|MPEG Audio Layer III|
|File extension :||
|MIME type :||audio / mpeg audio / MPA audio / mpa-robust|
|Magic number :||
\ xFF \ xFB
( ASCII-C notation )
|Standard (s) :||ISO / IEC 11172-3, ISO / IEC 13818-3|
MP3 , original spelling mp3 (name after the file name extension ; actually MPEG-1 Audio Layer III or MPEG-2 Audio Layer III) is a method for lossy compression of digitally stored audio data. MP3 uses psychoacoustics with the aim of storing only parts of the signal that can be perceived by humans. This enables a strong reduction in the amount of data if the audio quality is not (or hardly) perceived as being reduced.
With an example data rate of 192 kbit / s, which already enables high quality, the compression rate of an MP3 audio file is around 85% compared to an uncompressed audio CD . MP3 is the dominant method of storing and transferring music on computers, smartphones , on the Internet and on portable music players ( MP3 players ), although there are now a number of technologically advanced options. The process was mainly developed in Germany under the direction of Karlheinz Brandenburg and Hans-Georg Musmann .
In May 2017, the developers stopped licensing the format after the last patents in the US had expired (in Europe, MP3 had been patent-free since 2012). It is therefore now a freely available standard.
The MP3 format was developed from 1982 under the direction of Hans-Georg Musmann by a group led by Karlheinz Brandenburg at the Fraunhofer Institute for Integrated Circuits (IIS) in Erlangen and at the Friedrich-Alexander University Erlangen-Nuremberg in collaboration with AT&T Bell Labs and Thomson . From 1989 the development within ISO / IEC JTC1 SC29 WG11 (MPEG) was continued. In 1992 it was codified as part of the MPEG-1 standard. The history of the standardization and the appreciation of the contributions of the researchers is shown in Genesis of the MP3 Audio Coding Standard by Hans Georg Musmann in IEEE Transactions on Consumer Electronics, Vol. 52, No. 3, pp. 1043-1049, August 2006 . The Italian research center CSELT (Head of Media: Leonardo Chiariglione ) was the body that allowed standardization. The file name extension .mp3 (as an abbreviation for ISO MPEG Audio Layer 3 ) was determined on July 14, 1995 after an internal survey by the institute; previously the filename extension .bit was used internally . Brandenburg has received several awards for the development of this data format.
As early as the mid-1990s, players and software for PCs were in circulation that made it possible to save and play compressed MP3 files. The exchange of such files via the Internet has also been simplified: even with a simple ISDN speed, the transfer only required two to three times the playback time; With DSL lines, the transmission was even far below the playing time. This soon led to a lively exchange of audio files ( file sharing ) without observing the copyright of the respective artist or composer. Attempts by the music industry to take action against this have been characterized by only moderate success to this day, especially since the exchange systems are also developing further and based on the peer-to-peer principle without central, controllable authorities. At the end of the 1990s there were already large collections of music files on the Internet, for example on MP3.com or Napster , which caused the number of users to increase significantly. The first portable MP3 players appeared in stores in 1998 .
Patents and license disputes
The processes for MPEG coding (“MP3”) are now patent-free and can therefore be used freely. The original, almost finished standard MPEG-1 (Parts 1, 2 and 3) was published on December 6, 1991 as ISO CD 11172. In most countries, patents can no longer be registered if the "state of the art" has already been published. Patents lose their validity 20 years after the initial application; in some countries this period can be extended by up to 12 months, depending on the registration date. As a result, the patents required to implement MP3 technology lost their validity in most countries in December 2012, 21 years after the publication of the ISO CD 11172 standard.
An exception was the United States, where patents filed before June 8, 1995 were no longer valid after 17 years. However, it was possible to significantly delay the date of granting a patent by extending the filing period. The various MP3-related patents lost their validity in the United States between 2007 and 2017. The MP3 technology was patent-free in the USA on April 16, 2017 at the latest, when the US patent no.6009399 held by the Fraunhofer-Gesellschaft (and administered via Technicolor ) expired.
