Wired UK – September 2019

PH

OTO

GR

AP

HY:

GR

EG^

WH

ITE

024

If you know that a well-known brand of smoke

alarm sounds identical to a rare parrot, you may be

a fan of Chris Mitchell, the chief executive of Audio

Analytic. His Cambridge-based company augments

consumer technology with sound recognition. Just

as speech recognition makes out spoken words,

Audio Analytic’s sound recognition technology can

tell the difference between the bark of a dog and

the cry of a baby. “It’s the equivalent of a language

model, but for sounds,” explains Mitchell.

AuditoryNET – Audio Analytic’s sound-

recognition AI – has extremely broad uses. The

company worked with the German earphone

manufacturer Bragi to produce AI earbuds that

prioritise a user’s safety by recognising, for

instance, an ambulance’s siren and alerting the

wearer to the direction of its approach. The idea

is to protect users from “all sorts of dangerous

situations you find yourself in, because you made

yourself completely deaf in a city”, says Mitchell.

Audio Analytic has, moreover, embedded

AuditoryNET in smart home devices such as the Hive

Home Hub 360, which can detect the sound of smoke

alarms and carbon monoxide detectors, barking

dogs, or windows being broken. AuditoryNET

also doesn’t need an internet connection to run –

good news for the privacy-minded.

To train AuditoryNET’s algorithm, Audio Analytic

used Alexandria, the world’s largest collection of

audio data. “One of the things we’ve done recently

is map out the entirety of our sound universe,” says

Mitchell. The company has created a sound map,

comprising over six million audio files. It looks like

a rainbow honeycomb: it visually repre-

sents similar sounds clustered together

in coloured blocks, or “tonal islands”.

Mitchell suggests turning down the

volume when he gets to the “smashing”

island. “This ‘break’ sound was from

a dining room,” he says. “[The glass

being broken] was in a laminate-wood

frame – and there are six other kinds of

‘break’ sound associated with it.”

Alexandria’s data pool of audio

samples had to be built from the ground

up. The stock sounds used in movies

were useless, as they aren’t true to their

real-world equivalents. “If you go and

see the latest Avengers movie, I’m sure

the Hulk will drop something on a car,”

says Mitchell. “That car alarm going off

is not a real car alarm.” Using the film’s

sound would train AI on a fake reality:

disastrous if you need to detect someone

breaking into a real car. So, the company

built its own lab to capture raw sounds,

and sent out teams to record others in a

variety of environments. The project even

relied on networks of volunteers to come

into the sound lab, or to use recorders to

capture sounds in the physical world.

Audio Analytic’s plans for the future

are confidential, but the company’s

objectives are clear: to continue to

develop technology that will identify

sounds inside and outside the home

• and to meet the challenges of birds
  that sound like smoke alarms. “I doubt
  if evolution has conspired just to give me
  this problem,” reflects Mitchell. “But it
  is a problem that has to be addressed.”
  Will Bedingfield audioanalytic.com

Left: the howl of a dog is recorded for Audio Analytic’s

sound map. Below: capturing the crack of a gunshot

Radical

Corbofuturism

Fortnite action figures

New Music Friday

Article 13

TIRED

Y Combinator’s

Oakland UBI pilot

Pokémon trading cards

Black Friday

Article 50

EXPIRED

Fully Sustainable

Equitable Capitalism

Overwatch LEGO

FridaysForFuture

Article 17

WIRED

09-19-STSoundStudio.indd 24 24/07/2019 14:00