MP3 Format

Theory of the Standard

ELECTGON
www.electgon.com
ma_ext@gmx.net

03.06.2019
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 General Speci ications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1 Operating mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Sampling frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Bit‐rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Encoding Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3 Analysis Filter Bank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
4 MDCT Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4.0.1 Antialiasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4.0.2 Window Switching . . . . . . . . . . . . . . . . . . . . . . . . . 6
5 Scalefactors Bands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
6 Reordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
7 Bit Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
8 Huffman Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
9 Frame Packing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
MP3 Theory 1. Introduction

1 Introduction
Smart systems are exploited in many applications recently starting from home applications
up to industry and huge manufacturing sectors. These smart systems rely to high extend on
information exchange that is carried out using processing systems for analog and digital sig‐
nal. Audio streams are part of this exchanged information. Many applications are introduced
to facilitate dealing with this type of information. MP3 format is serving as best and most
common format for these audio application. In this document important introduction about
MP3 format is demonstrated to understand how this format works with high efϐiciency that
made it most common format for three decades so far.

2 General Speci ications

MPEG‐1 was the irst international standard for the perceptual coding. This standard was
developed by MPEG group in November 1992. It de ines coding scheme for the digital audio
using three possible techniques (or layers). These three layers are working similarly but the
higher layer has more sophisticated con iguration than others. MPEG‐1 layer III (which is the
scope of this project) has more complexity than other layers. It is known also as MP3 and it
has successfully found many applications in Video, CD, ISDN, Video games, broadcasting, etc.
In order to make this standard applicable in all these applications, MPEG‐1 Layer III de ined
a data representation including a number of options

2.1 Operating mode

MP3 Standard is designed to work with mono and dual channel signals. For Dual channel
signals, another extension is also added to cover ef icient combined coding of the left and
right channels of a stereophonic audio signal. Namely there is stereo joint stereo coding so
MP3 allows both mid/side stereo coding and intensity stereo coding. Thus The operating
modes are:

• single channel

• dual channel (two independent channels, for example containing different language
versions of the audio)

• stereo (no joint stereo coding)

• joint stereo.

2.2 Sampling frequency

MPEG audio compression works on a number of different sampling frequencies. In general

MPEG‐1 de ines audio compression at sampling frequencies 32 kHz, 44.1 kHz and 48 kHz.
This means that MP3 encoder is able to handle audio signals that are sampled at these frequencies.
MP3 standard is de ined to receive 1152 audio samples. If these samples were sampled at 32

3
MP3 Theory 3. Analysis Filter Bank

KHz, this means that the 1152 audio samples covers 36 ms of the audio source. 44.1 KHz
sampling covers each 26 ms. 48 KHz covers 24 ms.

2.3 Bit‐rate

Bit rate expresses capacity or number of bits used to represent audio samples. The implementer
of the MP3 encoder can choose which bitrate he can use to compress the samples in condition
that the selected bitrate within a range of bit‐rates from 8 kbit/s up to 320 kbit/s as this the
de ined range for the standard. MP3 standard gives the option to the implementer to switch
bitrate from audio frame to another frame.

2.4 Encoding Process

In very general terms, MP3 is encoding audio signal that are sampled into intervals, each
intervals lasts for 24 ms (in case of 48 KHz sampling frequency is used) and results in 1152
audio samples. These 1152 samples pass through ilter bank to be analyzed further in frequency
domain. This ilter bank will result in frequency coef icients that are quantized according
to psychoacoustic model. After that it undergoes to further coding using Huffman coding
algorithm. These coded frequency coef icients are packed into frames prior to transmission.
This process can be depicted in more details in igure 1.

Figure 1: MP3 Encoder Main Architecture[5]

3 Analysis Filter Bank

When we call this stage by “Analysis”, this means that this stage is trying to understand input
audio signal and analyze its nature for improving coding performance. For Example the
input signal is divided into 32 subbands, in parallel another analysis is done based on FFT to
calculate energy stored in each subband to know masker components and masking threshold.