As a consequence, the Fraunhofer-Gesellschaft discontinued its license program on April 23, 2017. The US patents administered and claimed by Sisvel, a large MP3 patent pool, had also expired by April 2017 (the last three patents still valid after 2015 were: US Patent No. 5878080, expired in February 2017, US Patent No. 5850456, expired February 2017, and U.S. Patent No. 5960037, expired April 9, 2017).
In May 2017, the Linux distribution Fedora announced that it would officially include MP3 decoders and encoders in the distribution, as the corresponding patents had expired.
The Fraunhofer-Gesellschaft and other companies owned by 2017 software patents on partial methods, which are used for MPEG encoding. There was no all-inclusive MP3 patent. The Fraunhofer-Gesellschaft contributed the largest part to the development of the MP3 standard and had some methods for MP3 coding patented. In a merger with Thomson , the two companies owned 18 MP3-related patents. From September 1998, after the MP3 standard was able to establish itself for six years, until April 2017, FhG / Thomson demanded license fees for the production of hardware and software that used the MP3 format.
Bell Laboratories' patents were originally thought to have been used to develop the format . At the time, these rights were held by Alcatel-Lucent , which Bell Labs had taken over. The company had filed patent lawsuits against Microsoft, Dell and Gateway around the turn of the millennium. In the proceedings against Microsoft in February 2007, Lucent was awarded 1.52 billion US dollars in the first instance. However, that judgment was overturned by the San Diego Federal District Court in August 2007. Sisvel also made patent infringement claims on behalf of Philips .
Like most lossy compression formats for music, the MP3 process uses psychoacoustic effects of human perception of tones and noises. For example, people can only distinguish between two tones when there is a certain minimum difference in pitch , before and after very loud noises they can perceive quieter noises less or not at all for a short time. So you do not need to store the original signal exactly; the signal components that the human ear can also perceive are sufficient. The task of the coder is to process the original sound signal according to fixed rules based on psychoacoustics so that it requires less storage space, but still sounds exactly like the original to the human ear. If the original and the MP3 version are completely identical, one speaks of transparency . In principle, however, due to the lossy compression, the original signal cannot be exactly reconstructed from the MP3 signal. There are also lossless methods for audio data compression such as FLAC , but these achieve significantly lower compression rates and are even less common - especially in the area of playback hardware.
When playing the MP3 signal generated in this way, the decoder uses the reduced data to generate an analog audio signal that sounds original for the majority of listeners, but which is not identical to the original signal, as information was removed during the conversion into MP3 format. If one were to compare the temporal waveform of the MP3 audio signal with the original, for example on the screen of an oscilloscope , clear differences would be seen. Because of the psychoacoustics of human perception mentioned above, the MP3 signal still sounds exactly like the original to a listener - provided that a mature encoder and a sufficiently high data rate (bit rate) is used during encoding.
While the decoding always follows a fixed algorithm, the coding can be done according to different algorithms (e.g. Fraunhofer encoder, LAME encoder) and accordingly provides different acoustic results. The question of whether some or many listeners perceive a loss of quality depends, among other things, on the quality of the encoder, the complexity of the signal, the data rate, the audio technology used ( amplifier , loudspeaker ) and ultimately also the hearing of the listener from. In addition to fixed data rates of 8 kbit / s up to 320 kbit / s, the MP3 format also allows any free data rates up to 640 kbit / s (Freeform MP3) in the freeformat mode. However, only a few MP3 player decoders are designed for higher bit rates than those from the ISO standard (currently up to 320 kbit / s).
The quality impressions are quite subjective and differ from person to person and from ear to ear. Most people can no longer distinguish the encoded material from the source material from a higher bit rate and when using a sophisticated encoder, even with concentrated listening. Nevertheless, in a listening test by the c't magazine, certain pieces of music, even at 256 kBit / s, could be distinguished from CD quality. However, the test was carried out in 2000 - since then the MP3 encoders have improved significantly. In people with "abnormal" hearing (e.g. with hearing damage due to impact trauma ), the mechanisms used sometimes do not work as intended, so that they tend to notice differences between the coded and source material (e.g. because loud tones affect the damaged hearing hears poorly, can no longer cover other tones well). The test person, who was best able to identify differences in the aforementioned test, even at high data rates, has damaged hearing.
