About MP3 Audio Coding

The MP3 audio format has been around for over thirty years, and while it has been largely superseded by other codecs, it continues to be used in many situations.

With metadata carried in ID3 tags and often embedded inside MP3 files, MP3 audio is widely supported on almost every computing platform, mobile device, TV receiver and tablet. It is ideal for distribution to consumers even though there are higher quality alternatives such as AAC, HE-AAC and MPEG-H.

MP3 is a mathematically lossy but perceptually lossless codec. It is never an optimal solution for creating, editing and archiving purposes. It is optimized for distributing content to end-users.

It is highly recommended that MP3 is NOT used in production workflows. Arguably, contributions delivered in MP3 format should be rejected and requested again in a lossless format instead. MP3 may be adequate for spoken content but not for music delivery.

MP3 is NOT an abbreviation for MPEG-3.

Why Did We Need The MP3 Codec?

The MP3 compression technology was devised because early Internet connections were unable to deliver audio files in a timely manner. A three-minute raw audio track extracted from a CD creates a 32MB file:

• Duration - 3 minutes (180 seconds).
• Sample rate - 44100 samples per second.
• Sample size - 16-bits.
• Number of tracks - 2 (Stereo).
• Resulting file size - 31.752 MB.

MP3 achieves a compression ratio of between 10:1 and 14:1 whilst still preserving (audibly) almost CD quality. The encoded output file would be approximately 3MB, depending on the content.

Audio delivered on DVD disks for movies and TV programs is sampled at a higher rate of 48kHz. This will create slightly larger files.

Audio compression was originally developed for the MPEG-1 standard.  It was then inherited into MPEG-2 and enhanced.  

MPEG-1 and MPEG-2 happily coexist and are not mutually exclusive. MPEG-2 does not replace MPEG-1 but inherits capabilities from it and augments them.

The bitstream syntax definition must be carefully reconciled across the MPEG-1 and MPEG-2 standards to remove any ambiguity.

How It Works

It’s not necessary to fully understand the inner workings of MP3 to utilize it. A brief overview here will suffice.

The MPEG Audio coding is organized into three layers which offer different levels of complexity and compression ratios.

Much of this work was done by Fraunhofer IIS collaborating with other international experts. Fraunhofer subsequently developed a (non-canonical) variant which they called MPEG 2.5 Layer III. This supports additional bit-rates but is not widely adopted.

The next generation AAC encoder refines these techniques to improve the quality and compression ratios.

MP3 players are expected to decode and play all MPEG content but may present an error message when encountering MPEG 2.5 Audio.

The original research was based on Psychoacoustic Analysis. This is the study of sound perception to compile statistics based on listener feedback about quality. A mathematical model is constructed from those statistical results.

MP3 Perceptual Encoding is driven by a statistical model that removes components the human ear would not perceive when they are masked by other louder components. This immediately makes MP3 a lossy codec. It reduces the complexity of the content to yield a better compression ratio and that discarded information cannot be restored by uncompressing the file.

Coding for layers I and II splits the audio into discrete sub-bands which it then analyzes individually. Rather like a spectrum analyzer displays a snapshot of the frequency response:

Layer III splits the audio into many more sub-bands and passes overlapping groups of them through a Modified Discrete Cosine Transform (MDCT) to apply a more fine-grained process. This trades complexity and computing load for better compression ratios but introduces more latency and artefacts that need to be removed by post-processing.

The DCT transform yields a list of coefficients which gradually decay down to zero. A quantization/entropy cut-off truncates the coefficients at the point where they are all at zero (and expected to remain so) for lossless compression.

Truncating earlier, when they are nearly zero, causes lossy compression.

The energy level of each band is measured and compared to a masking threshold value. This is fed back into the Quantizer which moves the entropy coder cut-off point earlier in the series of coefficients. Louder passages are coded less accurately and can withstand a higher level of quantization.

The entropy cut-off point determines the amount of quantization noise introduced as an audible artefact. It is constrained within the limits determined by the psychoacoustic statistical analysis to hide the artefacts.

MP3 Joint Stereo processing looks for similarities between the audio channels. Differences depend on how the sound sources are panned across the soundscape during mix-down. Coding a single channel plus the differences needed to recreate a second channel improves the compression ratio.

Audio Layers

Layers I and II were defined during the MPEG-1 standardization. Layer III was added for MPEG-2. The layer names use Roman numerals which may be confusing.

The term MP3 is an abbreviation of MPEG Audio layer III and describes MPEG-1 and MPEG-2 content.

The term MP2 describes MPEG Audio Layer II and is sometimes incorrectly used to describe other kinds of MPEG-2 Audio.

• Layer I - This is simpler than Layer II. The frame sizes are smaller which reduces coding delay (latency). It is useful for tele-conferencing applications and was designed for real-time encoding on early hardware systems. Layer I is now deemed to be obsolete.
• Layer II - Layer II performs well with orchestral content and delivers results nearly as good as AAC. Players decode this with less computational effort than Layer III. It is more complex than Layer I but yields a better compression ratio.
• Layer III - This is designed to operate at a lower bitrate than Layer II. It works quite differently with a much larger number of sub-bands which are processed in overlapping groups with a Modified Discrete Cosine Transform (MDCT) algorithm, as detailed above. Layer III does not handle transients quite as well as Layer II and needs additional pre-echo detection to increase the available bit rate during difficult passages. Additional post processing techniques are necessary to reduce artefacts which increases the computational workload.

For some content, Layer II performs better than Layer III even though it is less efficient.

These are the preferred file type extensions appropriate for the three layers:

Supported Bit Rates

MPEG-2 adds bit rates in the lower range that MPEG-1 does not already support. These are useful for implementing compression for spoken-word content rather than music. Higher bit-rates are undefined in MPEG-2 and are inherited from MPEG-1:

Supported Sample Rates

The lower sample rates added by MPEG-2 support more efficient speech encoding. The higher sample rates are undefined in MPEG-2 and are inherited from MPEG-1:

Channel Encoding Modes

The MPEG-1 standard only supports two channels. These can be configured for mono or stereo applications.

• Mono - Only one single channel is required.
• Stereo – two channels.
• Joint stereo - intensity encoded.
• Joint stereo - Mid/side encoded (Layer III only).
• Dual mono - Two uncorrelated mono channels.

MPEG-2 adds four more channels (six in all) to potentially carry 5.1 surround-sound content. Backwards compatible two channel stereo uses a sub-set of two out of the six available channels in the same way as MPEG-1:

• Front-Left - Used when two channel stereo is encoded.
• Front-Right - Used when two channel stereo is encoded.
• Front-Center - For dialog.
• Low frequency.
• Surround (rear) - Left.
• Surround (rear) - Right.

Surround systems with more channels than the 5.1 configuration cannot be delivered with MPEG-2.

File Name Extensions

Media Type Identifiers

When serving files with HTTP or uploading them from a browser, the media type header in the transaction will indicate the essence format.

Relevant Standards

These are the relevant standards that you should acquire to support your use of MP3:

Applying MP3

MPEG-2 does not replace MPEG-1 but augments it with additional features.

The MP3 audio file format is widely used and supported on most platforms and devices. Content creation is easy and metadata editing is a feature of many compatible applications.

The patents relevant to MP3 all expired in 2017. This facilitates the use of open-source libraries such as LAME without risk of litigation.

These Appendix articles contain additional information you may find useful: