快猫短视频

TO THE human ear, the robust tenor sax sound of Sonny Rollins is not easily
mistaken for the ethereal quality of a flute. But the differences in sound
between woodwind instruments can be quite subtle.

Now Judy Brown of Wellesley College in Massachusetts has trained a computer
to distinguish one woodwind instrument from another. Her system can correctly
identify the instrument about 80 per cent of the time鈥攚hich is about as
good as trained musicians.

One important future application of the technique, says Brown, will be to
allow Internet search engines to find music that features particular instruments
(快猫短视频, 17 March, p 34).
鈥淭his is a very difficult task,鈥� she says. 鈥淚t is at least five years away.鈥�

Working at the Massachusetts Institute of Technology鈥檚 media lab, Brown and
her colleagues at the Paris-based Institute for Musical and Acoustic Research
analysed the timbre鈥攐r sound quality鈥攐f the clarinet, flute, oboe
and saxophone.

Just as a room can resonate at a certain frequency, each of these instruments
has its own pattern of resonance. This resonance, which is a function of the
instrument鈥檚 construction, amplifies certain frequencies giving the instrument
an 鈥渁coustic fingerprint鈥�. Both the flute and the oboe, for example, tend to
amplify frequencies at around 1000 hertz, says Brown. But the oboe has an
additional bump at 1200 hertz, which gives it its nasal quality.

Very little research has been done to develop ways of recognising musical
instruments automatically, says Brown. So the researchers drew on statistical
techniques that were originally developed to recognise different voices.

One of the most promising techniques is to subtract the effect of loud and
soft notes and then examine the response of the instrument over the audible
spectrum. This reveals the instrument鈥檚 acoustic fingerprint.

The researchers then selected 25 sound samples from commercial recordings for
each of the four instruments being played unaccompanied. They then used
statistical pattern-recognition software to compare the acoustic fingerprints
from these samples with other recordings of solo woodwind instruments. The
computer was able to match the instruments in these unknown recordings with one
of the four known instruments with about 80 per cent accuracy.

In separate tests, the researchers asked 15 musicians to listen to many of
the same recordings. On average they identified 85 per cent of the instruments
correctly. Brown says that musicians should be better at identifying instruments
than untrained members of the public. 鈥淚 think computers can do as well as
humans in identifying woodwind instruments.鈥