How does Shazam work - Coding Geek http://coding-geek.

com/how-shazam-works/

A blog about I T, program m ing and Java

Menu

HOME »» ALGORI THM »» HOW DOES SHAZAM WORK

by Chr ist ophe | updat ed: August 6, 2015 | post ed: May 23, 2015

91 Com m ent s

1,755 369

Have you ever wondered how Shazam works? I asked m yself

this quest ion a few years ago and I read a research art icle
writ t en by Avery Li- Chun Wang, t he confounder of Shazam , t o
understand the m agic behind Shazam . The quick answer is
audio fingerprint ing, which leads t o anot her question: what is
audio fingerprint ing?

When I was student, I never took a course in signal processing.

To really understand Shazam ( and not j ust have a vague idea) I
had t o st art with t he basics. This art icle is a sum m ary of t he
search I did t o understand Shazam .

I’ll st art wit h t he basics of m usic theory, present

som e signal processing stuff and end wit h t he m echanism s
behind Shazam . You don’t need any knowledge to read this
art icle but since it involves com put er science and m at hem atics
it’s bet ter t o have a good scient ific background ( especially for

1 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

the last part s) . If you already know what the words “ oct aves” ,
“ frequencies” , “ sam pling” and “ spect ral leakage” m ean you can
skip t he first part s.

Since it ’s a long and t echnical art icle ( 11k words) feel free to
read each part at different t im es.

Cont ent s [ show ]

A sound is a vibrat ion t hat propagat es t hrough air ( or water)

and can be decrypt ed by ears. For exam ple, when you list en t o
your m p3 player the earphones produce vibrat ions that
propagate t hrough air unt il t hey reach your ears. The light is
also a vibration but you can’t hear it because your ears can’t
decrypt it ( but your eyes can) .

A vibrat ion can be m odeled by sinusoidal waveform s. In t his

chapt er, we’ll see how m usic can be physically/ t echnically
described.

Pure t one s v s re a l sounds

A pure tone is a tone with a sinusoidal waveform . A sine
wave is charact erized by:

I t s frequency: t he num ber of cycles per second. I t s unit is

t he Hertz ( Hz) , for exam ple 100Hz = 100 cycles per
second.

I t s am plit ude ( relat ed to loudness for sounds) : t he size of

each cycle.

Those charact erist ics are decrypt ed by t he hum an ear t o form a

sound. Hum an can hear pure tones from 20 Hz to 20 000 Hz
( for t he best ears) and t his range decreases over age. By
com parison, t he light you see is com posed of sinewaves from
4* 10^ 14 Hz t o 7.9* 10^ 14 Hz.

2 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

You can check t he range of your ears with youtube videos like
this one t hat displays all the pure tones from 20 Hz t o 20k Hz,
in m y case I can’t hear anything above 15 kHz.

The hum an perception of loudness depends on the frequency of

the pure tone. For inst ance, a pure t one at am plit ude 10 of
frequency 30Hz will be quieter t han a pure t one at am plitude 10
of frequency 1000Hz. Hum ans ears follow a psychoacoust ic
m odel, you can check t his article on Wikipedia for m ore
inform ation.

Not e: This fun fact will have consequences at the end of the
art icle.

pure sinewave at 20 Hz

In t his figure, you can see the representation of a pure sine

wave of frequency 20hz and am plit ude 1.

Pure t ones doesn’t nat urally exist but every sound in the world
is the sum a m ult iple pure t ones at different am plit udes.

3 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

com posit ion of sinewaves

In t his figure, you can see the representation of a m ore realist ic

sound which is t he com posit ion of m ultiple sinewaves:

a pure sinewave of frequency 20hz and am plit ude 1

a pure sinewave of frequency 40hz and am plit ude 2

a pure sinewave of frequency 80hz and am plit ude 1. 5

a pure sinewave of frequency 160hz and am plit ude 1

A real sound can be com posed of thousands of pure t ones.

M usica l N ot e s

A m usic part ition is a set of notes execut ed at a cert ain

m om ent . Those notes also have a durat ion and a loudness.

The notes are divided in oct a v e s. I n m ost occidental countries,

4 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

an octave is a set of 8 not es ( A, B, C, D, E, F,G in m ost English-

speaking count ries and Do, Re, Mi, Fa, Sol, La, Si in m ost Latin
occident al count ries) wit h t he following propert y:

The frequency of a not e in an oct ave doubles in t he next

oct ave. For exam ple, t he frequency of A4 ( A in t he 4 t h
oct ave) at 440Hz equals 2 t im es the frequency of A3 ( A in
t he 3 rd octave) at 220Hz and 4 t im es t he frequency of A2
( A in t he 2 nd oct ave) at 110Hz.

Many instrum ent s provides m ore than 8 not es by octaves, those

not es are called sem it one or halfst ep.

For t he 4 t h octave ( or 3 rd oct ave in Latin occidental countries) ,

the notes have the following frequency:

C4 ( or Do3) = 261.63Hz

D4 ( or Re3) = 293.67Hz

E4 ( or Mi3) = 329. 63Hz

F4 ( or Fa3) = 349. 23Hz

G4 ( or Sol3) = 392Hz

A4 ( or La3) = 440Hz

B4 ( or Si3) = 493. 88Hz

Though it m ight be odd, t he frequency sensit ivity of ears is

logarithm ic. I t m eans t hat:

bet ween 32. 70 Hz and 61. 74Hz ( t he 1st oct ave)

or between 261. 63Hz and 466. 16Hz ( 4 t h oct ave)

or between 2 093 Hz and 3 951.07Hz ( 7t h oct ave)

Hum an ears will be able t o det ect the sam e num ber of notes.

FYI , the A4/ La3 at 440Hz is a standard reference for t he

calibrat ion of acoust ic equipm ent and m usical inst rum ent s.

Tim bre

5 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

The sam e not e doesn’t sound exact ly t he sam e if it’s played by

a guit ar, a piano, a violin or a hum an singer. The reason is that
each instrum ent has its own t im b re for a given not e.

For each inst rum ent, t he sou n d p rod u ce d is a m u lt it u d e of

fre q u e n cie s t h a t sou n d s lik e a g iv e n n ot e ( t he scientific
term for a m usical not e is p it ch ) . This sound has a
fu n d a m e n t a l frequency ( t he lowest frequency) and m ultiple
ov e rt on e s ( any frequency higher t han t he fundam ental) .

Most instrum ent s produce ( close t o) h a rm on ic sou n d s. For

those inst rum ents, t he overt ones are m ult iples of t he
fundam ental frequency called h a rm on ics. For exam ple t he
com position of pure tones A2 ( fundam ental) , A4 and A6 is
harm onic whereas the com posit ion of pure t ones A2, B3, F5 is
in h a rm on ic.

Many percussion instrum ent s ( like cym bals or drum s) creat e

inharm onic sounds.

Not e: The pit ch ( t he m usical not e perceived) m ight not be

present in t he sound played by an instrum ent . For exam ple, if
an instrum ent plays a sound wit h pure t ones A4, A6 and A8,
Hum an brain will interpret t he result ing sound has an A2 not e.
This not e/ pitch will be an A2 whereas t he lowest frequency in
the sound is A4 ( t his fact is called the m issing fundam ent al) .

Spe ct rogra m
A m usic song is played by m ultiple inst rum ents and singers. All
those inst rum ents produce a com binat ion of sinewaves at
m ultiples frequencies and t he overall is an even bigger
com binat ion of sinewaves.

It is possible t o see m usic with a spectrogram . Most of t he tim e,

a spect rogram is a 3 dim ensions graph where:

on the horizontal ( X) axis, you have the tim e,

on t he vert ical ( Y) axis you have t he frequency of t he pure

t one

t he t hird dim ension is described by a color and it

represents the am plit ude of a frequency at a cert ain tim e.

6 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

For exam ple, here is a sound of a piano playing of C4 not e

( whose fundam ental frequency is 261. 63Hz)

00:00

And here is the associat ed spect rogram :

00:00

The color represent s the am plitude in dB ( we’ll see in a next

chapt er what it m eans) .

As I t old you in the previous chapt er, t hough t he not e played is

a C4 t here are ot her frequencies t han 261Hz in t his record: t he
overt ones. What ’s int erest ing is t hat t he ot her frequencies are
m ultiple of t he first one: the piano is an exam ple of a
h a rm on ic in st ru m e n t .

Anot her int erest ing fact is t hat t he int ensit y of t he frequencies
changes t hrough tim e. I t ’s another part icularit y of an
instrum ent that m akes it unique. I f you take the sam e artist but
you replace t he piano, t he evolution of frequencies won’t
behave the sam e and t he result ing sound will be slight ly
different because each art ist/ instrum ent has its own style.
Technically speaking, these evolutions of frequencies are
m odifying t he e n v e lop e of the sound signal ( which is a part of
the t im bre) .

To give you a first idea of Shazam m usic fingerprint ing

algorithm , you can see in t his spect rogram t hat som e

7 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

frequencies ( t he lowest ones) are m ore im port ant t han others.

What if we kept j ust t he strongest ones?

