Recorded music doesn’t sound as good as it used to. Recordings sound muddy, clipped and lack punch. This is due to the ‘loudness war’ that has been taking place in recording studios. To make a track stand out from the rest of the pack, recording engineers have been turning up the volume on recorded music. Louder tracks grab the listener’s attention, and in this crowded music market, attention is important. And thus the loudness war – engineers must turn up the volume on their tracks lest the track sound wimpy when compared to all of the other loud tracks. However, there’s a downside to all this volume. Our music is compressed. The louds are louds and the softs are loud, with little difference. The result is that our music seems strained, there is little emotional range, and listening to loud all the time becomes tedious and tiring.
I’m interested in looking at the loudness for the recordings of a number of artists to see how wide-spread this loudness war really is. To do this I used the Echo Nest remix API and a bit of Python to collect and plot loudness for a set of recordings. I did two experiments. First I looked at the loudness for music by some of my favorite or well known artists. Then I looked at loudness over a large collection of music.
First, lets start with a loudness plot of Dave Brubeck’s Take Five. There’s a loudness range of -33 to about -15 dBs – a range of about 18 dBs.
Now take a look at a track from the new Metallica album. Here we see a dB range of from about -3 dB to about -6 dB – for a range of about 3 dB. The difference is rather striking. You can see the lack of dynamic range in the plot quite easily.
Now you can’t really compare Dave Brubeck’s cool jazz with Metallica’s heavy metal – they are two very different kinds of music – so lets look at some others. (One caveat for all of these experiments – I don’t always know the provenance of all of my mp3s – some may be from remasters where the audio engineers may have adjusted the loudness, while some may be the original mix).
Here’s the venerable Stairway to Heaven – with a dB range of -40 dB to about -5dB for a range of 35 dB. That’s a whole lot of range.
Compare that to the track ‘supermassive black hole’ – by Muse – with a range of just 4dB. I like Muse, but I find their tracks to get boring quickly – perhaps this is because of the lack of dynamic range robs some of the emotional impact. There’s no emotional arc like you can see in a song like Stairway to Heaven.
Some more examples – The Clash – London Calling. Not a wide dynamic range – but still not at ear splitting volumes.
This track by Nickleback is pushing the loudness envelope, but does have a bit of dynamic range.
Compare the loudness level to the Sex Pistols. Less volume, and less dynamic range – but that’s how punk is – all one volume.
The Stooges – Raw Power is considered to be one of the loudest albums of all time. Indeed, the loudness curve is bursting through the margins of the plot.
Here in one plot are 4 tracks overlayed – Red is Dave Brubeck, Blue is the Sex Pistols, Green is Nickleback and purple is the Stooges.
There been quite a bit of writing about the loudness war. The wikipedia entry is quite comprehensive, with some excellent plots showing how some recordings have had a loudness makeover when remastered. The Rolling Stone’s article: The Death of High Fidelity gives reactions of musicians and record producers to the loudness war. Producer Butch Vig says “Compression is a necessary evil. The artists I know want to sound competitive. You don’t want your track to sound quieter or wimpier by comparison. We’ve raised the bar and you can’t really step back.”
The loudest artists
I have analyzed the loudness of about 15K tracks from the top 1,000 or so most popular artists. The average loudness across all 15K tracks is about -9.5 dB. The very loudest artists from this set – those with a loudness of -5 dB or greater are:
|Queens of the Stone Age||-3.23|
|All That Remains||-3.56|
|Fall Out Boy||-3.97|
|30 Seconds To Mars||-4.08|
|System of a Down||-4.51|
|My Chemical Romance||-4.81|
It is interesting to see that Avril Lavigne is louder than Metallica and Katy Perry is louder than Megadeth.
The Quietest Artists
Here are the quietest artists:
|The Velvet Underground||-14.05|
|Simon & Garfunkel||-14.03|
|The Very Hush Hush||-13.73|
|The Smashing Pumpkins||-12.83|
|A Tribe Called Quest||-12.52|
(note that I’m not including classical artists that tend to dominate the quiet side of the spectrum)
Again, there are caveats with this analysis. Many of the recordings analyzed may be remastered versions that have have had their loudness changed from the original. A proper analysis would be to repeat using recordings where the provenance is well known. There’s an excellent graphic in the wikipedia that shows the effect that remastering has had on 4 releases of a Beatles track.
Loudness as a function of Year
Here’s a plot of the loudness as a function of the year of release of a recording (the provenance caveat applies here too). This shows how loudness has increased over the last 40 years
I suspect that re-releases and re-masterings are affecting the Loudness averages for years before 1995. Another experiment is needed to sort that all out.
This table shows the histogram of Loudness:
Average Loudness per genre
This table shows the average loudness as a function of genre. No surprise here, Hip Hop and Rock is loud, while Children’s and Classical is soft:
|Funk / Soul||-9.83|
|Folk, World, & Country||-11.32|
|Stage & Screen||-14.29|
So, why do we care? Why shouldn’t our music be at maximum loudness? This Youtube video makes it clear:
Luckily, there are enough people that care about this to affect some change. The organization Turn Me Up! is devoted to bringing dynamic range back to music. Turn Me Up! is a non-profit music industry organization working together with a group of highly respected artists and recording professionals to give artists back the choice to release more dynamic records.
If I had a choice between a loud album and a dynamic one, I’d certainly go for the dynamic one.
Update: Andy exhorts me to make code samples available – which, of course, is a no-brainer – so here ya go: volume.py
#1 by Dan on March 24, 2009 - 11:17 am
There is another website devoted to the same subject, which also has an app you can download to measure the dynamic range of music yourself.
#2 by plamere on March 24, 2009 - 11:30 am
@dan – I hadn’t seen that site before … thanks for the link.
#3 by Andy Baio on March 24, 2009 - 11:22 am
Publishing your code samples, even for these small simple experiments, would really help people get acquainted with the Echo Nest API.
#4 by plamere on March 24, 2009 - 11:25 am
Good suggestion …. I’ll do that.
#5 by Nelson on March 24, 2009 - 11:58 am
Great analysis, thanks for posting it! One question; isn’t the range of the music more important than the mean loudness? I’m not sure if I care if the average of a track is -9.5 dB, what I care is that the entire track is compressed to -11 dB to -8dB.
Maybe a chart of standard deviation of loudness for each song would reveal something?
#6 by plamere on March 24, 2009 - 1:08 pm
@nelson – Good point – although since the loudness is averaged over frames, I worry about clipping for songs that don’t leave much headroom.
#7 by alf on March 24, 2009 - 3:20 pm
Code for drawing the graphs would be nice too, but it’s easy enough with Google Spreadsheets:
#8 by plamere on March 24, 2009 - 3:23 pm
I just gnuplot to make the plots:
gnuplot> set title “Loudness for song”
gnuplot> set xlabel “time”
gnuplot> set ylabel “dB”
gnuplot> plot “song.dat” using 1:2
#9 by Aaron on March 24, 2009 - 4:11 pm
I’m with Nelson; it seems like dynamic range is the real question. Something which only uses the space between -8db and -5db isn’t going to sound any worse if it’s boosted to be between -4 and -1.
#10 by plamere on March 24, 2009 - 4:55 pm
@Aaron – perhaps, but I’m not 100% sure … for one thing, the dB scale (and how the human ear/mind perceive loudness) is not linear, so I’m not sure if a 3db range at -8db to -5db is the same as between -4db and -1db (but I bet we can find someone who knows….). Second. I think that there may be some transient peaks that are more likely to be clipped if the sound is close to the edge.
#11 by GlebErty on June 20, 2009 - 12:44 am
Compression is not the same as normalizing, at all. Making quiet parts louder amplifies noise, but a studio commercial recording with audible noise in 2009 is a sort of nonsense, so the noise gate is applied, often masking silent parts at the same time; high level spikes are compressed too, making sound flat etc. There is nothing wrong with compression technology itself – the problem is in incredible high levels of compression in contemporary releases.
#12 by Joseph Rabb on March 25, 2009 - 1:02 am
I wish this wasn’t the same old argument, but I’ve (no pun intended) heard it before. I wish I could source this, but another argument against over compression I’ve heard is that the “wall of noise” can actually leave the listener feeling fatigued. I’m being clear about it being over compression, because this effect is desired by ‘heavy’ bands because it does lend to a chunky sound that they are looking for. It does drive me nuts however, because bands that do that make me wonder why they bother having a bassist.