4
MP3 Theory 4. MDCT Functions

This analysis result in what is called Signal to Mask Ratio (SMR). This analysis together with
frequency coef icients are provided to next stage for quantization and encoding. At decoder
side, the opposite process is done and it is called then “Synthesis” Filters.
Filter Bank that is used in MP3 encoding is Hybrid Filter that is composed of polyphase
ilter and MDCT ilter. Both ilters are producing frequency components at the output of
its transformation process. The frequency components are called subband samples if the
ilterbank has low frequency resolution, otherwise they are called frequency lines or frequency
coef icients. The polyphase has low frequency resolution so it is said that we get subbands
as a result of the polyphase iltering. Polyphase ilter is acting as band pass ilter. For MP3
it has 32 sub ilters. This means that its output is 32 frequency bands. MDCT ilter has more
resolution so it is able to analyze these 32 subbands into determined frequency coef icients.
Typically it produces 18 or 6 frequency coef icients for each subband. Figure 2 shows building
blocks for this Hybrid ilter.

Figure 2: Filter Bank Basic Blocks

4 MDCT Functions
4.0.1 Antialiasing

When reconstructing original signal at decoder side again, some aliasing effects can appear
because of missed frequency coef icient that were iltered out during encoding process. MDCT
is used to overcome this point by applying overlapping between current iltered frame and
the following frame. To make it clear, the Polyphase ilter is iltering frame by frame (each
frame is 1152 audio samples) but for the MDCT it processes two consecutive frames to make
overlapping between them. MDCT is performing 50% overlapping so eventually MDCT will
release frame by frame is can be demonstrated from igure 3 where M can be 1152 in our
case.

5
MP3 Theory 4. MDCT Functions

Figure 3: Overlapping in MDCT[6]

4.0.2 Window Switching

Another important role of the MDCT ilter comes from its window function. In MP3 standard,
MDCT is designed to apply four different window functions to the input spectrum as shown
in igure 4.

Figure 4: Window Functions of MDCT[7]

long, long‐to‐short, short‐to‐long windows are 36 points window functions, this means it
generate 36 frequency coef icients if these are applied. Short window is a 12 point window
function, this means it generate 12 frequency coef icients if it is applied. These generated 36
or 12 frequency coef icients are interpreted later into 18 or 6 frequency coef icients because
of overlapping property of the MDCT.
As explanation for the windowing operation, this can be described as declared in [5] “The
reason behind using window switching technique in MP3 can be explained as follows. A
crucial issue in frequency domain coding is the potential for preechoes. Consider the case
in which a silent period is followed by a percussive sound within the same coding block.
Such an onset (attack) will cause comparably large instantaneous quantization errors. This
preechoes can become distinctively audible. Preechoes can be masked by the time domain
effect of of premasking if the time spread is of short length (in the order of few milliseconds).
Therefore they can be reduced or avoided by using blocks of short length. However, a large
percentage of the total bit rate is typically required for the transmission of side information if
the blocks are shorter. A solution to this problem is to switch between block sizes of different
lengths. This means to use short blocks only to control preecho artifacts during nonstationary
periods of the signal, otherwise the coder uses long blocks. Figure 5 demonstrates the effect
of preecho when it is dealt with short and long blocks.”

6
MP3 Theory 4. MDCT Functions

Figure 5: Preecho Handling with Short and Long Blocks.[5]

(a) Source signal, (b) reconstructed signal with block size N=1024, and (c) reconstructed
signal with block size, N=256.

MP3 format is applying long window in MDCT for sample that has little change in its sound
level. Short window for samples that have considerable change in its sound level. Transition
from long window to short window is not happening immediately but as can be shown in
igure 6.

Figure 6: Sequence of Window Transition[5]

This windowing operation will result in three types of blocks at the output of MDCT. Long
Block which has 18 frequency coef icients. Short Block which has 6 frequency coef icients.
Mixed Block which has 18 frequency coef icients. It is important to point out here that MP3 is
using always three consecutive short blocks so that to keep order of number of coded sample

7
MP3 Theory 5. Scalefactors Bands

for each frame. So eventually one long block will have 32x18=576 frequency coef icients (and
also for Mixed Block). One short block is having 32x6=192 frequency coef icients, however
there is always three consecutive short blocks and they are grouped together to make 576
frequency coef icients.

5 Scalefactors Bands
At this stage we can summarize previous processes as: Filter Bank is iltering input spectrum
of each frame into 32 uniformly distributed subbands, each subband has 18 frequency coef icients
which leads to total of 576 frequency coef icients. Actually these 32 subbands are not emulating
critical bands of human ear. So Perceptual encoding shouldn’t be based on these bands as
it will not be able encode audible maskers within each critical band. Therefore frequency
coef icients are grouped into scalefactors bands that emulates critical bands of human ear.
These scalefactors bands depend on which sampling frequency is used for the audio. Also
these scalefactors bands depend on type of the block (long, mixed, short). For long blocks,
the 576 frequency coef icients are grouped into 22 scalefactors bands. For short block it is
13 scalefactors bands. For mixed block it is 18 scalefactors bands.
Tables 1 and 2 shows distribution of these frequencies on scalefactors bands.

