How human lives are saved by computers that can hear

"This article was published in our May 2017 newsletter". Sign up here.

One wrong step, that’s all it takes – and an old man suddenly finds himself lying on the floor in his living room. Injured by the fall, he is unable to get up again on his own. Now all he can do is hope that someone finds him as quickly as possible. Unless he has made provisions for just this kind of emergency, that is. By installing a digital sound detector, for instance.

What does an accident sound like?

acoustic emergency call systems can save lives

Researchers at the Fraunhofer Institute for Digital Media Technology IDMT in Oldenburg have developed precisely such a system. It is “trained” to pick up the kind of noises that typically accompany accidents in the home. The system detects emergency situations in inpatient care settings or the domestic environment acoustically and automatically triggers an emergency call. The idea is not entirely new. Acoustic emergency call systems are already available that can detect when patients or old people have an emergency in their homes and require assistance. “The problem is that false alarms occur very frequently”, says Fraunhofer engineer Dr Stefan Goetze. The new generation of electronic emergency call systems on which the Fraunhofer researchers are working is equipped with sensitive hearing. For one thing, their “electronic ears” are trained to recognise warning signs in the form of specific sounds such as severe coughing, calls for help or falling furniture. And for another thing, they are able to identify background noise – the sound of a bus passing outside, for instance – and filter it out as irrelevant.

Database of sounds

However, the emergency sounds that the technology is able to recognise are only a few of the more than 100 “target sounds” that the IDMT has stored in its database. The researchers do not simply file away the sounds they have recorded. They break down the acoustic signals, analyse them and give them a description – the metadata. On the basis of this data they develop algorithms. In turn, these algorithms are used to teach the recognition systems to react to different “target sounds” depending on the environment in which they are to be used. “One of the questions we are exploring in collaboration with doctoral students from the Hearing4All cluster of excellence at the University of Oldenburg is how acoustic signals can be described using specific attributes so as to ensure that speech and noise recognition systems function reliably”, reports Goetze.

Preventing crime

Certain criminals also make characteristic sounds, such as graffiti sprayers who illegally daub their artwork on trains: steps on the train tracks, the clicking sound of the mixing ball when the spray can is shaken, the hissing noise when the paint is forced through the nozzle of the spray can. One of the IDMT’s acoustic monitoring systems is also able to recognise these noises. If this electronic guard dog detects them, it will sound the alarm. Stefan Goetze and his colleagues developed the system so that railway operators can catch graffiti sprayers red-handed in the hope that this will keep them away from the trains long-term. After all, the damage they cause is considerable: in 2015, German rail operator Deutsche Bahn alone spent more than eight million euros having graffiti removed by hand and trains repainted.

Light switches not triggered by snoring

The artificial “ears” from Oldenburg already work almost as well as human ears. “When it comes to recognising noises, that is to say non-verbal acoustic events, we achieve nearly 80 percent of human recognition capabilities”, says Stefan Goetze. The achievement rate is even higher with speech recognition, where the acoustics researchers reach 85 to 95 percent. Such a high recognition rate is necessary to avoid misunderstandings – for example when it is a question of distinguishing between language and other human sounds. “After all, if a smart home misinterprets snoring and automatically switches on the lights, people will certainly not be willing to accept this kind of technology.” If researchers succeed in minimising the error rate, the technology could soon become widespread in all kinds of areas – making our daily lives easier.