Normalize

This is release 0.7.7 of Normalize, an audio file volume normalizer.

Copyright © 1999--2005, Chris Vaill <chrisvaill at gmail>

Normalize is a tool for adjusting the volume of audio files to a standard level. This is useful for things like creating mixed CD's and mp3 collections, where different recording levels on different albums can cause the volume to vary greatly from song to song.

Send bug reports, suggestions, comments to chrisvaill at gmail.

normalize is free software. See the file COPYING for copying conditions.


Installation synopsis

  1. ./configure options

  2. make

  3. make install

See the file INSTALL for more extensive directions. See the man page, normalize.1, for usage. Run "./configure --help" for configure options.


Dependencies

These dependencies are optional. Normalize doesn't require any other packages to compile and run.

MAD library (http://www.underbit.com/products/mad/)

Normalize will use the MAD MPEG Audio Decoder library if you have it (highly recommended). This gives normalize the ability to read mp3 files. MAD support in normalize was developed using MAD version 0.14.2b; earlier versions may not work.

You can run configure with the --without-mad option to turn off mp3 read support.

Audiofile library (http://www.68k.org/~michael/audiofile/)

Normalize can use the audiofile library if version 0.2.2 or later is available on your system. This gives normalize the ability to read and write AIFF, AIFF-C, WAV, NeXT/Sun .snd/.au, Berkeley/IRCAM/CARL, and whatever else the audiofile library people decide to implement in the future.

Audiofile support is not turned on by default, because the built-in WAV support is faster (only because it's specifically tailored for PCM WAVs), and because I'm guessing most people only ever need to normalize standard PCM WAV and mp3 files. If you only want to use normalize on standard PCM WAV and mp3 files, you don't need audiofile. If, however, you would like to be able to normalize all the different audio file formats that audiofile handles, run configure with the --with-audiofile option to turn on audiofile support.

XMMS (http://www.xmms.org/)

If you have xmms installed, the configure system will build the xmms-rva plugin, which honors the relative volume adjustment frames that normalize adds to ID3 tags. The option --disable-xmms prevents the plugin from being built.


Questions and Answers

1. What platforms does normalize work on?

I've tested normalize on GNU/Linux and FreeBSD on x86, Solaris on Sparc, and Irix on MIPS. I've heard that it works on GNU/Linux on Alpha and on BeOS R5. As far as Windows is concerned, you can compile it using the Cygwin toolkit. Question 8, below, contains a brief overview of this process.

I've tried to make the code as portable as possible, so I'd appreciate hearing whether normalize works on other platforms.

2. What is normalize useful for?

Example 1. Let's say you've got a bunch of wav files containing what are, in your estimation, Elvis's greatest hits, collected from various albums. You want to encode them as mp3's and add them to an established collection, but since they're all from different albums, they're all recorded at different volumes from each other and from the rest of your mp3 collection. If you've been using normalize on all your wav files before you encode them, your collection is normalized to the default volume level, and you want these new additions to be at the same level. Just run normalize with no options on the files, and each will be adjusted to the proper volume level:

            normalize "Hound Dog.wav" "Blue Suede Shoes.wav" \
                      "Here Comes Santa Claus.wav" ...

Example 2. Suppose now you've just extracted all the wav files from the Gorilla Biscuits album "Start Today," which, you may know, is recorded at a particularly low volume. We want to make the whole album louder, but individual tracks should stay at the same volume relative to each other. For this we use batch mode. Say the files are named 01.wav to 14.wav, and are in the current directory. We invoke normalize in batch mode to preserve the relative volumes, but otherwise, everything's the default:

            normalize -b *.wav
You can then fire up your mp3 encoder, and the whole album will be uniformly louder.

