The Best Lossy Music CompressionSun, Feb 21, 2016 in Music Reviews
I noticed recently that Google Play Music has a set of features I wasn’t aware of before, and decided to give it a try. Through various kinds of yak shaving, I ended up tackling a project I’d wanted to look into for a while: which formats and settings are really the best for lossy music compression?
What I learned was both fun and surprising, and ultimately highly practical. If you’re in a hurry, the summary is that high-quality variable-bitrate MP3 produced with the LAME encoder is probably the best all-around choice if you want broad compatibility; if you want the best sound quality, though, AAC (Apple’s native format) or Ogg Vorbis are much better than MP3.
But before I get to that, let’s sharpen some yak razors, shall we?
I have a large library of CDs. Although I listen to more music on Spotify now than on CD, many of my CDs are esoteric stuff that is not available on Spotify. For years I’ve tried keeping it in various music libraries and players so I can listen on-the-go. It was never a satisfactory solution. More recently I’ve been an Apple user, and gradually migrated all of my music into iTunes. Unfortunately, my collection is too large to sync to any mobile device or work laptop, so unless I’m at home, I can’t listen to my music.
Reevaluating Google Play Music
I’ve been aware of several services that let you upload your music library to the “cloud,” but was never happy with any of them. It always seemed that I could get most of what I wanted, but the lacking features were deal-killers in some way, so the situation remained: Spotify on the go, my personal library at home. As far as I could see, the choices looked kind of like this:
- Spotify: great for subscription-based streaming, cluttered UI, doesn’t have all my music, doesn’t support uploading my personal music.
- Apple services and devices: doesn’t fit all my music, too expensive, Music Match doesn’t seem to actually match anything, no subscription-based streaming of stuff I haven’t purchased (although that’s changed recently). It’s been a while; maybe I should take another look at this too.
- Amazon Cloud Player: doesn’t have a good library of stuff to stream, clunky UI, uploading features don’t work well (matching of songs doesn’t work).
- Pandora, et al: niche.
- Google Play Music: wasn’t really in my sphere of consciousness.
It all started with Neil Young. I just got nostalgic for his album Silver and Gold one day, and I couldn’t find it on Spotify. I am an Amazon Prime member, so I thought to look there, but it wasn’t available. I own it on CD and didn’t want to buy it again. While looking around to see where I could get it (I wasn’t near my home), I stumbled upon Google Play Music again.
Google Play Music turns out to have a surprisingly nice UI on both the web and mobile apps, a good library of music for streaming via subscription, and supports uploading 50,000 songs on the free plan (enough for me). It also supports offline listening via downloading, and permits exporting files that you uploaded. Could it be the perfect service?
Because I’m already a longtime Spotify customer, I didn’t want to switch away fully to another service like Google Play Music, but I thought perhaps it’d be worth trying it out for uploads only, since that adds $0 to my current monthly cost.
So I installed the Google Music Manager, let it upload my library, and went on a trip. While away, I cued up Silver and Gold. “Good to see you, good to see you again,” it started, and I smiled. Good to see you again too, Neil.
But then my smile faded. the sound quality was awful! I was so bummed. You might already know that I’m kind of picky about the sound quality of my music. I’ve taken some care in importing my music into iTunes, and although occasionally I can hear that a song’s quality isn’t quite as good as I’d like, this was another matter altogether. It was really painful.
Someday I’d like to learn about lossy compression in detail by reading Karlheinz Brandenburg’s paper MP3 and AAC explained more carefully, but my current understanding of it is that compression schemes such as MP3, Ogg Vorbis, and AAC discard portions of the sonic information that the human ear is less able to hear, in turn introducing errors that are designed to be hard to hear. These are based on “psychoacoustic models,” which I interpret in my layman’s terms as factors such as the following:
- When the music is loud, it’s harder to hear quiet sounds.
- When the music is complicated, we tend to focus on the big themes, not the details, and we don’t hear well at the high and low frequency limits.
- We don’t hear high and low frequencies well when they’re quiet.
- Compression artifacts can be masked by the music itself (and thus inaudible) if they’re quiet enough.
I know there are a variety of compression artifacts. For some reason, the one my ear picks up the most is what I call a “swishy” sound, which I hear in high frequencies such as sibilant vowels, cymbals, and the like. When these sounds change in amplitude, their frequency spectrum seems to change unnaturally if the compression is too lossy, and I cringe. I may be misinterpreting what’s happening here, but that’s how it sounds to my ears: loud cymbals sound bright, but as they die out, they become muddy.
I’ve also noticed that some songs tend to be hard edge-cases for lossy compression. For example, David Gray’s song Fugitive requires very high quality settings to sound good to my ears. It sounds awful at compression settings that make other songs sound pretty good to me.
All lossy compression works by throwing away information, so it’s not surprising that these effects show up every so often. I accept that as part of the deal. But when it’s as bad as my Neil Young album sounded, it’s unbearable for me.
I would have noticed if my original import into iTunes had sounded this bad. What happened? I speculated that Google Play Music had transcoded the already-compressed file into another format, adding generation loss. I found a help page on Google Play Music’s file format support that gave me a clue: any non-MP3 file is converted to an MP3 file at the same bit rate. Indeed, the song I was listening to downloaded (exported) as a 128k MP3 file, which is not very high quality, especially if it has been reencoded.