#13 by Joseph Rabb on March 25, 2009 - 1:24 am
plamere…Db is on a logarithmic scale, which actually it is linear since exponential functions are linear. The range of sound intensity the human ear can deal with between just barley hearing it to blowing your eardrums right out of your head is HUGE (its based on sustainable pressure) so it’s easier to rate it as a logarithm. 3dB is 3dB where ever you look. Don’t quote me, physics was 30 beers ago so feel free to refute.
#14 by plamere on March 25, 2009 - 4:41 am
@Joeseph – even though the db scale is linear, I don’t think the human perception of dB is linear – Fletcher-Munson curves are a good representation of how the human ear/brain perceives loudness – and these curves are non-linear: http://en.wikipedia.org/wiki/Fletcher-Munson_curves
#15 by taverner on June 25, 2009 - 2:30 am
fletcher-munson curves are a plot of the human ears frequency response at different amplitudes. the human ear doesnt have a flat frequency reponse, and thats what fletcher-munson curves prove
the dBSPL scale was invented to convert a non-linear set of numbers (the measure of air pressure (unit: pascals)) into a linear one. Joseph rabb is right, 3dB is 3dB no matter where in the dynamic range of human hearing you’re talking about.
#16 by james on March 26, 2009 - 6:04 pm
YEAH V-SNARES lol it is INTENSE THOUGH
#17 by enrique on March 26, 2009 - 7:09 pm
It’s funny that venetian snares is the #1 loudest artist since his music is literally just loud noise. (BTW I’m a fan of his music, don’t get me wrong).
I’m really annoyed at the loudness trend though. I believe I read about this somewhere a couple of years ago and it’s was kind of I knew there was something odd about some music, but I couldn’t put my finger on it. After reading about this it all made sense.
Now to figure out how we undo this thing.
#18 by lza on March 27, 2009 - 2:41 am
And you used MP3s to do this? MP3s you encoded yourself from a CD-Rip or MP3s downloaded from the internet? Because this is all completely useless info when you use MP3 because of something called MP3-Gain. Give me some of your files and I can make Dave Brubeck clip harder than Metallica or Muse together and make Venetian Snares appear even quieter than brian eno.
You failed, dude.
#19 by Iza's Mom on December 1, 2009 - 1:05 pm
You failed, retard. You can’t compress sound with “MP3_Gain” (Replay Gain).
#20 by Ian Shepherd on March 27, 2009 - 9:12 am
Interesting post ! Do you know what measure is used to judge the loudness ? RMS ? A-weighted RMS ? LEQ ? Or…?
I think the average-loudness-by-artist thing might be misleading – Metallica are lower than Avril Lavigne because some of their early albums are great, dynamic records, but all of hers are recent and over-compressed. Might be worth pointing that out.
The overall loudness is a bit of a red herring, too – people use their volume controls. If the Metallica, Stooges & Pistols examples were played back at equivalent volumes, their loudness would feel similar.
Finally, interesting as this analysis is, it doesn’t take into account HOW the records were made loud. For example, Anarchy sounds distorted but still better than “Death Magnetic”, because it hasn’t been digitally clipped. Analogue clipping is preferable to digital, on the whole.
On the other hand, “Some Might Say” by Oasis *has* been digitally clipped but doesn’t sound as crushed or over-compressed.
I posted some similar discussion a while back here.
#21 by Douglas Eck on March 29, 2009 - 10:03 am
Great post Paul. As for others in this thread, for me loudness is inversely related to music I like. I guess it’s another audio feature that predicts something interesting, but for boring reasons, like zero crossing rate.
@Iza : I would be worried about gain on the MP3s if Paul had encoded them purposefully for this demo and cheated. If he’s just sampling MP3s from the wild then I can’t imagine there will be any reliable relation between the use of gain and artist or genre. Anyway, this effect is well-documented and is well understood by anyone who knows how music is produced to arise from the overuse of compression in mastering, not from MP3 encoder parameters.
#22 by plamere on March 29, 2009 - 12:55 pm
@doug – thanks for the defense against Iza’s comments, I usually just ignore those type of comments.
#23 by LJ on March 29, 2009 - 11:52 am
@plamere – spooky, I did something almost exactly the same on exactly the same day last week. You can see my results here, with links to the source: http://www.flickr.com/photos/etdm/3383452670/
I found that you need to ignore anything pre 1990 to catch remastered releases, or ignore values falling outside +/- 1 standard deviation from the mean for each year. You need a rather large sample of source files though if you’re going to do the latter.