Example 3. Now suppose we want to encode the Converge album "When Forever Comes Crashing." This album has one song, "Ten Cents," that is really quiet while the rest of the songs have about the same (loud) volume. We'll turn up the verbosity so we can see what's going on:

        > normalize -bv *.wav
        Computing levels...
        Level for track01.cdda.wav: -9.3980dBFS (0.0000dBFS peak)
        Level for track02.cdda.wav: -9.2464dBFS (-0.1538dBFS peak)
        Level for track03.cdda.wav: -8.6308dBFS (-0.2520dBFS peak)
        Level for track04.cdda.wav: -8.7390dBFS (0.0000dBFS peak)
        Level for track05.cdda.wav: -8.1000dBFS (-0.0003dBFS peak)
        Level for track06.cdda.wav: -8.2215dBFS (-0.1754dBFS peak)
        Level for track07.cdda.wav: -8.9346dBFS (-0.1765dBFS peak)
        Level for track08.cdda.wav: -13.6175dBFS (-0.4552dBFS peak)
        Level for track09.cdda.wav: -9.0107dBFS (-0.1778dBFS peak)
        Level for track10.cdda.wav: -8.1824dBFS (-0.4519dBFS peak)
        Level for track11.cdda.wav: -8.5700dBFS (-0.1778dBFS peak)
        Standard deviation is 1.47 dB
        Throwing out level of -13.6175dBFS (different by 4.58dB)
        Average level: -8.6929dBFS
        Applying adjustment of -3.35dB...
The volume of "Ten Cents," which is track 8, is 4.58 decibels off the average, which, given a standard deviation of 1.47 decibels, makes it a statistical aberration (which I've defined as anything off by more that twice the standard deviation, but you can set a constant decibel threshold with the -t option). Therefore, it isn't counted in the average, and the adjustment applied to the album isn't thrown off because of one song. Although the aberrant song's volume is not counted in the average, it is adjusted along with the rest of the files.

Example 4. Finally, say you want to make a mixed CD of 80's songs for your mom or something. You won't allow any 80's songs to taint your hallowed mp3 collection, so the absolute volumes of these tracks don't matter, as long as they're all about the same, so mom doesn't have to keep adjusting the volume. For this, use the mix mode option,

        normalize -m *.wav
and each track will be adjusted to the average level of all the tracks.

3. How does normalize work?

A little background on how normalize computes the volume of a wav file, in case you want to know just how your files are being munged:

The volumes calculated are RMS amplitudes, which correspond (roughly) to perceived volume. Taking the RMS amplitude of an entire file would not give us quite the measure we want, though, because a quiet song punctuated by short loud parts would average out to a quiet song, and the adjustment we would compute would make the loud parts excessively loud.

What we want is to consider the maximum volume of the file, and normalize according to that. We break up the signal into 100 chunks per second, and get the signal power of each chunk, in order to get an estimation of "instantaneous power" over time. This "instantaneous power" signal varies too much to get a good measure of the original signal's maximum sustained power, so we run a smoothing algorithm over the power signal (specifically, a mean filter with a window width of 100 elements). The maximum point of the smoothed power signal turns out to be a good measure of the maximum sustained power of the file. We can then take the square root of the power to get maximum sustained RMS amplitude.

As for the default target amplitude of 0.25 (-12dBFS), I've found that it's pretty close to the level of most of my albums already, but not so high as to cause a lot of limiting on quieter albums. You may want to choose a different target amplitude, depending on your music collection (just make sure you normalize everything to the same amplitude if you want it to all be the same volume!).

Regarding clipping: since version 0.6, a limiter is employed to eliminate clipping. The limiter is on by default; you don't have to do anything to use it. The 0.5 series had a -c option to turn on limiting, but that limiter caused problems with inexact volume adjustment. The new limiter doesn't have this problem, and the -c option is considered deprecated (it will be removed in version 1.0).

Please note that I'm not a recording engineer or an electrical engineer, so my signal processing theory may be off. I'd be glad to hear from any signal processing wizards if I've made faulty assumptions regarding signal power, perceived volume, or any of that fun signal theory stuff.