In addition to coding with a constant data rate (= fluctuating quality, along with the complexity of the audio signal that changes over time), coding with constant quality (and thus fluctuating data rate) is also possible. This (to a large extent) avoids quality drops at difficult to encode music passages, but on the other hand saves on the data rate and thus on the final file size for calm or even completely silent passages of the audio stream. The quality level is specified, and in this way you get the minimum file size required for it.
- A first step in data compression is based, for example, on channel coupling of the stereo signal by forming the difference, since the data of the right and left channels are highly correlated , i.e. they are very similar. This is a lossless process, the output signals can be fully reproduced ( mid / side stereo ).
- According to the human auditory curve , signal components are represented in less precisely perceptible frequency ranges with less precision, in that the Fourier-transformed data material is quantized accordingly .
- So-called masking effects are used to store signal components that are less important for the auditory impression with reduced precision. These can be weak frequency components in the vicinity of strong overtones. A strong tone at 4 kHz can also mask frequencies up to 11 kHz. The greatest savings with MP3 encoding are therefore that the tones are only stored with just enough precision (with so many bits) that the resulting quantization noise is masked and therefore inaudible.
- The data, which are available in so-called frames , are finally Huffman-coded .
A flaw in the design is that the process is applied in blocks, which means that gaps can appear at the end of a file. This is annoying with audio books or live recordings, for example, in which a coherent lecture is broken down into individual tracks. Here, the last blocks stand out as annoying pauses (perceptible as cracks or a short drop-out ). This can be remedied by using the LAME encoder, which adds exact length information, in combination with a playback program that can handle this, such as foobar2000 or Winamp . However, some playback programs such as Windows Media Player do not support this method, known as gapless playback . Apple iTunes supports it from version 7.
Compression in detail
The compression consists of the following steps:
- Subband transformation of the signal
- MDCT transformation of the signal , then (!) The signal is divided into blocks.
- For stereo signals: Matrixing: decision for each block whether the signal is coded as a left-right or a middle-side signal
- Quantize the signal
- Huffman coding with fixed codebooks
Steps 4 and 5 take care of data reduction, with quantization being the lossy process.
Note: In the following text, the specified spectral widths and times refer to an audio signal with a sampling frequency of 48 kHz.
Subband transformation of the signal
With the subband transformation, the signal is broken down into 32 frequency bands of equal width using a polyphase filter bank (as with MPEG Layer 1, MPEG Layer 2 and dts). The filter bank works on a FIFO buffer with a size of 512 samples , to which 32 new samples are always fed in one step. This means that 16 filter windows always overlap on the audio signal.
The decision to use frequency bands of equal width simplifies the filters, but does not reflect human hearing, the sensitivity of which is non-linearly dependent on frequency.
Since there are no ideal filters in practice , the frequency ranges overlap, so that a single frequency can also occur in two adjacent sub-bands after filtering.
Subband filtering is encumbered by US Pat. No. 6,199,039.
MDCT transformation of the signal
The signals of the subbands are now transferred into the frequency domain by the modified discrete cosine transformation (MDCT). As a result, the frequency bands are further resolved spectrally. The MDCT can transform the bands into either short blocks (12 samples results in 6 frequency bands) or long blocks (36 samples, 18 frequency bands). Alternatively, the two lowest frequency bands can also be transformed with long blocks and the rest with short blocks. Long blocks have a better frequency resolution and are more suitable if the audio signal does not suddenly change within the corresponding frame (stationarity).
At the output of the MDCT, the signal is divided into blocks. From 576 input values (if you take into account the window width of the filter, there are actually a total of 1663 input values), through two transformations connected in series, either
- 576 spectral coefficients (long blocks),
- 3 × 192 spectral coefficients (short blocks) or
- 36 + 3 × 180 spectral coefficients (hybrid block, hardly used)
For 2-channel stereo signals, you can now decide whether the signal should be encoded either as mono (single-channel), stereo, joint stereo or dual-channel. In contrast to AAC or Ogg Vorbis, this decision has to be made globally for all frequencies.