Unless you’re a vinyl disk lover, when you list en t o m usic you’re
using a digit al file ( m p3, apple lossless, ogg, audio CD,
what ever ) . But when art ists produce m usic, it is analogical ( not
represent ed by bits) . The m usic is d ig it a liz e d in order t o be
st ored and played by elect ronics devices ( like com put ers,
phones, m p3 players, cd players …) . I n t his part we’ll see how
to pass from an analog sound t o a digit al one. Knowing how a
digit al m usic is m ade will help us t o analyse and m anipulat e
this digit al m usic in t he next part s.

Sa m pling
Analog signals are continuous signals, which m eans if you take
one second of an analog signal, you can divide this second into
[ put the great est num ber you can t hink of and I hope it’s a big
one ! ] part s t hat last a fract ion of second. I n t he digital world,
you can’t afford t o st ore an infinite am ount of inform at ion. You
need t o have a m inim um unit , for exam ple 1 m illisecond.
During t his unit of t im e t he sound cannot change so t his unit
needs t o be short enough so t hat the digit al song sounds like
the analog one and big enough t o lim it t he space needed for
st oring t he m usic.

For exam ple, t hink about your favorit e m usic. Now think about
it wit h the sound changing only every 2 seconds, it sounds like
nothing. Technically speaking the sound is a lia se d . In order t o
be sure that your song sounds great you can choose a very
sm all unit like a nano ( 10^ - 9) second. This tim e your m usic
sounds great but you don’t have enough disk space t o st ore it,
too bad.

This problem is called sa m p lin g .

The standard unit of t im e in digit al m usic is 44 100 unit s ( or

sa m p le s) p e r se con d . But where does t his 44,1kHz com e

8 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

from ? Well, som e dude t hought it would be good t o put 44100

unit s per second and t hat all … I ’m kidding of course.

In the first chapter I told you that hum ans can hear sounds from
20Hz to 20kHz. A theorem from Nyquist and Shannon st at es
that if you want t o digit alize a signal from 0Hz t o 20kHz you
need at least 40 000 sam ples per second. The m ain idea is that
a sine wave signal at a frequency F needs at least 2 points per
cycle to be ident ified. If t he frequency of your sam pling is at
least t wice than the frequency of your signal, you’ll end up wit h
at least 2 point s per cycle of the original signal.

Let ’s try to underst and with a picture, look at t his exam ple of a
good sam pling:

In t his figure, a sound at 20Hz is digit alized using a 40Hz

sam pling rat e:

t he blue curve represent s t he sound at 20 Hz,

t he red crosses represent t he sam pled sound, which

m eans I m arked t he blue curve wit h a red cross every 1/ 40
second,

t he green line an int erpolat ion of the sam pled sound.

Though it hasn’t the sam e shape nor t he sam e

am plit ude, t h e fre q u e n cy of t h e sa m p le d sig n a l re m a in s
t h e sa m e .

And here is an exam ple of bad sam pling :

9 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

In t his figure, a sound at 20 Hz is digitalized wit h a 30Hz

sam pling rat e. This tim e the fre q u e n cy of t h e sa m p le d
sig n a l is n ot t h e sa m e a s t h e orig in a l sig n a l: it ’s only 10Hz.
If you look carefully, you can see t hat one cycle in the sam pled
signal represent s two cycles in t he original signal. This case is
an under sam pling.

This case also shows som et hing else: if you want t o digit alize a
signal bet ween 0Hz and 20 kHz, you need rem ove from t he
signal its frequencies over 20kHz before t he sam pling.
Ot herwise those frequencies will be t ransform ed int o
frequencies bet ween 0Hz and 20Khz and t herefore add
unwanted sounds ( it’s called a lia sin g ) .

To sum up, if you want a good m usic conversion from analogic

to digital you have t o record the analog m usic at least 40000
tim es per second. HI FI corporat ions ( like Sony) chose 44,1kHz
during t he 80s because it was above 40000 Hz and com patible
wit h the video norm s NTSC and PAL. Other st andards exist for
audio like 48 kHz ( Blueray) , 96 kHz or 192 kHz but if you’re
neit her a professional nor an audiophile you’re likely t o list en t o
44. 1 kHz m usic.

Not e1: The theorem of Nyquist - Shannon is broader t han what I

said, you can check on Wikipedia if you want to know m ore
about it .

Not e2: The frequency of the sam pling rate needs t o be st rict ly
superior of 2 tim es t he frequency of t he signal to digitalize
because in the worst case scenario, you could end up wit h a

10 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

constant digitalized signal.

Qua nt iza t ion

We saw how t o digit alize t he frequencies of an analogic m usic
but what about t he loudness of m usic? The loudness is a
relat ive m easure: for t he sam e loudness inside t he signal, if
you increase your speakers the sound will be higher. The
loudness m easures t he variat ion bet ween t he lowest and t he
highest level of sound inside a song.

The sam e problem appears loudness: how to pass from a

continuous world ( with an infinite variat ion of volum e) to a
discrete one?

Im agine your favorite m usic wit h only 4 st at es of loudness: no

sound, low sound, high sound and full power. Even t he best
song in t he world becom es unbearable. What you’ve j ust
im agined was a 4- level quant ization.

Here is an exam ple of a low quantizat ion of an audio signal:

This figure present s an 8 level quantizat ion. As you can see, t he

result ing sound ( in red) is very alt ered. The difference bet ween
the real sound and t he quant ized one is called q u a n t iz a t ion
e rror or q u a n t iz a t ion n oise . Th is 8 le v e l q u a n t iz a t ion is
a lso ca lle d a 3 b it s q u a n t iz a t ion because you only need 3
bit s to im plem ent the 8 different levels ( 8 = 2^ 3) .

Here is t he sam e signal with a 64 levels quantizat ion ( or 6 bit s

11 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

quant ization)

Though t he result ing sound is still alt ered, it looks ( and sounds)
m ore like t he original sound.

Thankfully, hum ans don’t have extra sensitive ears. Th e

st a n d a rd q u a n t iz a t ion is cod e d on 1 6 b it s, which m eans
65536 levels. Wit h a 16 bits quantizat ion, the quantizat ion
noise is low enough for hum an ears.

Not e: I n st udio, the quantizat ion used by professionals is 24

bit s, which m eans t here are 2^ 24 ( 16 m illions) possible
variat ions of loudness between t he lowest point of t he track and
the highest .

Not e2: I m ade som e approxim at ions in m y exam ples

concerning the num ber of quantization levels.

Pulse Code d M odula t ion

PCM or Pulse Coded Modulat ion is a st andard that represent s
digit al signals. I t is used by com pact discs and m ost elect ronics
devices. For exam ple, when you list en t o an m p3 file in your
com put er/ phone/ t ablet , the m p3 is autom atically transform ed
int o a PCM signal and t hen send t o your headphones.

A PCM st ream is a stream of organized bit s. It can be com posed

of m ultiple channels. For exam ple, a st ereo m usic has 2

12 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

channels.

In a st ream , t he am plit ude of t he signal is divided int o

sam ples. The num ber of sam ples per second correspond to t he
sam pling rat e of t he m usic. For inst ance a 44, 1kHz sam pled
m usic will have 44100 sam ples per second. Each sam ple gives
t he ( quant ized) am plit ude of t he sound of t he corresponding
fract ion of seconds.

There are m ult iple PCM form at s but the m ost used one in audio
is the ( linear) PCM 44,1kHz, 16- bit dept h st ereo form at. This
form at has 44 100 sam ples for each second of m usic. Each
sam ple t akes 4 bytes:

2 byt es ( 16 bit s) for t he int ensit y ( from - 32, 768 t o 32, 767)
of t he left speaker

2 byt es ( 16 bit s) for t he int ensit y ( from - 32, 768 t o 32, 767)
of t he right speaker

In a PCM 44, 1kHz 16- bit depth stereo form at , you have 44100
sam ples like this one for every second of m usic.

You now know how t o pass from an analog sound to a digit al

one. But how can you get the frequencies inside a digital
signal? This part is very im portant since t he Shazam
fingerprinting algorit hm works only wit h frequencies.

For analog ( and therefore cont inuous) signals, t here is a

transform at ion called t he Cont iguous Fou rie r t ra n sform . This
funct ion t ransform s a function of t im e into a funct ion of
frequencies. I n ot her words, if you apply the Fourier transform

13 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

on a sound, it will give you t he frequencies ( and their

int ensit ies) inside t his sound.

But there are 2 problem s:

We are dealing with digital sounds and t herefore finit e

( none continuous) sounds.

To have a bet ter knowledge of t he frequencies inside a

m usic, we need to apply t he Fourier Transform on sm all
part s of the full lengt h audio signal, like 0. 1 second part s
so that we know what are t he frequencies for each 0. 1
second part s of an audio t rack) .

Thankfully, there is another m at hem atical funct ion,

the Discre t e Fou rie r Tra n sform ( DFT) , that works with som e
lim it at ions.

Not e: The Fourier Transform m ust be applied on only one

channel, which m eans that if you have a stereo song you need
to t ransform it int o a m ono song.

