Note for Amiga users: Please download and install the ixemul environment, available on Aminet under the name 'ixenv.lha'. It makes easier to install and increase compatibility with programs that use ixemul.library. Email me if you have problems: fr3dy@retemail.es ------------------------------------------------------------------------ This is release 0.6.1 of Normalize, a wave file volume normalizer. Copyright (C) 1999--2001, Chris Vaill Normalize is an overly complicated tool for adjusting the volume of wave 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. To build, just run: ./configure make make install See the file INSTALL for more extensive directions. See the man page, normalize.1, for usage. Send bug reports, suggestions, comments to cvaill@cs.columbia.edu. normalize is free software. See the file COPYING for copying conditions. ------------------------------------------------------------------------- 1 What platforms does this work on? I've tested normalize on GNU/Linux and FreeBSD on x86 and Solaris on sparc. I've heard that it works on GNU/Linux on alpha and on BeOS R5. As far as Windows is concerned, you may be able to compile it using the cygwin toolkit (see http://www.cygwin.com/), but I don't know. 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 this useful for? 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 differ­ ent 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 nor­ malized 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" ... 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. 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 ver­ bosity 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 deci­ bels 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 stan­ dard deviation, but you can set a constant decibel thresh­ old 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. 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 vol­ umes 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 it work? This is just 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 corre­ spond (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 maxi­ mum 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 dif­ ferent target amplitude, depending on your music collec­ tion (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 lim­ iter 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 elec­ trical engineer, so my signal processing theory may be off. I'd be glad to hear from any signal processing wiz­ ards 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 vol­ ume that your ear hears corresponds more closely with average RMS amplitude level than with peak level. There­ fore, making the RMS amplitude of two files equal makes their perceived volume equal. (Approximately equal, any­ way: certain frequencies sound louder at the same ampli­ tude because the ear is just more sensitive to those fre­ quencies. 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 normal­ ization 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. 5 Can you make normalize operate directly on mp3 files? The short answer is maybe -- I'm looking into it. MPEG layer III audio files store audio data as Huffman- coded frequency components. Theoretically, frequency com­ ponents can be analyzed and adjusted just as easily as time-domain samples (such as the PCM samples in a WAV file). However, because the samples are Huffman-coded, it is likely that decoding the samples, adjusting them, and re-encoding them will cause them to take more or less bits than before. This is a problem, since MPEG audio files have a set bitrate. Layer III files have a bit reservoir mechanism that could absorb some of the changes, but essentially there's no guarantee that adjusted audio data will still fit in the MP3 file. The process of getting the data to fit in the given bitrate is basically the same as the encoding process. That is, it's hard. A simpler task, which normalize may be able to do soon, is to analyze an MP3 file, and set its volume adjustment ID3 tag instead of actually adjusting the audio data. The only problem with this method is that I don't know of any MP3 players that honor this tag. The current situation is that to normalize an mp3 file, you can decode it to WAV, normalize the WAV, and re-encode it, but you incur the encoding loss a second time. If you don't mind the loss, you can do just this with the "normalize-mp3" script that is included in the distribution.