The stereo method (not joint stereo) (like dual-channel) is lossy due to the fact that even at 320 kbit / s only 160 kbit / s are available per channel, but depending on the complexity, one of the two can be selected Channels assigned different bit rates. Dual-Channel stores two independent mono tracks (e.g. bilingual text tracks) with the same bit rate coding; however, not necessarily every decoder reproduces both tracks at the same time.
There are two coding methods for joint stereo : intensity and mid / side stereo , which are also used in combination; Both methods form a middle channel (L + R) from the sum of both channels and the side channel (L − R) from the volume difference between the two channels. In contrast to the mid / side stereo method, the phase ( time difference ) of the signal is neglected in intensity stereo . The joint stereo method eliminates the frequent redundancy in the stereo channels in order to be able to encode the signals with a higher bit rate than with the stereo method; but if the channel signals are very dissimilar, the joint stereo method reverts to normal stereo coding.
Since the audio signal is first differentiated into frequency bands, the stereo information, if it can be used by the ear at all, must also be coded in a differentiated manner. Here, z. B. at depths or frequencies from 2 kHz, information content can be saved, in that the relevant non-localizable signals are no longer channel-true, but subsumed with adjacent frequency bands encoded (intensity stereo), or placed in the stereo center.
Due to the continuous further development of the codecs , the joint stereo method has recently been seen as the best solution for standard music, very similar stereo channels due to the better compression rate, higher bit rate coding and the lossless (except for low-frequency) stereo image.
The quantization is the essential step that is experiencing losses during encoding. It is mainly responsible for shrinking the amount of data.
Adjacent frequency bands are combined into groups of 4 to 18 bins . These are given a common scale factor s = 2 N / 4 , with which they are quantized. The scale factor determines the accuracy of the coding of this frequency band. Smaller scale factors result in a more precise coding, larger scale factors less precise (or no values unequal 0 at all).
From x 0 , x 1 ,…, x 17 the values N and Q 0 , Q 1 ,…, Q 17 become with the relation x i ~ Q i 4/3 2 N / 4 .
The non-linear coding Q 4/3 (for negative values: - (- Q) 4/3 ) was introduced for the first time in MP3 coding. MPEG layers 1 and 2 use linear coding.
This step is essentially responsible for the quality as well as the data rate of the resulting MP3 data stream. He is supported by a psychoacoustic model that tries to simulate the processes in the average human hearing and controls the control of the scale factors.
The scale factors N and the quantized amplitudes Q of the individual frequencies are Huffman-coded using fixed code tables .
The final MP3 file consists of a series of frames that begin with a start marker (sync) and contain one or two blocks created in the manner described above.
During decompression, the compression steps are carried out in reverse order. After Huffman decoding, the data are prepared for the inverse modified cosine transformation (IMCT) by means of inverse quantization. This forwards its data to an inverse filter bank, which now calculates the original samples (lossy due to the quantization in the coding process).
At that time, the industry was already working on the MDCT- based AAC, which is more cleanly designed and delivers better results with a comparable effort.
In addition (in the direction of high-quality coding) there are also further developments to achieve acceptable sound quality at very low data rates (less than 96 kbit / s). Representatives of this category are mp3PRO and MPEG-4 AAC HE and AAC +. With this method, however, transparency can only be achieved through High Definition (HD) AAC (AAC LC + SLS).
The MP3 surround format of the Fraunhofer Institute for Integrated Circuits IIS offers an extension to include multi-channel capabilities. MP3 surround allows 5.1 sound to be played back at bit rates that are comparable to those of stereo sound and is also fully backwards compatible. Thus, conventional MP3 decoder can decode the signal in stereo, MP3 surround decoder but full 5.1 - Surround generate sound stage.
To do this, the multi-channel material is mixed into a stereo signal and encoded by a regular MP3 encoder . At the same time, the surround sound information from the original is inserted as surround extension data into the "ancillary data" data field of the MP3 bit stream. The MP3 data can then be played back as a stereo signal by any MP3 decoder. The MP3 surround decoder uses the inserted extension data and reproduces the full multi-channel audio signal.
Further developments concern procedures for copyright protection , which could possibly be implemented in future versions.