Discre t e Fourie r Tra nsform

The DFT ( Discret e Fourier Transform ) applies t o discret e signals
and gives a discret e spectrum ( t he frequencies inside the
signal) .

Here is t he m agic form ula t o transform a digital signal int o

frequencies ( don’t run away, I ’ll explain it ) :

In t his form ula:

N is t he size of the w in d ow : t he num ber of sam ples t hat

com posed t he signal ( we’ll talk a lot about windows in t he
next part) .

X( n) represents t he nt h b in of fre q u e n cie s

14 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

x( k) is kt h sam ple of t he audio signal

For exam ple, for an audio signal with a 4096- sam ple window
this form ula m ust be applied 4096 t im es:

1 t im e for n = 0 t o com put e t he 0 t h bin a frequencies

1 t im e for n = 1 t o com put e t he 1 st bin a frequencies

1 t im e for n = 2 t o com put e t he 2 nd bin a frequencies

As you m ight have noticed, I spoke about bin of frequencies and

not frequency. The reason is that t he DFT gives a d iscre t e
sp e ct ru m . A bin of frequencies is t he sm allest unit of
frequency the DFT can com put e. The size of the bin ( called
sp e ct ra l/ sp e ct ru m re solu t ion or fre q u e n cy re solu t ion )
equals t he sam pling rat e of t he signal divided by t he size of t he
window ( N) . I n our exam ple, wit h a 4096- sam ple window and a
st andard audio sam pling rat e at 44. 1kHz, t he frequency
resolut ion is 10.77 Hz ( except the 0 t h bin t hat is special) :

t he 0t h bin represent s the frequencies bet ween 0Hz t o

5.38Hz

t he 1st bin represent s t he frequencies bet ween 5. 38Hz t o

16. 15Hz

the 2nd bin represents the frequencies between 16. 15Hz to

26. 92Hz

t he 3rd bin represent s t he frequencies bet ween 26.92Hz t o

37. 68Hz

That m eans t hat the DFT can’t dissociate 2 frequencies that are
closer t han 10.77Hz. For exam ple notes at 27Hz, 32Hz and
37Hz ends up in t he sam e bin. I f t he note at 37Hz is very
powerful you’ll j ust know that the 3 rd bin is powerful. This is
problem at ic for dissociating not es in the lowest oct aves. For
exam ple:

15 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

a A1 ( or La - 1) is at 55Hz whereas a B1 ( or Si - 1) is at
58. 27Hz and a G1 ( or Sol - 1) is at 49 Hz.

t he first note of a st andard 88- key piano is a A0 at 27. 5 Hz

followed by a A# 0 at 29. 14Hz.

You can im prove the frequency resolut ion by increasing the

window size but t hat m eans losing fast frequency/ not e changes
inside the m usic:

An audio signal has a sam pling rat e of 44, 1 kHz

I ncreasing t he window m eans t aking m ore sam ples and

t herefore increasing the t im e t aken by t he window.

Wit h 4096 sam ples, t he window duration is 0.1 sec and t he

frequency resolut ion is 10. 7 Hz: you can detect a change
every 0.1 sec.

Wit h 16384 sam ples, the window durat ion is 0. 37 sec and
the frequency resolution is 2.7 Hz: you can detect a
change every 0. 37 sec.

Anot her part icularit y for an audio signal is t hat w e on ly n e e d

h a lf t h e b in s com p u t e d b y t h e DFT. In t he previous
exam ple, the bin definition is 10. 7 Hz, which m eans t hat the
2047 t h bin represents t he frequencies from 21902, 9 Hz t o
21913, 6 Hz. But :

The 2048 t h bin will give t he sam e inform at ion as t he 0 t h

bin

The 2049 t h bin will give t he sam e inform at ion as t he 1t h

bin

The X+ 2048 t h bin will give the sam e inform ation as the Xt h
bin

..

If you want t o know why the bin resolut ion equals” t he sam pling
rat e” divided by “ t he size of t he window” or why this form ula is
so bizarre, you can read a 5- part art icle on Fourier Transform
on this very good websit e ( especially part 4 and part 5) which
is the best art icle for beginners t hat I read ( and I read a lot of

16 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

art icles on the m at t er) .

W indow funct ions

If you want to get the frequency of a one- second sound for each
0. 1- second part s, you have to apply t he Fourier Transform for
the first 0.1- second part, apply it for the second 0.1- second
part , apply it for t he t hird 0. 1- second part …

The problem

By doing so, you are im plicit ly applying a ( rect angular) window

funct ion:

For t he first 0. 1 second you are applying t he Fourier

t ransform on t he full one- second signal m ult iplied by a
funct ion t hat equals 1 bet ween 0 and 0.1second, and 0 for
t he rest

For t he second 0. 1 second you are applying t he Fourier

t ransform on t he full one- second signal m ult iplied by a
funct ion t hat equals 1 bet ween 0.1 and 0.2 second, and 0
for t he rest

For the third 0.1 second you are applying the Fourier
t ransform on t he full one- second signal m ult iplied by a
funct ion t hat equals 1 bet ween 0.2 and 0.3 second, and 0
for t he rest

Here is a visual exam ple of the window function t o apply t o a

digit al ( sam pled) audio signal t o get t he first 0. 01- second part :

17 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

In t his figure, t o get t he frequencies for t he first 0. 01- second

part , you need to m ult iply t he sam pled audio signal ( in blue)
with the window function ( in green) .

In t his figure, t o get t he frequencies for t he second 0. 01- second

part , you need to m ult iply t he sam pled audio signal ( in blue)
with the window function ( in green) .

By “ windowing” the audio signal, you m ultiply your signal

audio( t ) by a window funct ion window( t) . This window function
produces sp e ct ra l le a k a g e . Spect ral leakage is the apparition
of new frequencies t hat doesn’t exist inside the audio
signal. The power of the real frequencies is leaked t o others
frequencies.

Here is a non- form al ( and very light ) m at hem atical

explanat ion. Let ’s assum e you want a part of the full audio
signal. You will m ultiply t he audio signal wit h a window funct ion
that let pass the sound only for the part you want :

part _of_audio( t ) = full_audio( t ) . window ( t )

When you t ry t o get t he frequencies of the part of audio, you

apply t he Fourier t ransform on the signal

Fourier( part_of_audio( t ) ) = Fourier( full_audio( t ) . window ( t ) )

According t o the convolut ion t heorem ( * represent s t he

convolution operator and . t he m ult iplicat ion operat or)

Fourier( full_audio( t) . window ( t ) ) = Fourier( full_audio( t) ) *

18 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

Fourier( window ( t ) )

—> Fourier( part _of_audio( t ) ) = Fourier( full_audio( t) ) *

Fourier( window ( t ) )

—> The frequencies of the part _of_audio( t) depend on t he

window( ) function used.

I won’t go deeper because it requires advanced m athem at ics. I f

you want t o know m ore, look at this link on page 29, the
chapt er “ t he truncat e effect s” present s t he m athem at ical effect
of applying a rect angular window on a signal. What you need t o
keep in m ind is t hat cutt ing an audio signal int o sm all part s to
analyze t he frequencies of each part produces spectral leakage.

different types of windows

You ca n ’t a v oid sp e ct ra l le a k a g e b u t y ou ca n h a n d le h ow
t h e le a k a g e w ill b e h a v e by choosing t he right window
funct ion: inst ead of using a rect angular window funct ion, you
can choose a t riangular widows, a Parzen window, a Blackm an
window, a Ham m ing window …

The rect angular window is t he easiest window t o use ( because

you j ust have to “ cut” the audio signal into sm all part s) but for
analyzing t he m ost im port ant frequencies in a signal, it m ight
not be t he best t ype of windows. Let ’s have a look of 3 types of
windows: rect angular, Ham m ing and Blackm an. I n order to
analyse t he effect of the 3 windows, we will use t he following
audio signal com posed of:

A frequency 40 Hz wit h an am plit ude of 2

A frequency 160 Hz with an am plit ude of 0.5

A frequency 320 Hz with an am plit ude of 8

A frequency 640Hz wit h an am plit ude of 1

A frequency 1000 Hz wit h an am plitude of 1

A frequency 1225 Hz wit h an am plitude of0.25

A frequency 1400 Hz wit h an am plitude of 0. 125

A frequency 2000 Hz wit h an am plitude of 0. 125

A frequency 2500Hz with an am plit ude of 1.5

19 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

In a perfect world, t he Fourier t ransform of this signal should

give us t he following spect rum :

This figure shows a spectrum wit h only 9 vertical lines ( at 40

Hz, 160 Hz, 320 Hz, 640 Hz, 1000 Hz, 1225 Hz, 1400 Hz, 2000
Hz and 2500 Hz. The y axis gives the am plitude in decibels ( dB)
which m eans the scale is logarithm ic. Wit h t his scale a sound at
60 dB is 100 t im es m ore powerful t han a sound at 40 dB and
10000 tim es m ore powerful than a sound at 20 dB. To give you
an idea, when you speak in a quiet room , t he sound you
produce is 20- 30 dB higher ( at 1 m of you) than the sound of
the room .