32 KHz 44.1 KHz 48 KHz

Scalefactor Band Width Freq. Width Freq. Width Freq.
0 4 0‐3 4 0‐3 4 0‐3
1 4 4‐7 4 4‐7 4 4‐7
2 4 8‐11 4 8‐11 4 8‐11
3 4 12‐15 4 12‐15 4 12‐15
4 4 16‐19 4 16‐19 4 16‐19
5 4 20‐23 4 20‐23 4 20‐23
6 6 24‐29 6 24‐29 6 24‐29
7 6 30‐35 6 30‐35 6 30‐35
8 8 36‐43 8 36‐43 6 36‐41
9 10 44‐53 8 44‐51 8 42‐49
10 12 54‐65 10 52‐61 10 50‐59
11 16 66‐81 12 62‐73 12 60‐71
12 20 82‐101 16 74‐89 16 72‐87
13 24 102‐125 20 90‐109 18 88‐105
14 30 126‐155 24 110‐133 22 106‐127
15 38 156‐193 28 134‐161 28 128‐155
16 46 194‐239 34 162‐195 34 156‐189
17 56 240‐295 42 196‐237 40 190‐229
18 68 296‐363 50 238‐287 46 230‐275
19 84 364‐447 54 288‐341 54 276‐329
20 102 448‐549 76 342‐417 54 330‐383
21 26 550‐575 158 418‐575 192 384‐575

Table 1: Scalefactors Bands De ined in MP3 ‐ Long Blocks Type

8
MP3 Theory 5. Scalefactors Bands

32 KHz 44.1 KHz 48 KHz

Scalefactor Band Width Freq. Width Freq. Width Freq.
0 4 0‐3 4 0‐3 4 0‐3
1 4 4‐7 4 4‐7 4 4‐7
2 4 8‐11 4 8‐11 4 8‐11
3 4 12‐15 4 12‐15 4 12‐15
4 6 16‐21 6 16‐21 6 16‐21
5 8 22‐29 8 22‐29 6 22‐27
6 12 30‐41 10 30‐39 10 28‐37
7 16 42‐57 12 40‐51 12 38‐49
8 20 58‐77 14 52‐65 14 50‐63
9 26 78‐103 18 66‐83 16 64‐79
10 34 104‐137 22 84‐105 20 80‐99
11 42 138‐179 30 106‐135 26 100‐125
12 12 180‐191 57 135‐191 66 126‐191

Table 2: Scalefactors Bands De ined in MP3 ‐ Short Blocks Type

Grouping 576 frequency coef icients from subbands into scalefactors bands means nothing
is changing the frequency coef icients itself. It means only that the new grouping will undergo
to special treatment afterward. This treatment is applying scalefactor values. It means that
values of frequency coef icients will be adjusted so that it can be represented correctly with
suitable number of bits. In other words, frequency coef icients undergo to scaling by these
scalefactor values. MP3 standard is using limited and prede ined scalefactor values; within
the range 0 to 4. Table 3 shows scalefactor values that are assigned to each scalefactor band
in case of long block type and sampling frequency is 44.1 KHz.

9
MP3 Theory 5. Scalefactors Bands

SF SF Freq.
SF SF Freq.
Value Band
Value Band
0‐3
0 0‐3
0 4‐7
1 4‐7
8‐11
2 8‐11
12‐15
3 12‐15
1 16‐19
4 16‐19
SF scfsi SF Freq. 20‐23
5 20‐23
Value Band Band 24‐27
6 24‐29
2 28‐31
0 0‐3 7 30‐35
32‐35
1 4‐7 slen1 slen1 36‐39
36‐39
2 8‐11 8 40‐43
0 3 40‐43
3 12‐15 44‐47
44‐47
4 16‐19 48‐53
48‐53
slen1 5 20‐23 9 54‐59
4 54‐59
6 24‐29 60‐65
60‐65
7 30‐35 66‐73
66‐73
1 8 36‐43 10 74‐81
5 74‐81
9 44‐51 82‐89
82‐89
10 52‐61 90‐99
90‐99
11 62‐73 11 100‐109
6 100‐109
12 74‐89 110‐119
110‐119
2 13 90‐190 120‐131
120‐131
14 110‐133 12 132‐143
7 132‐143
15 134‐161 144‐155
slen2 144‐155
16 162‐195 156‐169
156‐169
17 196‐237 13 170‐183
8 170‐183
3 18 238‐287 184‐197
184‐197 slen2
19 288‐341 slen2 198‐215
198‐215
20 342‐417 14 216‐233
9 216‐233
21 418‐575 234‐251
234‐251
252‐273
252‐273
15 274‐295
10 274‐295
296‐317
296‐317
318‐347
318‐347
16 348‐377
11 348‐377
378‐407
378‐407
17 408‐575
12 408‐575

