libmad - MPEG audio decoder library


Except where otherwise noted, all code was authored by: Robert Leslie <rob at underbit com>
Copyright (C) 2000-2008 Underbit Technologies, Inc.
See the file CREDITS for a list of contributors.


MAD (libmad) is a high-quality MPEG audio decoder. It currently supports MPEG-1 and the MPEG-2 extension to Lower Sampling Frequencies, as well as the so-called MPEG 2.5 format. All three audio layers (Layer I, Layer II, and Layer III a.k.a. MP3) are fully implemented.

MAD does not yet support MPEG-2 multichannel audio (although it should be backward compatible with such streams) nor does it currently support AAC.

MAD has the following special features:

Because MAD provides full 24-bit PCM output, applications using MAD are able to produce high quality audio. Even when the output device supports only 16-bit PCM, applications can use the extra resolution to increase the audible dynamic range through the use of dithering or noise shaping.

Because MAD uses integer computation rather than floating point, it is well suited for architectures without a floating point unit. All calculations are performed with a 32-bit fixed-point integer representation.

Because MAD is a new implementation of the ISO/IEC standards, it is unencumbered by the errors of other implementations. MAD is NOT a derivation of the ISO reference source or any other code. Considerable effort has been expended to ensure a correct implementation, even in cases where the standards are ambiguous or misleading.

Because MAD is distributed under the terms of the GPL, its redistribution is not generally restricted, so long as the terms of the GPL are followed. This means MAD can be incorporated into other software as long as that software is also distributed under the GPL. (Should this be undesirable, alternate arrangements may be possible by contacting Underbit.)

About the code

The code is optimized and performs very well, although specific improvements can still be made. The output from the decoder library consists of 32-bit signed linear fixed-point values that can be easily scaled for any size PCM output, up to 24 bits per sample.

The API for libmad can be found in the mad.h header file. Note that this file is automatically generated, from the following files:

so it will not exist until after you have built the library.

There are two APIs available, one High-Level API, and the other Low-Level API. With the low-level API, each step of the decoding process must be handled explicitly, offering the greatest amount of control. With the high-level API, after callbacks are configured, a single routine will decode an entire bitstream.


The file minimad.c contains an example usage of the libmad API that shows only the bare minimum required to implement a useful decoder. It expects a regular file to be redirected to standard input, and it sends decoded 16-bit signed little-endian PCM samples standard error and decoding continues. Note that the scale() to standard output. If a decoding error occurs, it is reported to routine in this code is only provided as an example; it rounds MAD's high-resolution samples down to 16 bits, but does not perform any dithering or noise shaping. It is therefore not recommended to use this routine as-is in your own code if sound quality is important.

Integer Performance

To get the best possible performance, it is recommended that an assembly version of the fixed-point multiply and related routines be selected. Several such assembly routines have been written for various CPUs.

If an assembly version is not available, a fast approximation version will be used. This will result in reduced accuracy of the decoder.

Alternatively, if 64-bit integers are supported as a datatype by the compiler, another version can be used that is much more accurate. However, using an assembly version is generally much faster and just as accurate.

More information can be gathered from the ` fixed.h ' header file.

MAD's CPU-intensive subband synthesis routine can be further optimized at the expense of a slight loss in output accuracy due to a modified method for fixed-point multiplication with a small windowing constant. While this is helpful for performance and the output accuracy loss is generally undetectable, it is disabled by default and must be explicitly enabled.

Under some architectures, other special optimizations may also be available.

Audio Quality

The output from MAD has been found to satisfy the ISO/IEC 11172-4 computational accuracy requirements for compliance. In most configurations, MAD is a Full Layer III ISO/IEC 11172-3 audio decoder as defined by the standard.

When the approximation version of the fixed-point multiply is used, MAD is a limited accuracy ISO/IEC 11172-3 audio decoder as defined by the standard.

MAD can alternatively be configured to produce output with less or more accuracy than the default, as a tradeoff with performance.

MAD produces output samples with a precision greater than 24 bits. Because most output formats use fewer bits, typically 16, it is recommended that a dithering algorithm be used (rather than rounding or truncating) to obtain the highest quality audio. However, dithering may unfavorably affect an analytic examination of the output (such as compliance testing); you may therefore wish to use rounding in this case instead.

Portability Issues

GCC is preferred to compile the code, but other compilers may also work. The assembly code in ` fixed.h ' depends on the inline assembly features of your compiler. If you're not using GCC or MSVC++, you can either write your own assembly macros or use the default (low quality output) version.

The union initialization of ` huffman.c ' may not be portable to all platforms when GCC is not used.

The code should not be sensitive to word sizes or byte ordering, however it does assume A % B has the same sign as A.

Building and Installing

Windows Platforms

MAD can be built under Windows using either MSVC++ or Cygwin. A MSVC++ project file can be found under the ` msvc++ ' subdirectory.

To build libmad using Cygwin, you will first need to install the Cygwin tools:

You may then proceed with the following POSIX instructions within the Cygwin shell.

Note that by default Cygwin will build a library that depends on the Cygwin DLL. You can use MinGW to build a library that does not depend on the Cygwin DLL. To do so, give the option --host=mingw32 to ` configure '.

POSIX Platforms (including Cygwin)

The code is distributed with a ` configure ' script that will generate for you a ` Makefile ' and a ` config.h ' for your platform. See the file ` INSTALL ' for generic instructions.

The specific options you may want to give ` configure ' are:

optimize for speed over accuracy
optimize for accuracy over speed
do not compile with debugging support, and use more optimizations
do not build a shared library

Note that you need not specify one of --enable-speed or --enable-accuracy; in its default configuration, MAD is optimized for both. You should only use one of these options if you wish to compromise speed or accuracy for the other.

By default the package will build a shared library if possible for your platform. If you want only a static library, use --disable-shared.

It is not normally necessary to use the following options, but you may fine-tune the configuration with them if desired:

use the ARCH-specific version of the fixed-point math assembly routines (current options are: intel, arm, mips, sparc, ppc; also allowed are: 64bit, approx)
use the subband synthesis optimization, with reduced accuracy
do not use certain architecture-specific optimizations

By default an appropriate fixed-point assembly routine will be selected for the configured host type, if it can be determined. Thus if you are cross-compiling for another architecture, you should be sure either to give ` configure ' a host type argument (--host) or to use an explicit --enable-fpm option.

If an appropriate assembly routine cannot be determined, the default approximation version will be used. In this case, use of an alternate --enable-fpm is highly recommended.

Experimenting and Developing

Further options for ` configure ' that may be useful to developers and experimenters are:
enable diagnostic debugging support and debugging symbols
generate gprof profiling code
enable code using the EXPERIMENTAL preprocessor define


Please read the ` COPYRIGHT ' file for copyright and warranty information. Also, the file ` COPYING ' contains the full text of the GNU GPL.

Send inquiries, comments, bug reports, suggestions, patches, etc. to:

Underbit Technologies, Inc. <support at underbit com>

See also the MAD home page on the Web:

Generated on Tue Jun 10 12:14:17 2008 for libmad by  doxygen 1.5.5