In order t o plot t his “ perfect” spect rum , I applied the Fourier

Transform with a very long window: a 10- second window. Using
a very long window reduces t he spectrum leakage but 10
seconds is too long because in a real song the sound changes
m uch fast er. To give you an idea of how fast t he m usic
changes:

here is a video wit h 1 change ( or beat ) per second, it

sounds slow but it ’s a com m on rhyt hm for classical m usic.

here is a video wit h 2. 7 changes per second, it sounds

m uch fast er but t his rhyt hm is com m on for elect ro m usic

here is a video wit h 8. 3 changes per second, it ’s a very

( very) fast rhyt hm but possible for sm all parts of songs.

In order t o capt ure t hose fast changes, you need t o “ cut ” t he

sound int o very sm all part s using window funct ions. I m agine

20 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

you want to analyze the frequencies of a sound every 1/ 3

second.

In t his figure, you can m ult iply t he audio signal wit h one of t he
3 window t ypes t o get t he part of the signal bet ween 0. 333sec
and 0. 666 sec. As I said, using a rectangular window is like
cutt ing t he signal bet ween 0. 333sec and 0.666sec whereas wit h
the Ham m ing or t he Blackm an windows you need t o m ult iply
the signal wit h t he window signal.

Now, here is t he spect rum of t he previous audio signal wit h a

4096- sam ple window:

The signal is sam pled at 44100Hz so a 4096- sam ple window

represent s a 93- m illisecond part ( 4096/ 44100) and a frequency
resolut ion of 10.7 Hz.

21 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

This figure shows t hat all windows m odify t he real spect rum of
the sound. We clearly see t hat a part of t h e p ow e r of t h e re a l
fre q u e n cie s is sp re a d t o t h e ir n e ig h b ou rs. The spect rum
from t he rect angular window is t he worst since t he spectrum
leakage is m uch higher than the 2 ot hers. I t ’s especially true
bet ween 40 and 160 Hz. The Blackm an window gives the
closest spect rum from t he real spectrum .

Here is t he sam e exam ple wit h a Fourier Transform of a 1024

window:

The signal is sam pled at 44100Hz so a 1024- sam ple window

represent s a 23- m illisecond part ( 1024/ 44100) and a frequency
resolut ion of 43 Hz.

This t im e t he rect angular window gives t he best spect rum . With

the 3 windows t he 160 Hz frequency is hidden by t he spect rum
leakage produced by t he 40 Hz and 320 Hz frequencies. The
Blackm an window gives t he worst result with a 1225 Hz
frequency close t o invisible.

Com paring bot h figures shows that t he spectrum leakage

increases ( for all t he window funct ion) as the frequency
resolut ion increases. The fingerprint algorithm used by Shazam
look for the loudest frequencies inside an audio t rack. Because
of spectrum leakage, we can’t j ust t ake t he X highest
frequencies. I n the last exam ple, t he 3 loudest frequencies are
approxim at ely 320 Hz, 277 Hz ( 320- 43) and 363 Hz ( 320+ 43)
whereas only the 320 Hz frequency exist s.

22 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

Which window is the best ?

There are no “ best ” or “ worst ” windows. Each window has its

specificities and depending on t he problem you m ight want to
use a cert ain type.

A rectangular window has excellent resolut ion charact erist ics

for sinusoids of com parable strengt h, but it is a poor choice for
sinusoids of disparat e am plit udes ( which is t he case inside a
song because t he m usical not es don’t have the sam e loudness) .

Windows like Blackm an are bet ter t o prevent from t he case

where spectrum leakage of st rong frequencies hides weak
frequencies. But , t hese windows deal badly wit h noise since a
noise will hide m ore frequencies than rectangular window. This
is problem at ic for an algorithm like Shazam that needs t o
handle noise ( for inst ance when you Shazam a m usic in a bar or
out door t here are a lot of noise) .

A Ham m ing window is bet ween these t wo extrem es and is ( in

m y opinion) a bet t er choice for an algorit hm like shazam .

Here are som e useful links t o go deeper on window functions

and spect rum leakage:

ht tp: / / en. wikipedia. org/ wiki/ Spectral_leakage

ht tp: / / en. wikipedia. org/ wiki/ Window_funct ion

ht tp: / / web. m it . edu/ xiphm ont / Public/ windows. pdf

Fa st Fourie r Tra nsform a nd t im e com ple x it y

the problem

23 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

If you look again at t he DFT form ula ( don’t worry, it ’s t he last

tim e you see it ) , you can see that t o com pute one bin you need
to do N additions and N m ult iplicat ions ( where N is t he size of
the window) . Gett ing t he N bins requires 2 * N^ 2 operat ions
which is a lot .

For exam ple, let’s assum e you have a three- m inut e song at
44, 1 kHz and you com pute t he spectrogram of t he song with a
4096- sam ple window. You’ll have t o com pute 10.7
( 44100/ 4096) DFT per second so 1938 DFTs for the full song.
Each DFT needs 3. 35* 10^ 7 operat ions ( 2* 4096^ 2) . To get t he
spectrogram of t he song you need to do 6, 5* 10^ 10 operat ions.

Le t ’s a ssu m e y ou h a v e a m u sic colle ct ion of 1 0 0 0 t h re e -

m in u t e s- lon g son g s, you’ll need 6,5* 10^ 13 operations to get
the spect rogram s of your songs. Even wit h a good processor, it
w ou ld t a k e d a y s/ m on t h s t o g e t t h e re su lt .

Thankfully, there are fast er im plem ent at ions of t he DFT called

FFT ( Fast Fourier Transform s) . Som e im plem ent at ions require
j ust 1. 5* N * log( N) operat ions. For t he sam e m usic collection,
u sin g t h e FFT in st e a d of t h e DFT requires 340 tim es less
additions ( 1.43* 10^ 11) and it w ou ld t a k e m in u t e s/ h ou rs t o
g e t t h e re su lt .

This exam ple shows another t radeoff: though increasing the

size of the window im proves the frequency resolut ion, it also
increases t he com put at ion t im e. For t he sam e m usic collect ion,
if you com pute t he spectrogram using a 512 sam ple window
( frequency resolution of 86 Hz) , you get t he result wit h t he FFT
in 1.07* 10^ 11 operat ions, approxim at ely 1/ 4 t im e fast er than
wit h a 4096 sam ple window ( frequency resolut ion of 10.77 Hz) .

This t im e com plexity is im port ant since when you shazam a

sound, your phone needs to com pute the spectrogram of the
recorded audio and a m obile processor is less powerful t han a
deskt op processor.

downsam pling

Thankfully, there is a t rick t o keep the frequency resolut ion and

reduce t he window size at the sam e tim e, it ’s called

24 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

downsam pling. Let ’s t ake a st andard song at 44100 Hz, if you

resam ple it at 11025 Hz ( 44100/ 4) you will get t he sam e
frequency resolut ion whet her you do a FFT on t he 44.1kHz song
with a 4096 window or you do a FFT on the 11kHz resam pled
song wit h a 1024 window. The only difference is t hat t he
resam pled song will only have frequencies from 0 to 5 kHz. But
the m ost im port ant part of a song is between 0 and 5kHz. In
fact m ost of you won’t hear a big difference bet ween a m usic at
11kHz and a m usic at 44. 1kHz. So, t he m ost im port ant
frequencies are st ill in t he resam pled song which is what
m att ers for an algorit hm like Shazam .

Downsam pling a 44.1 kHz song t o a 11. 025 kHz one is not very
difficult: A sim ple way to do it is t o t ake t he sam ples by group
of 4 and to t ransform this group into j ust one sam ple by t aking
the average of t he 4 sam ples. The only t ricky part is t hat before
downsam pling a signal, you need t o filt er t he higher
frequencies in t he sound t o avoid aliasing ( rem em ber the
Nyquist - Shannon t heorem ) . This can be done by using a digit al
low pass filter.

FFT

But let ’s go back t o the FFT. The sim plest im plem ent at ion of t he
FFT is t he radix 2 Cooley–Tukey algorit hm which is a divide a
conquer algorit hm . The idea is that inst ead of direct ly
com put ing the Fourier Transform on t he N- sam ple window, t he
algorithm :

divides t he N- sam ple window int o 2 N/ 2- sam ple windows

com putes ( recursively) t he FFT for t he 2 N/ 2- sam ple

windows

25 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

com putes efficient ly t he FFT for t he N- sam ple windows

from t he 2 previous FFT

The last part only cost s N operat ions using a m at hem at ical t rick
on the root s of unit y ( the exponent ial term s) .

Here is a readable version of t he FFT ( writ t en in python) t hat I

found on Wikipedia

from cmath import *

def fft(x):
N=len(x)
if N==1: return x

even=fft([x[k] for k in range(0,N,2)])

odd= fft([x[k] for k in range(1,N,2)])

M=N/2
l=[ even[k] + exp(‐2j*pi*k/N)*odd[k] for k in ran
r=[ even[k] ‐ exp(‐2j*pi*k/N)*odd[k] for k in ran

return l+r

For m ore inform at ion on the FFT, you can check this art icle on
Wikipedia.