#24 by plamere on March 29, 2009 - 12:53 pm
@lj Indeed, very spooky – I saw the same thing too, where certain releases were likely remasters that had been made louder, but I was uncomfortable just throwing away the data that didn’t fit. The best thing to do really is to only deal with tracks where one knows exactly which release it is coming from.
#25 by Joseph Rabb on April 1, 2009 - 1:48 pm
@plamere – thanks for the link. I’m looking at the curves and it’s throwing frequency into the mix. Db is still Db, but your point of perception is taken. You also have to consider the curves are going to change based on per person since hearing is unique to each person. That’s going to leave you a three variable problem with infinite solutions. Ouch. Yeah, most of those solutions are going to fit in a pattern, but I don’t know. Seems like a lot to think about. You might just have to linearize what you have just to make the data manageable.
#26 by Richard Tollerton on April 3, 2009 - 12:23 am
Hi Paul: I just came across your link. Excellent work, particularly with the average loudness vs. time plot.
I’m obliged to link to a free tool I wrote a year ago that also does dynamic range estimation like the Pleasurize tool does, but uses a considerably more useful loudness model, and has a much more comprehensive set of analyses. And was first :)
#27 by D. Prince on April 20, 2009 - 5:16 pm
The issue is fairly old (but is of course still a huge problem). Robinson proposed a solution around 2001 and wrote some code that was integrated into Media Jukebox probably around 2003. http://replaygain.hydrogenaudio.org/ includes code in the links
#28 by mellows on May 30, 2009 - 5:10 pm
Wow. Great article. Your graphs represent the problem better than the normal waveform images. I’ve also become really troubled by this, especially since the YouTube video that showed the level of dynamic compression in Rdiohead’s Nude.
I think recording studios/artists that are producing loud albums need to look at the successful artists of the past and currently that produce ‘quiet’ albums. They can release loud versions to MTV and the radio and whatever but for the people that buy the CDs and like the music… why not keep it good? Glad to see Justin Vernon and Tool way up there. If Tool, as a loud band – i.e. prog metal, meant to be played loudly – can make great, dynamic albums, what’s holding back everyone else in that genre area?
Here’s my blog post with the Radiohead video: http://mellowsmusings.blogspot.com/2009/05/loudness-war.html
#29 by Paul A. Teseny on July 23, 2009 - 10:38 pm
Compression may very well be the reason people (audiophiles) say they can hear a difference between CD’s and records or simply prefer to listen to records.
This happens to be one area where audiophiles do notice something; the problem is most people don’t know what it is. For lack of knowledge, they say is because vinyl sounds better than CDs.
When you go to a live performance, the dynamic range might well be around 100 db for classical music and something less for rock music; albeit the average level for rock music may be higher. This should not be an issue for CDs, since 16 bits is about 96 db (6 db per bit).
According to the IEEE, vinyl records tend to support only a 70 db dynamic range (which is only 11.6 bits). This means that even the best vinyl recordings do not exceed the dynamic range capability of a CD. Sorry Audiophiles – it’s math!
Audiophiles spend a great deal of time listening to recorded material; much more time listening than the average person and really listening to the music, not just using music as background. The common lament among audiophiles is that CDs are less pleasurable to listen to than vinyl records. You hear comments like records sound warmer! I can’t listen to my CDs as long as I can listen to my records. Compression!
I have a business that digitizes records and burns CDs (CD-Rs) for people, mostly audiophiles. In blind tests that I conducted, (and I’m not biased either way – I am an engineer) the listeners (both audiophiles and casual listeners) could NOT distinguish between the record and the CD. Is it real or is it …?
One reason I was given for the compression of CD’s, is that during playback of wide dynamic range material in cars, the sound will drop below audiability during quiet passages when the loud passages are set for normal volume. So, why not put the compressor in the car CD player?
It’s a real shame, that we don’t use the available dynamic range of CD’s. imagine how great recorded music would sound if we did.
#30 by Cantonman on August 6, 2009 - 1:05 pm
In recent testing of HD radio, I discovered that the dynamic range is horrible! Now to determine if it’s the receiver or station that is destroying the signal. I used digital optical out from receiver to digital optical line-in on soundcard. Was expecting great waveform but it’s narrow, loud, with no range. Disappointed but research continues.