4. Why don't you normalize using peak levels instead of RMS amplitude?

Well, in early (unreleased) versions, this is how it worked. I found that this just didn't work well. The volume that your ear hears corresponds more closely with average RMS amplitude level than with peak level. Therefore, making the RMS amplitude of two files equal makes their perceived volume equal. (Approximately equal, anyway: certain frequencies sound louder at the same amplitude because the ear is just more sensitive to those frequencies. I may try to take this into account in a future version, but that opens up a whole new can of worms.)

"Normalizing" by peak level generally makes files with small dynamic range very loud and does nothing to files with large dynamic ranges. There's not really any normalization being done, it's more of a histogram expansion. That said, since version 0.5, you can use the --peak option to do this in normalize if you're sure it's what you really want to do.

5. Can normalize operate directly on mp3 files?

Version 0.7 and up can operate directly on MPEG audio files. An mp3 file is decoded (using Robert Leslie's MAD library) and analyzed on the fly, without the need for large temporary WAV files. The mp3 file is then "adjusted" by setting its relative volume adjustment information (technically, an "RVA2" frame is set in its ID3v2 tag). The advantage of this method is that the audio data doesn't need to be touched, and you don't incur the cost of re-encoding. The disadvantage is that your mp3 player needs to read and use relative volume adjustment ID3 frames. The normalize distribution now includes a plugin for xmms that honors volume adjustment frames. If you use an mp3 player other than xmms, you'll have to bug the author to support RVA2 frames in ID3 tags.

If you'd rather change the volume of the mp3 audio data itself, you still have to decode to WAV, normalize the WAV, and re-encode. A script, normalize-mp3, is included in the normalize distribution to do this for you.

6. Can normalize operate on ogg vorbis files?

Version 0.8 will at least be able to read vorbis audio files. Adjusting is harder, though: the problem is that, unlike with ID3, as far as I know there's no standardized volume adjustment tag for ogg. I could just use, say, "VOLUME_ADJUST=X.XXdB" as an ogg comment, but there would be no reason for players to support it.

It may be possible to twiddle the vorbis data itself to alter the volume in a lossless way. I'm looking into this, but it would be a big undertaking, not something that would be finished anytime soon.

The current situation is that you have to decode to WAV, normalize the WAV, and re-encode. The normalize-ogg script is included in the normalize distribution to do this for you.

7. How do I normalize a whole tree of files recursively?

The "unix way" to do this is to use find:

find . -type d -exec sh -c "normalize -b \"{}\"/*.mp3" \;
will go directory by directory, running normalize -b on all mp3 files in each. If you don't want batch mode, just:
find . -name \*.mp3 -exec normalize {} \;
will run normalize on each mp3 file separately. If you want to run normalize in batch or mix mode on all files in the directory tree, use:
find . -name \*.mp3 -print0 | xargs -0 normalize -b

A built-in recurse option has been a very popular request, so I'm adding support for it in version 0.8.

8. How do I use normalize in Windows?

"I click on INSTALL but nothing happens. What's wrong?" Okay, here's the deal: normalize is free software, written for free operating systems such as Linux and FreeBSD. These happen to be unix-style operating systems, so normalize generally works on other non-free flavors of unix as well. Unlike Windows software, unix software such as normalize is meant to run on many different operating systems on many different architectures, so usually it comes in source code form and you have to compile it for your particular setup. If you are running some form of unix, normalize should compile right out of the box (let me know if it doesn't!). For other operating systems, such as Amiga, BeOS, OS/2, or Windows, you may have to jump through some hoops to get it to compile.

A discussion of compiling unix software for Windows is way beyond the scope of this FAQ, but here's a quick rundown:

  1. You first need the Cygwin toolkit. After installing, start up a cygwin bash shell.

  2. Go to the directory where you unzipped the normalize archive -- it would be named something like normalize-x.y.z.

  3. Type "./configure", then "make", then "make install"

  4. If there were no errors, you can run normalize by typing "normalize" at the prompt. Normalize is a command-line utility, so you have to pass it command line options. Run "normalize --help" for a synopsis.


Copyright © 1999--2005, Chris Vaill <chrisvaill at gmail>

Permission is granted to copy, distribute, and/or modify this document under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.