We’ve seen a lot of st uff during t he previous parts. Now, we’ll

put everyt hing t ogether t o explain how Shazam quickly
identifies songs ( at last ! ) . I ’ll first give you a global overview of
Shazam , then I’ll focus on t he generation of t he fingerprint s and
I’ll finish wit h the efficient audio search m echanism .

Not e: From now on, I assum e t hat you read the part s on
m usical not es, FFT and window funct ions. I’ll som etim es use
the words “ frequency” , “ bin” , ” not e” or t he full expression “ bin
of frequencies” but it’s t he sam e concept since we’re dealing
wit h digit al audio signals.

Globa l ov e rv ie w
An a u d io fin g e rp rin t is a digit al sum m ary t hat can be used t o
identify an audio sam ple or quickly locat e sim ilar it em s in an

26 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

audio database. For exam ple, when you’re hum m ing a song to
som eone, you’re creating a fingerprint because you’re
ext racting from t he m usic what you think is essential ( and if
you’re a good singer, t he person will recognize t he song) .

Before going deeper, here is a sim plified architect ure of what

Shazam m ight be. I don’t work at Shazam so it ’s only a guess
( from the 2003 paper of the co- founder of Shazam ) :

On t he server side:

Shazam precom putes fingerprints from a very big dat abase

of m usic t racks.

All t hose fingerprint s are put in a fingerprint dat abase

which is updat ed whenever a new song is added in t he
song dat abase

On t he client side:

when a user uses t he Shazam app, t he app first records

t he current m usic wit h t he phone m icrophone

t he phone applies t he sam e fingerprint ing algorit hm

as Shazam on t he record

t he phone sends the fingerprint t o Shazam

Shazam checks if t his fingerprint m at ches wit h one of its

fingerprint s
If no it inform s the user t hat t he m usic can’t be found

If yes, it looks for the m et adat a associat ed with t he

fingerprints ( nam e of the song, I Tunes url, Am azon url

27 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

…) and gives it back to t he user.

The key points of Shazam are:

being Noise/ Fault tolerant:

because t he m usic recorded by a phone in a
bar/ out door has a bad qualit y,

because of t he art ifact due t o window funct ions,

because of t he cheap m icrophone inside a phone t hat

produces noise/ dist ort ion

because of m any physical st uff I ’m not aware of

fingerprint s needs t o be tim e invariant : t he fingerprint of a

full song m ust be able t o m atch wit h j ust a 10- second
record of t he song

fingerprint m at ching need t o be fast : who wants t o wait

m inutes/ hours to get an answer from Shazam ?

having few false posit ives: who want s t o get an answer

t hat doesn’t correspond t o the right song?

Spe ct rogra m filt e ring

Audio fingerprints differ from st andard com put er fingerprints
like SSHA or MD5 because t wo different files ( in term s of bits)
that cont ain t he sam e m usic m ust have the sam e audio
fingerprint. For exam ple a song in a 256kbit ACC form at
( I Tunes) m ust give t he sam e fingerprint as t he sam e song in a
256kbit MP3 form at ( Am azon) or in a 128kbit WMA form at
( Microsoft ) . To solve t his problem , a u d io fin g e rp rin t in g
a lg orit h m s u se s t h e sp e ct rog ra m of audio signals t o
ext ract fingerprints.

Get t ing our spectrogram

I told you before t hat to get the spect rogram of a digit al sound
you need to apply a FFT. For a fingerprinting algorithm we need
a good frequency resolution ( like 10. 7Hz) to reduce spect rum
leakage and have a good idea of t he m ost im portant notes
played inside t he song. At t he sam e t im e, we need t o reduce
the com putation t im e as far as possible and t herefore use t he

28 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

lowest possible window size. I n t he research paper from

Shazam , they don’t explain how they get t he spectrogram but
here is a possible solut ion:

On t he server side ( Shazam ) , the 44. 1khz sam pled sound ( from
CD, MP3 or what ever sound form at) needs t o pass from st ereo
to m ono. We can do t hat by t aking t he average of the left
speaker and t he right one. Before downsam pling, we need t o
filt er t he frequencies above 5kHz t o avoid aliasing. Then, t he
sound can be downsam pled at 11. 025kHz.

On t he client side ( phone) , t he sam pling rat e of t he m icrophone

that records the sound needs t o be at 11. 025 kHz.

Then, in bot h cases we need to apply a window funct ion t o t he

signal ( like a ham m ing 1024- sam ple window, read t he chapter
on window funct ion to see why) and apply t he FFT for every
1024 sam ples. By doing so, each FFT analyses 0. 1 second of
m usic. This gives us a spect rogram :

from 0 Hz t o 5000Hz

wit h a bin size of 10.7Hz,

512 possible frequencies

and a unit of t im e of 0. 1 second.

Filtering

At t his st age we have the spect rogram of t he song. Since

Shazam needs t o be noise tolerant, on ly t h e lou d e st n ot e s
a re k e p t . But you can’t j ust keep t he X m ore powerful
frequencies every 0. 1 second. Here are som e reasons:

I n the beginning of t he art icle I spoke about

psychoacoust ic m odels. Hum an ears have m ore difficulties

29 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

t o hear a low sound ( < 500Hz) t han a m id- sound ( 500Hz-

2000Hz) or a high sound ( > 2000Hz) . As a result low
sounds of m any “ raw” songs are art ificially increased
before being released. I f you only take t he m ost powerful
frequencies you’ll end up wit h only the low ones and I f 2
songs have t he sam e drum partit ion, they m ight have a
very close filtered spectrogram whereas t here are flut es in
t he first song and guit ars in the second.

We saw on t he chapt er on window funct ions t hat if you

have a very powerful frequency ot her powerful frequencies
close to t his one will appeared on t he spectrum whereas
t hey doesn’t exist ( because of spectrum leakage) . You
m ust be able t o only take t he real one.

Here is a sim ple way t o keep only strong frequencies while

reducing t he previous problem s:

st ep1 – For each FFT result, you put t he 512 bins you inside 6
logarithm ic bands:

the very low sound band ( from bin 0 to 10)

t he low sound band ( from bin 10 t o 20)

the low- m id sound band ( from bin 20 to 40)

t he m id sound band ( from bin 40 t o 80)

t he m id- high sound band ( from bin 80 to 160)

t he high sound band ( from bin 160 to 511)

st ep2 – For each band you keep the st rongest bin of

frequencies.

st ep3 – You t hen com pute t he average value of t hese 6

powerful bins.

st ep4 – You keep the bins ( from t he 6 ones) that are above this
m ean ( m ult iplied by a coefficient) .

The step4 is very im portant because you m ight have:

an a cappella m usic involving soprano singers wit h only

m id or m id- high frequencies

a j azz/ rap m usic with only low and low- m id frequencies

30 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

..

And you don’t want to keep a weak frequency in a band j ust

because t his frequency is t he st rongest of it s band.

But this algorithm has a lim itation. I n m ost songs som e parts
are very weak ( like t he beginning or t he end of a song) . If you
analyze t hese part s you’ll end up wit h false st rong frequencies
because t he m ean value ( com puted at st ep 3) of t hese part s is
very low. To avoid that , instead of taking the m ean of t he 6
powerful beans of the current FFT ( t hat represents only 0.1sec
of the song) you could take t he m ean of t he m ost powerful bins
of t he full song.

To sum m arize, by applying t his algorit hm we’re filt ering t he

spectrogram of t he song t o keep the peaks of energy in the
spectrum that represent t he loudest notes. To give you a visual
idea of what t his filtering is, here is a real spect rogram of a
14- second song.

This figure is from t he Shazam research art icle. In t his

spectrogram , you can see t hat som e frequencies are m ore
powerful t han others. If you apply t he previous algorit hm on t he
spectrogram here is what you’ll get:

31 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

This figure ( st ill from the Shazam research art icle) is a filt ered
spectrogram . Only t he strongest frequencies from t he previous
figure are kept . Som e part s of the song have no frequency ( for
exam ple between 4 and 4.5 seconds) .

The num ber of frequencies in t he filt ered spect rogram depends

on the coefficient used wit h the m ean during step4. I t also
depends on the num ber of bands you use ( we used 6 bands but
we could have used anot her num ber) .

At t his st age, the intensity of the frequencies is useless.

Therefore, t his spect rogram can m odeled as a 2- colum n t able
where

t he first colum n represent s the frequency inside t he

spect rogram ( t he Y axis)

t he second colum n represent s t he tim e when the

frequency occurred during t he song ( t he X axis)

This filtered spectrogram is not t he final fingerprint but it’s a

huge part of it . Read t he next chapt er t o know m ore.

Not e: I gave you a sim ple algorit hm t o filt er t he spect rogram . A

bet t er approach could be t o use a logarit hm ic sliding window
and t o keep only the m ost powerful frequencies above t he
m ean + the st andard deviation ( m ultiplied by a coefficient ) of a
m oving part of t he song. I used t his approach when I did m y
own Shazam prot ot ype but it ’s m ore difficult t o explain ( and
I’m not even sure t hat what I did was correct …) .

32 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

St oring Finge rprint s

We’ve j ust ended up with a filt ered spect rogram of a song. How
can we st ore and use it in an efficient way? This part is where
the power of Shazam lies. To understand the problem , I ’ll
present a sim ple approach where I search for a song by using
direct ly t he filt ered spect rogram s.