Audio raw material requires a lot of storage space (1 minute stereo in CD quality about 10 MB) and high data transfer rates or a lot of time for transfer (for example via the Internet) . Lossless compression does not reduce the amount of data to be transferred as much as lossy methods, which for most cases (exceptions are, for example, studio applications or archiving) still deliver acceptable quality. The MP3 format for audio data quickly achieved the status that JPEG compression has for image data.
MP3 became known to the general public primarily through music exchanges . In the warez scene, many DVD rips use the MP3 audio format as the sound track. With CD ripper programs it is possible to extract the music from audio CDs and output them to MP3 files. There are also many programs that make it possible to convert MP3 to another format, but also vice versa (example: the audio track of a YouTube video ( FLV ) is converted to an MP3 file). Another focus of application were MP3 players , with which you can listen to music on the go. Most smartphones nowadays also support MP3 files.
In the WWW there are numerous applications for MP3 technology, from self-composed music on (even) audiobooks spoken, radio plays, bird calls and other sounds through to find podcasting . Musicians can now distribute their music worldwide without a sales organization and make sound recordings available on a website without much effort (apart from the GEMA fees, including their own compositions that are registered with GEMA) . Users can use search engines to find all imaginable (non-commercial) sounds and styles of music.
Even with multimedia software, especially with PC games, the often numerous audio files are stored in MP3 format. In addition, MP3 is used by numerous - mostly smaller - online music stores .
In contrast to more modern codecs, MP3 files originally offered no way of storing metadata (for example title, artist, album, year, genre) about the piece of music they contained.
Regardless of the developer of the format, a solution was found that is supported by almost all software and hardware players: The ID3 tags are simply attached to the beginning or the end of the MP3 file. In the first version (ID3v1) they are appended at the end and are limited to 30 characters per entry and a few standard entries. The much more flexible version 2 (ID3v2) is not supported by all MP3 players (especially hardware players), since the tags are inserted at the beginning of the MP3 file. There are also considerable differences within ID3v2. The most widespread are ID3v2.3 and ID3v2.4, although only ID3v2.4 officially allows the use of UTF-8 encoded characters (previously only ISO-8859-1 and UTF-16 were allowed). However, many hardware players only display UTF-8 tags as jumbled characters . Since ID3v2 tags are at the beginning of the file, this data can also be read during transmission via HTTP, for example, without first reading the entire file or requesting several parts of the file. In order to avoid having to rewrite the entire file when changes are made, padding is usually used , which means that space is reserved for these changes in advance.
The metadata from the ID3 tag can be used, for example, to display information about the track currently being played, to sort the tracks into playlists or to organize archives.
|Byte 1||Byte 2||Byte 3||Byte 4|
|Sync||11 bit||all bits are set to 1|
|ID||2 bits||0 = MPEG version 2.5
1 = reserved
2 = MPEG version 2
3 = MPEG version 1
|Layer||2 bits||0 = reserved
1 = layer III
2 = layer II
3 = layer I
|Protection||1 bit||0 = 16-bit CRC after the header
1 = no CRC
|Bit rate||4 bit||according to the bit rate table|
|Sampling frequency||2 bits||according to the sampling table|
|Padding||1 bit||0 = frame is not filled
1 = frame is filled with extra
slot Slot size: Layer I = 32 bits; Layer II + III 8 bits
|Private||1 bit||only informative|
|Channel mode||2 bits||0 = stereo
1 = joint stereo
2 = 2 mono channels
3 = one channel (mono)
|Fashion extension||2 bits||(only for joint stereo)
according to the mode extension table
|copyright||1 bit||0 = without copyright
1 = with copyright
|original||1 bit||0 = copy
1 = original
|Emphasis||2 bits||0 = none
1 = 50/15 ms
2 = reserved
3 = ITU-T J.17
Bit rate table (information in kbps)
|value||MPEG 1||MPEG 2 / 2.5|
|Layer I||Layer II||Layer III||Layer I||Layer II / III|
Sampling frequency table (specifications in Hz)
|value||MPEG 1||MPEG 2||MPEG 2.5|
Mode extension table
