What do different dog barks mean & can humans beat computers at understanding dog barks?
While previous studies have concluded that dogs can learn over 200 human words, this 2008 study is one of the few to statistically assess the meaning of different dog's barks.
Have you ever wondered if that bark was a good or bad bark, play or protect? Then possibly here, are some answers for 6,000 barks, human V computer, across six bark categories.
What do the SIX different dog barks mean?
For any dog owner, this is the holy grail of learning to communicate with your dog.
In the experiment they attempted to define "context specific" and "individual-specific" features of dog barks using a new machine-learning algorithm they developed. The analysis used a database of 6,000 barks from 14 dogs, "which were recorded in six different communicative situations.
The aim was for the computer to be able to classify all of the barks into the six categories.
"The recognition rates we found were highly above chance level: the algorithm could
categorize the barks according to their recorded situation with an efficiency of 43% and with an efficiency of 52% of the barking individuals. These findings suggest that dog barks have context-specific and individual-specific acoustic features."
The researchers followed on from previous studies were owners were asked to classify barks. In that study there was as great deal of various between the owner's abilities to identify the context of barks, particularly from dogs other than their own.
The machine learning process involved a lot of programming and mathematics to build the correct models. It is one of the most rigorous statistically based models I have seen on dog behaviour.
The whole basis for this study was on the barks of the "Mudi breed (a Hungarian sheep dog) .. they recorded barks from 14 individuals, sex ratio (male/female): 4/10, age: 4.21 § 3.17 years. The total sample size of barks analysed was N = 6,646."
How did they cause the different types of barks?
They collected bark recordings in seven different behavioural contexts, most of which could be arranged at the homes of the owners. The exceptions for this were two of the six bark categories: the “Fight” situation, which was staged at dog training schools and the ‘Alone’ situation, held in a street or in a park.
The SIX dog BARK situations are as follows:
“Stranger” (N = 1802): The experimenter (male, age 23 years) was the stranger for all the dogs.
He appeared in the owner's garden or at the front door their home in the absence of the owner.
“Fight” (N = 1118): the dog were encouraged to bark aggressively and to bite the glove on the trainer’s PADDED arm. The owner was present and had their dog on a leash.
“Walk” (N = 1231): The owner was requested to go through their usual dog walk procedure. This included visual and aural cues such as them saying 'walkies'.
“Alone” (N = 752): The owner tied their dog to a tree or similar restraint and walked out of sight.
“Ball” (N = 1001): The owner held a ball or favourite toy at a height of around 1.5 m in front of the dog.
“Play” (N = 742): The owner was requested to play their dogs favourite game such as 'tug-of-war', chasing or
The highest recognition rates for both Computer and human were achieved for the barks recorded in the “Fight” (74% computer, human 50%) and “Stranger” (75% computer, 58% human).
One important difference between computer and humans was that humans were exposed to several barks in sequence coming from the same context. The computer made its analysis on “one” isolated bark.
Previous studies have found that the intervals between the barks have an effect on the human categorization of the dog’s motivational state, That is barks with shorter intervals were considered more “aggressive” while barks with longer intervals were considered more “fearful”, “desperate” and “happy”. This might suggest that humans would do better at defining aggressive barks such as stranger and fight, and while they recognize these best, they are not as good at it as the computer.
The next four categories were much lower recognition for human and computer.
The computer outperformed humans (once the statistical error was considered) in categorizing "walk" and "ball". although both these categories were less than 50%.
In the “Play” and “Alone” contexts human showed much better performance., though again recognition was below 50%. Note the other 50% of error categories were divided up between the other five possible bark states, meaning that 50% was relatively good.
The discussion should perhaps centre on why humans are so good at recognizing stranger and fight barks. The obvious answer might be that these are the most aggressive of barks. You should also note that during the domestication of the wolf into the domestic dog, these two barks are most likely what the wolf/ dogs were chosen for. Dogs were initially used as protection of the village and individual sleeping areas was paramount.
These 'stranger' and 'fight' barks are warning barks with high volume and some level of distress. Even so, the computer was able to recognize these barks approximately 20% better than humans! These very high results should be tempered by the fact that the computers were fine tuned a very high level to separate out spectral information in the barks well before the tests were conducted, while the humans only went into the tests with their current level of skill.
The third highest recognition by humans was the "Alone" bark. While the other barks were warning barks to alert the human, these are barks from the dog, trying to locate the human or show its own distress. Human empathy seems to come into play with the relatively high recognition rate.
the fourth highest recognized bark, the PLAY bark was recognized about four times better by a human than the computer. While this comes down to the skill of the engineering and computer algorithms, they were fairly complex attempts and refinement made on the fly. Again the humans self interest, the joy of interacting with their dog in a fun way might suggest that we are fully attentive to the positive feedback that the dog is giving us, more than what the machine can ascertain.
The BALL bark was similar to the PLAY bark (both positive fun things for the dog to do), but at a less energetic level. While the human easily outscored the computer in figuring out the PLAY bark they did substantially worse than the computer on the more muted BALL bark.
The worst recognized bark by humans was the WALK bark. By comparison to the urgency of fight or play barks it might be understandable how this was harder for people to recognize. another consideration might be that not all people find dog walking a joy and therefore it is something that they don't do very often or don't want to respond to.
Looking at the sound spectrum used by dogs for the different barks it was found that the domestic dogs could modify their sounds within the same vocalization type (ie the difference between aggressive and play barks were not purely frequency dependent such as low growls for warning). This suggests a profound evolution from its originator the wolf, that only tends to bark to warn of an attack on the den or the pack. It is not know if the dogs have had an evolution of their vocal chords or have changed barks in accordance with their learned experience.
It is likely that the computers and algorithms will only get smarter, while the human recognition will remain the same. if this is the case, then the application for such a hand held device might be in understanding the state of an approaching dog, understanding your own dog better or diagnosis of illnesses in dogs. Imagine if it could tell an owner when the dog wanted a walk?
Article by Bruce Dwyer. If you wish to use any of this information please refer to the article as a reference and provide a link to http://www.dogwalkersmelbourne.com.au
If you are interested on being kept up to date on dog article releases or healthy dog treat specials, Join us on http://www.facebook.com/HealthyDogTreatsShop
Classification of dog barks: a machine learning approach; Csaba Molnár · Frédéric Kaplan · Pierre Roy · et al. 2008