Sim ple search approach

Pre- st ep: I precom put e a dat abase of filt ered spect rogram s for
all t he songs in m y com puter

St ep 1: I record a 10- second part of a song from TV in m y

com put er

St ep 2: I com put e the filtered spectrogram of t his record

St ep 3: I com pare t his “ sm all” spectrogram wit h t he “ full”

spectrogram of each songs. How can I com pare a 10- second
spectrogram wit h a spect rogram of a 180- second song? I nst ead
of losing m yself in a bad explanat ion, here is a visual
explanat ion of what I need t o do.

Visually speaking, I need t o superpose t he sm all spectrogram

everywhere inside t he spect rogram of the full song t o check if
the sm all spect rogram m at ches with a part of t he full one.

33 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

And I need t o do t his for each song unt il I find a perfect m at ch.

In t his exam ple, t here is a perfect m atch bet ween t he record

and t he end of t he song. I f it’s not t he case, I have t o com pare
t he record wit h anot her song and so on unt il I find a perfect
m atch. I f I don’t find a perfect m atch I can choose t he closest
m at ch I found ( in all t he songs) if t he m at ching rat e is above a
threshold. For inst ance, if t he best m at ch I found gives m e a
90% sim ilarit y between t he record and a part of a song, I
can assum e it’s the right song because the 10% of none
sim ilarit y are cert ainly due t o ext ernal noise.

Though it works well, t his sim ple approach requires a lot of

com put at ion tim e. I t needs t o com put e all t he possibilit ies of
m atching bet ween the 10- second record and each song in t he
collection. Let ’s assum e on average m usic cont ains 3 peak
frequencies per 0. 1 seconds. Therefore, t he filt ered

34 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

spectrogram of t he 10- second record has 300 t im e- frequency

points. I n t he worst case scenario, you’ll need 300 * 300 * 30*
S operat ions t o find t he right song where S is t he num ber of
second of m usic in your collect ion. I f like m e you have 30k
songs ( 7 * 10^ 6 seconds of m usic) it m ight t ake a long tim e
and it ’s harder for Shazam wit h its 40 m illion songs collection
( it ’s a guess I couldn’t find t he current size of Shazam ) .

So, how Shazam does it efficient ly?

Target zones

Inst ead of com paring each point one by one, the idea is to look
for m ultiple point s at t he sam e t im e. I n t he Shazam paper, this
group of point is called a t a rg e t z on e . The paper from Shazam
doesn’t explain how t o generat e t hese t arget zones but here is
a possibility. For t he sake of com prehension I ’ll fix t he size of
t he t arget zone at 5 frequency- t im e point s.

In order t o be sure t hat bot h the record and the full song will
generat e the sam e target zones, you need an order relat ion
bet ween t he tim e- frequency points in a filt ered spect rogram .
Here is one:

I f two tim e- frequency points have the sam e tim e, t he

t im e- frequency point wit h t he lowest frequency is before
t he ot her one.

I f a t im e t im e- frequency point has a lower t im e t han

anot her point one t hen it is before.

Here is what you get if you apply t his order on t he sim plified
spectrogram we saw before:

35 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

In t his figure I labeled all t he t im e- frequency point s using this

order relation. For exam ple:

The point 0 is before any other point s in the spect rogram .

The point 2 is aft er point 0 and 1 but before all t he ot hers.

Now t hat the spect rogram s can be inner- ordered, we can

creat e t he sam e t arget zones on different spectrogram wit h t he
following rule: “ To generate t arget zones in a spectrogram , you
need for each t im e- frequency point t o creat e a group com posed
of this point and t he 4 point s after it ” . We’ll end up with
approxim at ely the sam e am ount of target zones as the num ber
of points. This generat ion is t he sam e for t he songs or t he
record

In t his sim plified spect rogram , you can see the different t arget
zones generat ed by t he previous algorit hm . Since t he t arget
size is 5, m ost of the point s belong t o 5 t arget zones ( except
the point s at t he beginning and t he end of t he spectrogram ) .

Not e: I didn’t underst and at first why for the record we needed
to com pute that m uch target zones. We could generate target
zones wit h a rule like “ for each point whose label is a m ultiple
of 5 you need t o create a group com posed of t his frequency and
the 4 frequencies after it ” . Wit h t his rule, t he num ber of t arget
zones would be reduced by 5 and so t he search t im e ( explained
in t he next part ) . The only reason I found is t hat com puting all
the possible zones on both t he record and t he song increases a
lot t he noise robust ness.

36 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

Address generation

We now have m ult iple t arget zones, what do we do next ? We

creat e for each point an a d d re ss based on those t arget zones.
In order t o creat e t hose addresses, we also need an a n ch or
p oin t per t arget zone. Again, t he paper doesn’t explain how t o
do it. I propose t his anchor point t o be the 3 rd point before t he
target zone. The anchor can be anywhere as long as the way it
is generat ed is reproducible ( which it is t hanks to our order
relat ion) .

In t his pict ure I plot t ed 2 t arget zones wit h their anchor points.
Let ’s focus on t he purple target zone. The address form ula
proposed by shazam is following one:

[ “ frequency of t he anchor” ; ” frequency of the point” ; ” delt a

tim e bet ween t he anchor and t he point ” ] .

For t he purple target zone:

t he address of point 6 is [ “ frequency of 3” ; ” frequency of

point 6” ; ” delta_t im e between point 3 & point 6” ] so
concretely [ 10; 30; 1] ,

t he address of point 7 is [ 10; 20; 2] .

Both point s appeared also in the brown t arget zone, their

addresses wit h this target zone are [ 10; 30; 2] for point 6 and
[ 10; 20; 3] for point 7.

37 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

I spoke about addresses, right ? That m eans that t hose

addresses are linked t o som et hing. I n t he case of t he full songs
( so only in t he server side) , t hose addresses are linked t o t he
following couple [ “ absolut e tim e of the anchor in t he song” ; ” Id
of the song” ] . In our sim ple exam ple wit h the 2 previous point s
we have t he following result :

[ 10; 30; 1] –> [ 2; 1]

[ 10; 30; 2] –> [ 2; 1]

[ 10; 30; 2] –> [ 1; 1]

[ 10; 30; 3] –> [ 1; 1]

If you apply t he sam e logic for all the point s of all the t arget
zones of all the song spectrogram s, you’ll end up with a very
big table wit h 2 colum ns:

t he addresses

t he couples ( “ tim e of anchor” ; “ song I d” ) .

Th is t a b le is t h e fin g e rp rin t d a t a b a se of Sh a z a m . I f on
average a song contains 30 peak frequencies per second and
the size of t he t arget zone is 5, t he size of t his t able is 5 * 30
* S where S is t he num ber of seconds of the m usic collect ion.

If you rem em ber, we used an FFT wit h 1024 sam ples which
m eans t hat there are only 512 possible frequency values. Those
frequencies can be coded in 9 bit s ( 2^ 9 = 512) . Assum ing t hat
the delt a t im e is in m illiseconds, it will never be over 16
seconds because it would im ply a song wit h a 16- second part
wit hout m usic ( or very low sound) . So, t he delta t im e can be
coded in 14 bit s ( 2^ 14 = 16384) . Th e a d d re ss ca n b e cod e d
in a 3 2 - b it in t e g e r:

9 bit s for t he “ frequency of t he anchor”

9 bit s for t he ” frequency of t he point ”

14 bits for t he ” delt a t im e bet ween the anchor and t he

point ”

Using t he sam e logic, t h e cou p le ( “t im e of a n ch or” ; “son g

I d ”) ca n b e cod e d in a 6 4 - b it in t e g e r ( 32 bit for each part ) .

The fingerprint t able can be im plem ent ed as a sim ple array of

38 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

list of 64- bit int egers where:

t he index of t he array is t he 32- bit int eger address

t he list of 64- bit s int egers is all t he couples for t his address
.

In ot her words, we t ransform ed the fin g e rp rin t t a b le into an

in v e rt e d look - u p t hat allows search operation in O( 1) ( ie.
very effect ive search t im e) .

Note: You m ay have noticed that I didn’t choose the anchor

point inside the t arget zone ( I could have chosen the first point
of the t arget Zone for exam ple) . I f I did it would have
generat ed a lot of addresses like [ frequency anchor; frequency
anchor; 0] and t herefore t oo m any couples( “ t im e of anchor” ;
“ song I d” ) would have an address like [ Y,Y,0] where Y is t he
frequency ( between 0 and 511) . In ot her words, t he look- up
would have been skewed.

Se a rching And Scoring t he finge rprint s

We now have a great data struct ure on t he server side, how
can we use it ? I t’s m y last quest ion, I prom ise!

Search

To perform a search, t he fingerprinting st ep is perform ed on the

recorded sound file to generate an address/ value st ructure
slight ly different on t he value side:

[ “ frequency of t he anchor” ; ” frequency of the point” ; ” delt a

tim e bet ween t he anchor and t he point ” ] - > [ “ absolute t im e of
the anchor in t he record” ] .