|value||Layer I / II||Layer III|
|0||Sub-bands 4 to 31||Intensity stereo: off; M / S stereo: off|
|1||Subbands 8 to 31||Intensity stereo: one; M / S stereo: off|
|2||Subbands 12 to 31||Intensity stereo: off; M / S stereo: a|
|3||Subbands 16 to 31||Intensity stereo: one; M / S stereo: a|
The frame data (possibly first CRC), which contains the encoded audio data, follow the frame header. A frame has a playing time of 1152 samples at a sample rate of 32,000 to 48,000 samples per second; at lower sample rates (16,000 to 24,000 samples per second) it is only 576. At 48,000 samples per second this corresponds to 24 ms. The amount of data in a frame can be calculated according to the properties specified in the header. The size of a frame in bytes can then be calculated using the following formula, whereby the division is to be carried out as an integer division:
- Frame size = (144 · bit rate): sample rate + padding [bytes]
If the amount of data cannot be stored in a frame for complex pieces of music, MP3 offers a so-called bit reservoir. This memory area is intended as additional space for the file and extends the data in the corresponding frame. For this purpose, the encoder encodes previous music passages with a lower data rate and thus does not completely fill out earlier frames, the bit reservoir is created. This created free storage space can now be used for the larger amount of data of more complex music passages. The maximum size of this data reservoir is 511 bytes, whereby only previous frames may be filled.
The encoder of the Fraunhofer-Gesellschaft and the encoder of the open source project LAME are available for coding MP3 files . There is also the ISO dist10 reference encoder and other projects such as Xing , blade and Gogo .
In addition to MP3, there are numerous other audio formats . The Vorbis format is open source and has been designated by the developers as patent-free. (Vorbis was published 15 years before the MP3 patents expired.) Vorbis has proven to be superior to MP3 in technical analyzes and blind tests, especially in the low and medium bit rate ranges . The qualitative advantage of Vorbis is only slightly noticeable in the high bit rate range (around 256 kbit / s). Ogg-Vorbis also offers multi-channel support, and Ogg can also accept video and text data as a container format . The latter is only supported by very few MP3 players and radios.
The free Musepack (formerly MPEGPlus) based on MP2 algorithms was developed to provide even better quality than the MP3 format at bit rates above 160 kbit / s. However, it was not able to gain wide acceptance because it is more aimed at use by enthusiasts in the high-end sector and is hardly supported in the commercial sector. Files in Musepack format can be recognized by the extension mpc or mp + .
Advanced Audio Coding (AAC) is a standardized method within the framework of MPEG-2 and MPEG-4 that was developed by several large companies. Apple and RealMedia use this format for their online music stores, and Nero AG provides an encoder for the format. A free encoder is also available with faac . At low bit rates of up to around 160 kbit / s, AAC is superior to MP3 in terms of sound quality - the lower the bit rate, the clearer it is - it allows multi-channel sound and is widely supported by the industry (for example in cell phones and MP3 players ).
Windows Media Audio (WMA) is an audio format developed by Microsoft and is widely used for DRM -protected downloads. While it can be played on many common platforms, it has not stood up to the MP3 format.
The team around Brandenburg made the first practical tests with the a cappella version of the song Tom's Diner by Suzanne Vega . In his search for suitable test material, Brandenburg read in a hi-fi magazine that their testers were using the song to assess loudspeakers and found the piece to be a suitable challenge for audio data compression.
- Franz Miller: The mp3 story: A German success story , Carl Hanser Verlag GmbH & Co. KG, ISBN 978-3-446-44471-3 .
- Roland Enders: The home recording manual. The way to optimal recordings. 3rd, revised edition, revised by Andreas Schulz. Carstensen, Munich 2003, ISBN 3-910098-25-8 .
- Thomas Görne: Sound engineering. Fachbuchverlag Leipzig in Carl Hanser Verlag, Munich et al. 2006, ISBN 3-446-40198-9 .
- Hubert Henle: The recording studio manual. Practical introduction to professional recording technology. 5th, completely revised edition. Carstensen, Munich 2001, ISBN 3-910098-19-3 .