(a) (b) (c)

Table 3: Scalefactors Distribution at 44.1 KHz

(a) Long Block Type, (b) Short Block Type (c) Mixed Block Type

slen1 and slen2 are parameters that have values between 0 and 4. Selection of speci ic
values for slen1 and slen2 is done during quantization process that will be explained shortly.
Note that these tables are at 44.1 KHz.Other operating frequencies (32 and 48 KHz) have
similar pattern but distribution of frequency coef icients is according to its de ined table
in ISO‐11172‐3. Note also that distribution of short block type in table 3 is shown after
consolidation of three successive short windows. this is applied also for the mixed block

10
MP3 Theory 6. Reordering

table.
In a nutshell there are critical bands of human ear, subbands which are output of Polyphase
ilter, scalefactors bands which are used by Quantization and Encoding process to encode
suitable frequency coef icients, scfsi bands which are used for further compression.

6 Reordering
Since short block types are consolidated together when it is used in MDCT to compensate for
frequency number of each frame, the order of this consolidation has to be known. It is not
mentioned explicitly in ISO‐11172‐3 what the order of these three blocks is. However it is
mentioned in section 2.4.3.4.8 in MP3 standard in Decoding section as follows “If short blocks
are used (block_type==2), the rescaled data xr[scf_band][window][freq_lines] (as described
in huffmancodebits() in 2.4.1.7) shall be reordered in subband order, xr[subband][window][freq_line],
prior to the IMDCT operation.”
Then in Huffman declaration section it is mentioned that “The Huffman encoded data
are given for successive scalefactor bands, beginning with scalefactor band 0 and ending
with scalefactor band 11. Within each scalefactor band, the data is given for successive time
windows, beginning with window 0 and ending with window 2. The data values within each
window are arranged in order of increasing frequency.”
What can be understood then is; each scalefactor band will expand to consolidate the
same scalefactor band of the three windows. The clause “The data values within each window
are arranged in order of increasing frequency.” may make confusion understanding what is
meant by increasing frequency; Taking for example scalefactor band 9. This scalefactor band
has width of 18 for each window, so after consolidation it will be 54. Does it mean then to
arrange ascending frequencies of each band by irst 18 of irst window then irst 18 of second
window then irst 18 of the third window. Or does it mean to arrange the content of each
window so that irst window will have irst frequency of irst window then irst frequency of
second window then irst frequency of third window then second frequency of irst window,
second frequency of second window, second frequency of third window, etc. This confusion
may be removed if the clause is meant to be “The data values of each window are arranged
in order of increasing frequency.” In this case, it means that frequencies of irst, second and
third windows are arranged in increasing order together. So the second understanding will
be followed. Moreover, a lot of previous context and researches are based on this conception.
Figure 7 demonstrates this reorder operation at encoder side. Note that at decoder side, the
opposite operation is needed before IMDCT iltering.

11
MP3 Theory 7. Bit Allocation

Figure 7: Reorder Process at Encoder

7 Bit Allocation
Quantization process is doing mainly two tasks; irst task is to allocate (quantize) input frequency
coef icients. Second Task is to scale some frequency coef icients if it found that number of
bits assigned to it is too much so that after scaling these frequency coef icients, number of
assigned bits can be reduced. These two task are executed with a constraint to keep the
quantization noise below the masking threshold. The allocation control algorithm suggested
for the layer III encoder uses non uniform and dynamic quantization. Non uniform means
step size between quantization levels is not equal. Dynamic means that this non‐linearity
differs from one scalefactor band to another (i.e. same non‐linearity is not applied for each
scalefactor band). Quantization is applied to 576 spectral values at a time. This is done
iteratively in two nested loops, a distortion control loop (outer loop) and a rate control loop
(inner loop). The following explanation of these loops is declared by [8].