This dat a is t hen sent to t he server side ( Shazam ) . Let ’s take

the sam e assum ption t han before ( 300 t im e- frequency point s in
the filtered spectrogram of t he 10- second record and t he size of
the t arget zone of 5 points) , it m eans t here are approxim at ely
1500 dat a sent to Shazam .

Each address from t he record is used t o search in the

39 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

fingerprint dat abase for t he associat ed couples [ “ absolut e tim e

of the anchor in t he song” ; ” Id of t he song” ] . I n t erm s of t im e
com plexit y, assum ing t hat the fingerprint dat abase is
in- m em ory, t he cost is the search is proport ional t o the num ber
of address sent t o Shazam ( 1500 in our case) . This search
ret urns a big am ount of couples, let ’s say for t he rest of the
art icle it ret urns M couples.

Though M is huge, it’s way lower t han t he num ber of not es

( t im e- frequency point s) of all the songs. Th e re a l p ow e r of
t h is se a rch is t h a t in st e a d of look in g if a on e n ot e e x ist s
in a son g , w e ’re look in g if 2 n ot e s se p a ra t e d from
d e lt a _ t im e se con d s e x ist in t h e son g . At t he end of t his part
we’ll t alk m ore about t im e com plexity.

Result filt ering

Though it is not m ent ioned in t he Shazam paper, I t hink t he

next t hing t o do is to filt er the M result s of the search by
keeping only the couples of t he songs that have a m inim um
num ber of t arget zones in com m on wit h t he record.

For exam ple, let’s suppose our search has ret urned:

100 couples from song 1 which has 0 target zone in

com m on wit h the record

10 couples from song 2 which has 0 t arget zone in com m on

wit h t he record

50 couples from song 5 which has 0 t arget zone in com m on

wit h t he record

70 couples from song 8 which has 0 t arget zone in com m on

wit h t he record

83 couples from song 10 which has 30 t arget zones in

com m on wit h the record

210 couples from song 17 which has 100 target zones in

com m on wit h the record

4400 couples from song 13 which has 280 t arget zones in

40 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

com m on wit h the record

3500 couples from song 25 which has 400 t arget zones in

com m on wit h the record

Our 10- second record has ( approxim at ely) 300 t arget s zone. I n
the best case scenario:

song 1 and t he record will have a 0% m at ching rat io

song 2 and t he record will have a 0% m at ching rat io

song 5 and t he record will have a 0% m at ching rat io

song 8 and t he record will have a 0% m at ching rat io

song 10 and the record will have a 10% m at ching rat io

song 17 and the record will have a 33% m at ching rat io

song 13 and the record will have a 91. 7% m at ching rat io

song 25 and the record will have a 100% m at ching rat io

We’ll only keep the couples of song 13 and 25 from t he result .

Alt hough songs 1 2,5 and 8 have m ult iples couples in com m on
wit h the record, none of them form at least a t arget zone ( of 5
points) in com m on wit h the record. This st ep can filt er a lot of
false result s because t he fingerprint dat abase of Shazam has a
lot of couples for t he sam e address and you can easily end up
wit h couples at t he sam e address t hat don’t belong t o the sam e
target zone. I f you don’t underst and why, look t he last pict ure
of the previous part : the [ 10; 30; 2] address is used by 2
tim e- frequency points t hat doesn’t belong to t he sam e t arget
zone. I f the record also have an [ 10; 30; 2] , ( at least ) one of t he
2 couples in t he result will be filt ered in t his step.

This st ep can be done in O( M) with the help of a hash table

whose key is t he couple ( songI D; absolut e tim e of the anchor in
the song) and value t he num ber of t im e it appears in the result:

We it erat e t hrough t he M results and count ( in t he hash

t able ) t he num ber of t im e a couple is present

We rem ove all the couples ( i.e. the key of t he hash table)
t hat appear less than 4 t im es ( in ot her words we rem ove
all t he points t hat doesn’t form a t arget zone) *

We count the num ber X of tim es t he song I D is part of a

41 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

key in the hash t able ( i.e we count t he num ber of com plet e
t arget zones in t he song. Since t he couple com e from t he
search, those t arget zones are also in t he record)

We only keep t he result whose song num ber is above

300* coeff ( 300 is the num ber of target zone of t he record
and we reduce t his num ber wit h a coeff because of t he
noise) .

We put the rem aining result s in a new hash t able whose

index is the songI d ( this hashm ap will be useful for t he
next step)

* The idea is t o look for t he t arget zone creat ed by an anchor

point in a song. This anchor point can be defined by t he id of
the song it belongs and t he absolut e t im e it occurs. I m ade an
approxim at ion because in a song, you can have m ult iple anchor
points at the sam e tim e. Since we’re dealing wit h a filt ered
spectrogram , you won’t have a lot of anchor point s at t he sam e
tim e. But t he key [ songID; absolut e t im e of t he anchor in t he
song] will gather all t he target zones creat ed by these t arget
points.

Not e: I used 2 hash tables in t his algorit hm . If you don’t know

how it works, j ust see it as a very efficient way t o st ore and get
dat a. I f you wan’t t o know m ore, you can read m y article on the
HashMap in Java which is j ust an efficient hash t able.

Tim e coherency

At t his st age we only have songs t hat are really close t o t he

record. But we st ill need t o verify t he t im e coherency bet ween
the not es of t he record and t hese songs. Let ’s see why:

42 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

In t his figure, we have 2 target zones t hat belong t o 2 different

songs. I f we didn’t look for tim e coherency, those target zones
would increase t he m at ching score bet ween t he 2 songs
whereas they don’t sound alike since t he not es in t hose target
zones are not played in t he sam e order.

This last st ep is about t im e ordering. The idea is:

t o com put e for each rem aining song t he not es and t heir
absolut e tim e posit ion in t he song.

t o do the sam e for the record, which gives us t he notes

and their absolut e tim e posit ion in t he record.

if t he not es in t he song and t hose in t he record are t im e

coherent, we should find a relation like t his one: “ absolut e
t im e of t he note in the song = absolut e tim e of the not e in
record + delt a” , were delta is t he st art ing t im e of t he part
of t he song t hat m at ches wit h t he record.

for each song, we need t o find t he delt a t hat m axim izes

t he num ber of not es t hat respect this tim e relat ion

Then we choose t he song t hat has t he m axim um num ber of

t im e coherent not es wit h t he record

Now t hat you get t he idea let’s see how to do it t echnically. At

this st age we have for t he record a list of address/ value:

[ “ frequency of t he anchor” ; ” frequency of the point” ; ” delt a

tim e bet ween t he anchor and t he point ” ] - > [ “ absolute t im e of

43 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

the anchor in t he record” ] .

And we have for each song a list address/ value ( stored in t he

hash table of the previous st ep) :

[ “ frequency of t he anchor” ; ” frequency of the point” ; ” delt a

tim e bet ween t he anchor and t he point ” ] - > [ “ absolute t im e of
the anchor in t he song” ; ” Id of t he song” ] .

The following process needs t o be done for all t he rem aining

songs:

For each address in the record, we get t he associated

value of the song and we com put e delt a = “ absolut e t im e
of t he anchor in t he record” – “ absolut e t im e of t he anchor
in t he song” and put t he delt a in a “ list of delta” .

I t is possible t hat t he address in the record is associat ed

wit h m ult iples values in t he song ( i. e. m ult iple point s in
different t arget zones of the song) , in t his case we
com pute t he delt a for each associated values and we put
t he deltas in t he “ list of delta”

For each different value of delt a in t he “ list of delt a” we

count it s num ber of occurrence ( in ot her words, we count
for each delt a t he num ber of not es t hat respect the rule
“ absolute t im e of t he not e in t he song = absolut e t im e of
t he note in record + delt a” )

We keep t he great est value ( which gives us the m axim um

num ber of not es t hat are t im e coherent bet ween t he
record and the song)

From all the songs, we keep t he song with t he m axim um tim e

coherent notes. If this coherency is above “ t he num ber of note
in t he record” * “ a coefficient ” then this song is the right one.

We j ust have t o look for t he m et adata of t he song ( “ artist

nam e” ,” song nam e, “ It unes URL” , ” Am azon URL” , …) wit h the
Song ID and gives t he result back to t he user.

Let ’s talk about com plexity!

44 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

This search is really m ore com plicated t hat the sim ple one we
first saw, let’s see if this is worth it . The enhanced search is a
st ep by st ep approach t hat reduces t he com plexit y at each
st ep.

For t he sake of com prehension, I ’ll recall all the assum pt ions
( or choices) I m ade and m ake new ones t o sim plify the
problem :

We have 512 possible frequencies

on average a song contains 30 peak frequencies per

second

Therefore t he 10- sec record cont ains 300 tim e- frequency

point s

S is t he num ber of seconds of m usic off all the songs

The size of the t arget zone is 5 not es

( new) I assum e t hat t he delta t im e between a point and it s

anchor is et her 0 or 10 m sec

( new) I assum e t he generat ion of addresses is uniform ly

dist ributed, which m eans t here is t he sam e am ount of
couple for any address [ X,Y,T] where X and Y are one of
t he 512 frequencies and T is either 0 or 10 m sec

The first step, t he search, only requires 5 * 300 unitary

searches.