- Michael Dickreiter, Volker Dittel, Wolfgang Hoeg, Martin Wöhr (eds.): Manual of the recording studio technology. Walter de Gruyter, Berlin / Boston 2014, ISBN 978-3-11-028978-7 or e- ISBN 978-3-11-031650-6 .
- MP3 put to the test
- Audio interview with Dr. Karlheinz Brandenburg on the genesis and background of MP3 from March 16, 2004 (48 minutes)
- Short interview with Hans-Georg Musmann about the development of MP3 as part of the experiment of the week
- The history of MP3 - memories of Ernst F. Schröder, one of the MP3 developers
- Podcast "Green Luck": Suzanne Vega on her role as the "mother of mp3"
- The MP3 revolution began in Germany
- Karl-Gerhard Haas: 25 years of MP3: It's the Klang-Quetsche's birthday. In: Heise online . July 14, 2020 .
- Imre Grimm: Music is in the air: How eight German engineers wrote history with mp3 25 years ago. Sunday - the weekend magazine of the Leipziger Volkszeitung , online portal, July 25, 2020. Accessed on August 2, 2020 .
- RFC 3003 , RFC 3555 , RFC 5219
- MP3 turns 10. (No longer available online.) July 12, 2005, archived from the original on February 12, 2016 ; Retrieved February 3, 2011 . Info: The archive link was inserted automatically and has not yet been checked. Please check the original and archive link according to the instructions and then remove this notice.
- MP3 is officially dead: Fraunhofer has stopped licensing , curved.de from May 15, 2017; Accessed May 16, 2017.
- Fraunhofer IIS: License program for MP3 ends , Heise online from May 15, 2017; Accessed May 16, 2017.
- Miller, Franz. The mp3 story: A German success story. Carl Hanser Verlag GmbH Co KG, 2015.
- Performance of a Software MPEG Video Decoder, Ketan Patel, Brian C. Smith, and Lawrence A. Rowe, ACM Multimedia 1993 Conference, Berkeley.edu (in English)
- THE MPEG FAQ ¦Version 3.1 - May 14, 1994¦ PHADE SOFTWARE Leibnizstr. 30, 10625 Berlin, GERMANY - Owner Frank Gadegast
- A Big List of MP3 Patents (and supposed expiration dates) . In: tunequest . February 26, 2007 (in English).
- Patent US5703999 : Process for reducing data in the transmission and / or storage of digital signals from several interdependent channels. Registered on November 18, 1996 , published on December 30, 1997 , applicant: Fraunhofer-Gesellschaft, inventor: Jürgen Herre, Dieter Seitzer, Karl-Heinz Brandenburg, Ernst Eberlein.
- mp3. Retrieved February 15, 2018 .
- US MPEG Audio patents. (PDF) October 27, 2016, accessed on October 27, 2016 (English).
- Patent US5850456 : 7-channel transmission, compatible with 5-channel transmission and 2-channel transmission. Filed February 8, 1996 , published December 15, 1998 , Applicant: US Philips Corporation, Inventor: Warner RT Ten Kate, Leon M. Van De Kerkhof.
- Full MP3 support coming soon to Fedora - Fedora Magazine . In: Fedora Magazine . May 5, 2017 ( fedoramagazine.org [accessed February 15, 2018]).
- Microsoft sees hundreds of companies affected by MP3 patent dispute , heise.de
- Microsoft achieves success in the dispute over MP3 patents , heise.de
- Carsten Meyer: Kreuzverhörtest , heise.de. In: c't 6/2000, section award ceremony
- AudioHQ over MP3 quality
- What is Gapless Playback (English), Apple.com
- David Salomon: Data Compression . The Complete Reference. 4th edition. Springer, 2007, ISBN 978-1-84628-602-5 , pp. 815 (English).
- OGG vs. LAME (English), Pricenfees.com
- MPC vs VORBIS vs MP3 vs AAC at 180 kbps, 2nd checkup with classical music (English), hydrogenaudio.org
- Freeware Advanced Audio Coder , SourceForge.net
- Many hi-fi fans can apparently hear the grass growing - interview with the “MP3 inventor” Karlheinz Brandenburg on Tagesanzeiger.ch , accessed on January 20, 2015.