It is marketed as a toy: a dog collar equipped with a wireless microphone that records a pet’s barks, interprets them as emotions and transmits them as text messages like ”I’m bored, let’s play.”
The $139 device, the Bowlingual, ”comes a step closer to realizing everyone’s childhood dream of conversing with their pet,” its Japanese manufacturer says.
Though it all sounds very Dr. Dolittle, animal researchers have in fact been using voice-recognition technology for decades. The Bowlingual’s developer says its technology is based on 2,000 voiceprints from roughly a thousand dogs. And it turns out that the utterances of many species can be analyzed and identified using many of the same techniques that have allowed human voice recognition to make the leap from high-tech novelty to valuable application.
In Braunschweig, Germany, for example, researchers at the Institute of Technology and Biosystems Engineering have recently been able to decipher, with about 90 percent accuracy, what cows mean when they moo: hunger, thirst, need for milking and so on.
Dr. Gerhard Jahns, a control engineer who helped devise the project, said that about 700 ”vocalizations” were recorded from about 20 cows, a process he described as ”extremely time-consuming.” Cows can go for hours without making a sound, Dr. Jahns said, ”and it’s hard to get them to speak into the microphone.”
The system records the moos and scans them for specific tonal frequencies or aberrations: a cough at a certain pitch, for example, might indicate that Bessie is coming down with the flu. Dr. Jahns stressed, however, that the goal was ”information, and not diagnosis” of the animals; it is ultimately up to the farmer to decide what to do.
Wojtek Kowalczyk, a professor at the Free University in Amsterdam who helped develop the project’s software, said there was a big demand for such a system, although he estimated that wide-scale deployment was still three to five years away. Surprisingly, he said, the system’s most commercially attractive feature is its ability to recognize estrus, or the period when the animal is in heat, which may last only a few hours. ”If the cow is not inseminated during this period, a lot of money — a few thousand dollars per cow — is lost,” he said. Conventional methods, which rely on analyzing milk, are less accurate, he said.
To make any such system practical, it has to be able to recognize unfamiliar voices. So-called ”speaker independent” recognition is difficult for any voice application, as anyone who has tussled with a voice-based telephone directory knows too well. Professor Kowalczyk said that theoretically, this should be much easier to achieve with animals. Most of their vocabularies, or ”repertoires,” are smaller; the cow system, for example, recognizes about a dozen vocalizations.
But in practice the situation is quite different, he said, because the meaning of the source material ultimately remains elusive: ”How can we know what a cow wanted to say?” Animal physiologists try to deduce the answer by imposing specific conditions: not milking a cow for a long period, for example, will increase the likelihood that a moo emitted under those circumstances means ”I want to be milked.” As the database of vocalizations expands, the results will improve.
An important advantage of the system developed in Germany is its emphasis on simple hardware. In the interest of accessibility, it is based on consumer-grade PC’s using generic sound cards and microphones.
Indeed, that computers have become so fast and cheap has been pivotal to this kind of research. ”None of my current work could have existed even 18 months ago,” said Jim Nollman, founder of a nonprofit organization called Interspecies Communication, who has worked on voice recognition for whales. ”It’s ironic and self-defeating that the study of whale sounds has been developing for 20 years using equipment that only records one-tenth of the frequency spectrum some of these species use.”
That is partly a result of the limited hearing range of human beings: human ears peak out at frequencies of around 22 kilohertz. Many animals, however, communicate at levels well above this (or below it, in the case of elephants, for example). Recording just one second of 100-kilohertz sound, a common frequency among marine mammals, takes a megabyte; an all-day field recording, then, requires serious drive space.
These ultra-hi-fi recordings have helped reveal that the meaning of animal language is often generated not by the sounds themselves but rather by the way those sounds are modulated. A second of whale song, for example, contains thousands of Morse-code-like beats that whales use to communicate. And many birds seem to have a vocabulary of basically one call, which takes on different meanings based on its frequency or volume. ”It’s not syntax per se, but sequence,” said Brenda McCowan, an assistant professor at the Veterinary Medicine Teaching and Research Center at the University of California at Davis.
Dan Weary, a professor of animal welfare at the University of British Columbia, has used this knowledge of context to interpret pig calls. The pitch of a piglet’s squeal, for instance, has been correlated with pain when it crosses the one-kilohertz threshold. This knowledge might allow farmers to prevent outbursts of tail-biting, in which frenzied piglets can wind up killing each other. If the software detects a certain number of piglets emitting very high-pitched squeals at the same time, it can automatically send the farmer a warning.
Other applications include call-recognition programs that count bird, bat and other populations. A recent study in Australia actually used solar-powered PC’s to monitor the rapid proliferation of the cane toad.
Not all voices are so sanguine about this burgeoning field. ”An electronic ear can never replace the human ears,” said Colin Johnson, a livestock specialist at Iowa State University. He doubts similar systems will find widespread acceptance in American agribusiness, especially in the current economic climate. ”Producers aren’t looking to invest in unproven technology,” he said.
The Bowlingual has put off its American debut, mainly because of legal obstacles rather than economic ones, said Katsuhisa Oda, a spokesman for the Index Corporation, which developed the device with the Takara toy company.
The device could, after all, misinterpret a menacing mood. ”If the dog tries to bite a kid,” Mr. Oda said, ”then that could be difficult” as a liability issue in the United States — a consideration absent in Japan’s less litigious culture.
If the Bowlingual does catch on, could meow-activated cat doors or drug-sniffing police dogs whose barks can be decoded be far behind? ”If you can classify the morphology of the signal, it can be synthesized,” said Marc Hauser, a professor of psychology and neurosciences at Harvard University who imagined a system that would let pet owners track down runaway dogs. Playing aloud recorded samples of Rover’s bark while canvassing the neighborhood would elicit responses from nearby dogs, whose voiceprints could then be analyzed instantly for a match. If this sounds far-fetched, remember that Americans spent $30 billion on their pets last year, more than twice as much as in 1994.
On the other hand, these techniques could also lead to further automation of agriculture, with ”Matrix”-style systems robotically tending to endless rows of cattle or kennel puppies. Professor Weary regards such a vision as far from utopian, commenting, ”I don’t think the way to make the world a better place is to put more computers in pens.”
(Originally published in The New York Times in October 2002)