight or distress, and nothing in the acoustics will.

How we built this into the app

PawParley measures the same parameters those studies measured: pitch and its contour, harshness, duration, and the gaps between bursts. It shows you those numbers alongside the reading, so you can see what the description was built from rather than taking it on trust.

It also refuses in the places the research refuses. There is no confident happy label, because two good studies disagree on the pitch effect behind it. Play and isolation are not offered as a confident distinction, because Yin's field data says they overlap. When the sound is ambiguous the app says the sound is ambiguous, which is a less impressive product and a more truthful one.

What you get is a description of what a sound acoustically resembles, with the measurements attached. Not a sentence in your dog's voice. The dog is the only one who has that.