#31 by Drew on August 19, 2009 - 7:26 am
Notice it’s all the ones whose listeners probably wouldn’t care about sound quality? Nice comparisons up there.
#32 by Jamie on October 27, 2009 - 1:15 pm
I like the list for the various artists, it would be great if you could review new music based on the merits of the production value. I recently bought the new Chevelle Sci-Fi Crimes and its a crime indeed. You can’t tell if there is a rhythm section…its sooooo loud. Their previous albums aren’t as bad. Anyway, I bought the CD really wishing I hadn’t spent the money. If this crap continues I will stop buying new music. If Paul can manage to review new music when its released then post it…wouldn’t that be a powerfull motivator to record execs to stop putting out noise? Just a thought.
#33 by sad but true on December 23, 2009 - 10:04 am
Actually from your shitty boses etc the shitty mixes sound “acceptable” but once you get some real speakers&pay attention for positioning&room effects you just can’t listen to those overly loud albums anymore.
same goes for headphones, i occasionally dig my old shitty headphones to listen some really bad quality recordings, becouse their “shitty” soundquality waters it down to acceptable.
but hey what can you do when common folk doesn’t pay attention to at all speaker is just a speaker, positioning doesnt matter and recording quality is shait anyways so if they buy better speakers what’s the upgrade?
#34 by Justin on December 29, 2009 - 8:56 am
I just found this site because I’m listening to a Commodores record (vinyl) from 1975 that actually distorts because it’s too loud. The vocal don’t sound that great like that – but by the looks of it it could be worse. Thanks for the research!
#35 by Justin on December 29, 2009 - 9:03 am
As an addition – sounds like a mastering problem rather than a recording problem.
#36 by David on January 7, 2010 - 5:55 am
This could ne another reason why we enjoy a lot listening to CD of 10-15 years ago, as they sound “fresh” with all the instruments and deails, without distortion or quality loss.
#37 by Daniel Saner on January 31, 2010 - 10:44 am
Nice short summary of the issue. What the people responsible for this fail to realise is that while you might catch people’s attention for the first couple of seconds with its loudness standing out against the other tracks (which in itself depends on the idea that not all others are competing in the loudness wars as well), the most common reaction of people is to turn the volume down, switch the music off altogether, change the channel, etc. In addition, listening to loud mastering is tiresome, so while you might stand out among other records, people enjoy (and buy!) the well-mastered music much more. And it becomes even more pointless today: not in the least due to the loudness wars, customers now depend on technologies such as ReplayGain to even out playback volume, and demand for devices supporting these ideas increases.
And those stupid asses still wonder why no one wants to pay money for their music anymore.
#38 by Andrew on February 9, 2010 - 10:36 am
This is great! Thanks for doing this, Paul – I think I’m going to show it to my class today.
One thing, though – are you using an internal echonest library for this? I don’t see the ‘audio’ module in the pyechonest package.
#39 by Paul on February 9, 2010 - 10:44 am
Look at the Track object in pyechonest. Look at the track.loudness() method to get the overall loudness of track and look at track.segments() to get the individual audio segments. Each segment has a loudness as well. Also, look at this alpha API at:
That allows you to get loudness info without having to upload tracks.
#40 by Jimmy on February 10, 2010 - 2:44 pm
Very interesting analysis. It comes as little surprise that recordings are getting louder, but one question I am left with is “Are recordings becoming less dynamic?” It would be really interesting to extend your analysis to see plots of dynamic range vs. recording date, as well as overall loudness vs. dynamic range. It seems to me that over -compression and the resulting loss of dynamic range is more the problem than overall loudness, but the two are probably closely correlated.
Back in the days of vinyl, mastering engineers worked to manage volume so needles would stay in the groove — too much amplitude and the record would be unplayable. Digital formats have no such limitation, so they can be mastered louder.
#41 by Daniel Saner on February 11, 2010 - 6:19 am
Yes they are correlated, because one is a means of achieving the other. The limitation of digital formats is that they have a clearly defined and limited dynamic range (which is, however, pretty awesomely large for CDs). As you amplify your signal you approach a barrier, a clearly defined 0.0dB mark. Anything that exceeds this is just clipped away to fit the specification. So in order to push the signal further and further, compression is used and overused so the average loudness can be increased while minimising clipping.