The size of t he result M is t he sum of t he result of t he 5 * 300

unit ary searches

M = ( 5 * 300) * ( S * 30* 5 * 300) / ( 512 * 512 * 2)

The second step, t he result filt ering can be done in M

operat ions. At the end of this st ep t here are N not es dist ributed
in Z songs. Without a statist ical analysis of the m usic collect ion,
it’s im possible t o get t he value of N and Z. I feel N is really
lower t han M and Z represent only a few songs, even for a
40- m illion song dat abase like Shazam .

The last step is t he analysis of the t im e coherency of the

Z songs. We’ll assum e t hat each song as approxim at ely t he
sam e am ount of notes: N/ Z. I n the worst case scenario ( a
record that com es from a song t hat cont ains only one not e

45 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

played continuously) , the com plexity of one analysis is ( 5* 300)

* ( N/ Z) .

The cost of the Z songs is 5 * 300 * N.

Since N< < M, the real cost of t his search is M = ( 300 * 300 *
30* S) * ( 5 * 5) / ( 512 * 512 * 2)

If you rem em ber, the cost of t he sim ple search was: 300 * 300
* 30* S.

Th is n e w se a rch is 2 0 0 0 0 t im e s fa st e r

Not e: The real com plexity depends on distribut ion of

frequencies inside the songs of t he collection but t his sim ple
calculus gives us a good idea of the real one.

Im provem ent s

The Shazam paper is from 2003 which m eans t he associat ed

research is even older. I n 2003, 64- bit s processors were
released t o the m ainst ream m arket . Inst ead of using one
anchor point per t arget zone like t he paper proposes ( because
of the lim it ed size of a 32- bit int eger) , you could use 3 anchor
points ( for exam ple t he 3 point s j ust before t he t arget zone)
and st ore t h e a d d re ss of a point in t he target zone in a
6 4 - b it in t e g e r. Th is w ou ld d ra m a t ica lly im p rov e t h e
se a rch t im e . I ndeed, t he search would be t o find 4 not es in a
song separated from detla_t im e1, detla_t im e2 and det la_tim e3
seconds which m eans the num ber of results M would be very
( very) lower than the one we j ust com puted.

A great advant age of t his fingerprint search is its h ig h

sca la b ilit y :

I n st e a d of h a v in g 1 fin g e rp rin t d a t a b a se y ou ca n
h a v e D d a t a b a se s, each of them cont aining 1/ D of t he
full song collect ion

You can search at t he sam e t im e for t he closest song of the

record in t he D dat abases

Then you choose t he closest song from t he D songs

Th e w h ole p roce ss is D t im e s fa st e r.

46 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

Tradeoffs

Anot her good discussion is the noise robust ness of t his

algorithm . I could easily add 2k words j ust for this subj ect but
aft er 11k words I t hink it’s bet ter not to speak about it … or
j ust a few words.

If you read carefully, you not iced t hat I used a lot of thresholds,
coefficient s and fixed values ( like t he sam pling rat e, the
duration of a record, …) . I also chose/ m ade m any algorithm s
( t o filter a spectrogram , t o generat e a spect rogram , …) . They all
have an im pact on t he noise resist ance and t he t im e
com plexit y. The real challenge is t o find the right values and
algorithm s t hat m axim ize:

The noises resist ance

The tim e com plexit y

The precision ( reducing the num ber of false posit ive

results)

I hope you now understand how Shazam works. It took m e a lot

of tim e to underst and t he different subj ect s of t his art icle and I
st ill don’t m ast er t hem . This art icle won’t m ake you an expert
but I hope you have a very good pict ure of t he processes
behind Shazam . Keep in m ind t hat Shazam is j ust one possible
audio fingerprint ing im plem ent at ion.

You should be able to code your own Shazam . You can look at
this very good art icle t hat focuses m ore on how to code a
sim plified Shazam in Java t han t he concept s behind it . The
sam e author m ade a present at ion at a Java conference and t he
slides are available here. You can also check t his link for a
Mat Lab/ Oct ave im plem ent at ion of Shazam . And of course, you
can read by yourself the paper from Shazam co- founder
Avery Li- Chun Wang by clicking right here.

47 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

The world of m usic com puting is a very int erest ing field with
touchy algorit hm s t hat you use every day without knowing it.
Though Shazam is not easy to underst and it ’s easier than:

query by hum m ing: for exam ple SoundHound , a concurrent

of Shazam , allows you to hum / sing t he song you’re looking
for

speech recognit ion and speech synt hesis: im plem ent ed by

Skype, Apple “ Siri” and Android “ Ok Google”

m usic sim ilarity: which is t he ability to find t hat 2 song are

sim ilar. I t’s used by Echonest a st art - up recent ly acquired
by Spot ify

If you’re interested, t here is an annual cont est between

researchers on those t opics and the algorit hm s of each
part icipant are available. Here is the link t o t he MI REX cont est .

I spent approxim ately 200 hours during t he last 3 years t o

understand the signal processing concept s, t he m at hem at ics
behind them , to m ake m y own Shazam prot ot ype, to fully
understand Wang’s paper and im agine t he processes t he paper
doesn’t explain. I wrot e t his article because I have never found
an art icle t hat really explains Shazam and I wished I could have
found one when I began t his side proj ect in 2012. I hope I didn’t
writ e too m any t echnical m ist akes. The only t hing I ’m sure of is
that despite m y effort s t here are m any gram m ar and spelling
m ist akes ( alas! ) . Tell m e what you t hink of this art icle on t he
com m ents.

1,755 369

48 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

Algorit hm algorithm m usic

Related Posts

How does a relational dat abase work

Mem ory opt im isat ion: Cust om Set

« « JVM m em ory m odel Leonard Susskind’s Quant um

Mechanics course » »

Leave a Reply

9 1 Com m ent s on " H ow does Shazam w or k"

Zeehtbert November 25, 2016 10:02 am

Thank you so m uch. Didn’t read t he whole t hing

Guest
yet , but definit ely will continue lat er. .
enlight ening! ! !

Reply - Share

stino November 4, 2016 6:09 pm

Great art icle! Thanks for t he explanation. I was

Guest
always wondering how shazam found t he result s
so fast! Alt hough som e years ago I had DSP course
and practical class where we had t o m ake MATLAB
exam ples t o im plem ent t he fft a dft. Am azingly
how fast people forget t hings t hey have learned in
life if they don’t use it in daily life. Thanks for the
refreshm ent , really enj oyed it !

Reply - Share

49 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

Anonymous November 3, 2016 8:43 pm

Am azing work. So m uch inform at ion to digest. All

Guest of m y fifth sem ester’s com puter networking notes
have been applied int o Shazam

Reply - Share

Jomars October 18, 2016 7:58 pm

Bravo! ! !am azing! !!

Guest

Reply - Share

ĝródmieĞcie October 16, 2016 12:30 pm

Good j ob! I’m t hinking to im plem ent sim ilar

Guest
algorit hm for m y m aster t hesis, and I believe that
Your article will be a great st art ! Thanks!

Reply - Share

Anonymous September 12, 2016 3:16 pm

i dindt com e here to know how shazam does t heir

Guest
t hing i j ust want t o know how it works m an lm fao

Reply - Share

Anonymous August 26, 2016 8:57 pm

t hat ’s one of t he best articles i read in m y life. .

Guest
t hank you so m uch

Reply - Share

Load Mor e Com m ent s

50 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

Top 10 articles

How does Shazam work

195, 619 views | 91
com m ent s

How does a relat ional

dat abase work
166, 687 views | 127
com m ent s

How does a HashMap work

in JAVA
58, 420 views | 52
com m ent s

Design Pat t ern: fact ory

pat terns
11, 155 views | 8
com m ent s

JVM m em ory m odel

6,950 views | 8 com m ent s

Machine Learning: Andrew

NG’s course from coursera
5,321 views | 0 com m ent s

Design pat t ern: singleton,

prot ot ype and builder
5,059 views | 4 com m ent s

How t o conduct t echnical

interviews?
4,718 views | 13
com m ent s

t he best program m ing

language
4,038 views | 6 com m ent s

What is a good
applicat ion?

51 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

2,824 views | 2 com m ent s

Categories

Algorit hm ( 3)

Design Pat t ern ( 3)

Java ( 2)

JVM ( 2)

Methodology ( 3)

review ( 2)

Tool ( 1)

Tags

big dat a

builder cousera dat abase

eclipse fact ory garbage

collect or hashm ap hashset

HashTable j vm
Liskov LSP m achine learning

m em or y m ooc m usic

prot ot ype

quant um m echanics reference

runt ine dat a areas set short cut

key singlet on sof t ref erence

W eakHashMap weak reference

work

Sear ch t his sit e...

Sear ch

52 of 53 11/30/2016 12:26 PM
How does Shazam work - Coding Geek http://coding-geek.com/how-shazam-works/

Proudly pow ered by WordPress Prem ium St yle Them e by w w w.gopiplus.com

53 of 53 11/30/2016 12:26 PM