rate control loop: The rate loop does the quantization of the frequency domain samples
and thus also determines the required quantization step size. Furthermore the subdivision
of the big values (see Huffman part below) into regions, the Huffman table selection decision
for each region and the calculation of the boundaries between the regions take place here.
To begin with the samples are quantized with an increasing step size until the quantized
values can be coded using one of the available Huffman code tables. A larger step size leads to
smaller quantized values. Then the overall Huffman coded bit sum is calculated and compared
with the number of bits available. If the calculated bit sum exceeds the number of bits available
the quantization step size is further increased and the entire procedure is repeated until the
available bits are suf icient. The non‐linearity is achieved raising each sample to the power

12
MP3 Theory 8. Huffman Encoding

of 3/4.

distortion control loop: This loop controls the quantization noise which is produced by
the quantization of the frequency domain lines within the rate control loop. The aim is to keep
the quantization noise below the masking threshold (allowed noise given by the psychoacoustic
model) for each scalefactor band. To shape the quantization noise scalefactors are applied to
the frequency lines within each scalefactor band. The scalefactors of all scalefactor bands and
the quantization step size are then saved before the rate control loop in called. After the inner
loop the quantization noise is calculated. This is repeated until there is no more scalefactor
band with more noise than allowed by the threshold. The values of the scalefactors belonging
to bands that are too noisy are increased for each iteration loop. Finally the noise caused
by the quantization will not be audible by a human and the loop will exit. There are still
situations where both loops can go on forever depending on the calculated threshold. To
avoid this there are several conditions in the distortion control loop that can be checked to
stop the iterations more early.

8 Huffman Encoding
Huffman coding algorithm is considered a lossless encoding method which means it is able
to represent input samples in fewer bits without losing any information. What is important
to point out here is the input to Huffman encoder is 576 quantized frequency coef icients.
These 576 lines are classi ied into three regions; big values, count1, zero regions. Big values
region includes frequencies that needs higher gain as it is audible by human ear so it is better
to represent it clearly and it is further divided into region0, region1, region2. Count1 region
includes frequencies with values between ‐1 and 1. Zero region is not coded and approximated
to 0 as it in not audible by human ear. Figure 8 shows these different regions.

Figure 8: Huffman Regions[9]

13
MP3 Theory 9. Frame Packing

9 Frame Packing
Encoding process resulted in frequency coef icients represented with Huffman bits that were
scaled with scalefactors values. These are mainly what is needed to be sent to a receiver in
order to be able to decode original samples back. Some associated information has to be sent
also to clarify to the receiver what are number of bits representing the whole frequencies
(part2_3_length), limits of Huffman regions (big_values, count1, etc), sampling frequency, etc.
Figure 9 shows anatomy of the MP3 frame.

Figure 9: MP3 Frame Structure[10]

It is important to point out here that one MP3 frame carries 1152 frequency samples. i.e.
two 576 frequencies. First 576 frequencies are labeled as granule 0, the second is labeled as
granule 1. What can be concluded then from this frame anatomy that it starts with Header
which is needed by the receiver to know synchronization of bits sequence and to know additional
tags like which layer is this frame (layer I, II, III), bit rate and sampling frequency, etc. Then
Side Information is sent to inform the receiver about no of Main Data bits, big values region
length, count1 length, which table is used in Huffman algorithm, scalefactors bands length,
etc. All these details can be reviewed in ISO 11172‐3 standard section 2.4.1.7.

14
Bibliography

[1] S. Committe, Health Risk from Exposure to Noise from Personal Music Players, sep 2008.

[2] http://www.intropsych.com, June 2015.

[3] https://www.wikipedia.org/, June 2015.

[4] V. G. Oklobdzija, Ed., Digital Systems and Applications. CRC Press, 2008.

[5] V. Madisetti, Ed., Digital Signal Processing Handbook. CRC Press, 2009.

[6] A. S. Ted Painter, “Perceptual coding of digital audio,” Telecommunications Research

Center, Arizona State University, Tech. Rep.

[7] C.‐J. Tsai, Audio Codecs, National Chiao Tung University, dec 2012.

[8] R. Raissi, “The theory behind mp3,” Tech. Rep., December 2002.

[9] M. Schmidt, “Hardware modelle fü r die mp3‐huffman dekodierung,”, Hochschule

Bremen, feb 2014.

[10] K. C. Pohlmann, Principles of Digital Audio, 5th ed. McGraw‐Hill, 2005.

15