If not for the compression and loss in dynamics, the loudness war would not be a problem, as you could always level out the difference on your amplifier. But sadly, what’s happening today is labels sacrificing dynamics for perceived loudness.
#42 by audiofile on February 10, 2010 - 6:21 pm
This is a great article. It would be especially helpful if I had a list of the loudest recordings (artist & song title). Does anyone know where I can find such a list?
#43 by Daniel Saner on February 13, 2010 - 7:31 pm
There is an unofficial, user-contributed database collecting dynamic range ratings (using PleasurizeMusic’s DR Meter) for a number of albums.
Low number means bad case of loudness war, although to some degree it also depends on the genre (it can be okay, say, for a Breakcore record). Sort the table by Album DR and you’ve got some nasty offenders right there. Venetian Snares is probably okay acceptable up there, but Metallica’s Death Magnetic for instance is widely considered to be one of the worst-mastered CDs ever.
#44 by witheld on February 13, 2010 - 9:06 pm
Word. It’s time the public became aware of the truth. It’s time the a new generation experiences the true sound and dynamics that have been stolen from their music. It’s time to let every sellout A$R and audio enginot producing this overhyped crap know….
#45 by Flattydog on February 23, 2010 - 8:45 pm
There is no way hip hip is the loudest genre you have not looked all of the genres.
i would say that drum and bass, hard techno and gabba would now be the loudest as the tracks are rammed right up to 0bd a solid brick wall of loudness.
And whats more – they still sound puncy and dynamic – becuase if they didnt, no DJ would touch them !!
#46 by Flattydog on February 23, 2010 - 8:58 pm
sorry i forgot to add paul i suggest you analyze some modern Drum and Bass tracks [for example – sub focus/chase and status/noisia etc] i think you will have to rethink your idea of how loud music is these days .. great article tho v interesting ..
#47 by wesman on March 1, 2010 - 4:34 pm
hi – I loved your post – I’m very interested in production, python and charts so this is perfect for me :)
I ran your volume.py over a few tracks and I’m not clear on what you’re outputting. The second column is the dynamic range, but what is the first column? I assumed it was some indication of time, but I’m not clear on the unit. Is it seconds? samples?
More importantly, when I ran this, I found multiple lines in the output that had the same value in the first column. Can you explain why that would happen?
#48 by Paul on March 1, 2010 - 9:08 pm
wesman – the first column is deci-seconds from the beginning of the song. Some segments are very short, even shorter than a 10th second – which is why, you may see dups in the first column.
#49 by scrag on April 12, 2010 - 2:06 pm
Very good article. I am working on a project related with the loudness war but I’m not so familiar with Echo Nest Remix API. The graphs on this article seems to be useful in my project also. Can you share the code that you used in python with me? I will appreciate so much if you share.
#50 by Cliff M. on April 23, 2010 - 10:32 pm
I think a large part of the loudness wars are the fact that many musicians, newer engineers and the like have not studied or considered “Dynamics” in music. A simple understanding of crescendos and decrescendos is what I think is missing in their music education. Those two basic things in music are the heart and soul or should I say the emotion within. I would say a record could be compressed a little at it’s highest full scale section (Chorus or loudest crescendo) that way all the soft parts would never be in danger of being clipped and the emotion of the song would still exist in the song or piece. Imagine if an orchestra was brick wall limited and you might get an understanding of the importance of De and crescendos. Club music ruined music with that ol’ Bass and Snare peak level stuff. See what Disco did? :)
#51 by Tracey Larvenz on May 20, 2010 - 12:02 am
I’m a post production audio engineer. I’ll never forget the first time I pulled up Oasis back at the start of the loudness war. I double checked to make sure that I didn’t accidentally import a 1k tone because the waveform looked eerily similar – completely filled out and without any peaks. I’d love to see a return to mixing and mastering with dynamic range.
#52 by mavallarino on March 30, 2010 - 5:33 am
same here, 28 “quiets”, of which coincidentally I have the first 15 – sorry Alphex Twin ;-). Loud is only good when you are doing something… this should explain why the “louds” are good to listening while doing activities (sport, gym etc.). Not many have the patience to listen to music.