It seems straightforward to me that the only way to avoid this problem is to upload MP3 files to Google Play Music in the first place. I haven’t looked at my iTunes library, but I’m sure it is a variety of formats, including lots of AAC files (the iTunes default). I ripped those CDs over many years with many different programs. I’m not eager to throw it away and do it again.
Comparing Lossy Compression Formats
This piqued my curiosity, though: how good are the compression formats, anyway? I have been interested in lossy compression for many years. I remember being fascinated when MP3 first appeared on the scene, and learning about alternative formats such as VQF (TwinVQ), Real Audio. I remember when WMA arrived, and then Ogg Vorbis, and the newest contender, Opus. The claimed benefit is usually better sound quality at lower bit rates (smaller file sizes). My unqualified, untested assumptions were that:
- MP3 wasn’t very good relative to its successors, but improved over time.
- Vendor-specific formats such as WMA were not much better.
- Ogg Vorbis was very good, but would never be widely accepted as a standard.
- AAC I wasn’t sure about. I thought it was a proprietary Apple format; I was wrong.
While browsing around a bit, an additional nuance caught my attention: the major MP3 encoding software is the free/opensource LAME and the proprietary, patented Fraunhofer encoding software. Apparently iTunes uses Fraunhofer software, not LAME, and it seems that lots of people believe LAME produces better sounding MP3 files than Fraunhofer.
At this point I had a thought: leaving aside my subjective opinions of audio quality, it shouldn’t be hard to get a firm answer on how well various formats perform. Just encode, then decode and subtract from the original signal, and look at the resulting errors. Right?
I need to caveat what follows. It is not that simple. To do that misses the point of perceptual audio coding to a large extent. However, it is still fun and interesting to do.
I encoded the first 30 seconds of Fugitive in a variety of formats and settings. Then I inverted the original signal, and added the original and decoded signals together. (It’s necessary to ensure that the two signals are exactly aligned in time, or this won’t work). This is easy to accomplish in Audacity. The result is the encoding error, or roughly the combination of the signal that’s discarded plus signals that the encoder adds (quantization noise and the like). In a way, it’s like the audio version of looking at a photographic negative overlaid with a positive, seeing what “leaks” through.
Listening to the error signal is interesting. With high quality compression, it essentially sounds like some high-frequency static, in which you can recognize the original song’s rhythm. At lower quality, it sounds like a muffled version of the song, with the midrange elided. You can also hear many of the compression artifacts quite easily, such as dropouts in frequency bands. Two of the most dramatic examples are the error signals from these compression settings:
- iTunes using MP3 with custom settings: highest quality, VBR, 96k minimum.
- LAME using 320k constant bit rate.
I combined the channels into mono and amplified the signals by 10dB to make their waveforms clearer, and took a screenshot of the result. In order, these are iTunes AAC defaults, iTunes MP3 at fixed 320k bitrate, iTunes default MP3 settings, iTunes MP3 at highest quality preset, iTunes MP3 at highest quality with VBR at 96k minimum bitrate, LAME at 320k bitrate, LAME at VBR 0 (highest quality) setting, Ogg Vorbis at Audacity’s highest quality setting (10 out of 10), and Ogg Vorbis at Audacity’s default quality setting of 5 out of 10.
The more information in the error, the more of the original music is missing or distorted by added non-music signals. However, it is not true that this means the music will sound better. As I said before, the point of perceptual audio coding as a compression technique is to throw away as much information as possible, thus adding as much error as possible, while trying to keep it below the threshold at which a human can perceive it. If successful, ideally this actually creates a large error signal but the music still sounds good. Poor-quality encoders might produce a small error signal but make the music sound bad. So there isn’t a direct relationship between the error signal and the quality of the music.
Another interesting way to look at the error signal is as a spectrogram:
Here it’s quite clear that all of the iTunes MP3s are missing a lot of bass. Contrast the LAME 320 error spectrum to the AAC spectrum, and you can see that AAC retains a lot more high-frequency data too.
What’s The Best Lossy Compression?
You can probably guess that I’m going to say this, but it depends. AAC and Ogg Vorbis are clearly more advanced than the older MP3 standard, and are capable of producing better sounding audio at lower bitrates. However, that’s not the only thing you might care about:
- Compatibility. I don’t know why Google Play Music doesn’t support AAC natively, but it doesn’t. Re-encoding a file from AAC into “equivalent bitrate” MP3 files is a terrible idea; the resulting music sounds awful. Likewise, I’d be pretty happy if everyone supported Ogg Vorbis, but that’s never going to happen. (Tangent: I’ve heard that Spotify uses Ogg Vorbis.)
- The sweet spot. Although a lot of people seem to believe that 320k MP3 files are the pinnacle, in reality there’s a nonlinear bitrate-versus-quality curve for both MP3 and AAC compression (and probably others as well.) There’s a point of diminishing returns, after which higher bitrates just make the files a lot larger without adding much in perceived sound quality. The paper I linked to earlier has more information on this, and notes that for MP3 the sweet spot is about 128kbit/s, whereas AAC’s sweet spot is a bit lower bitrate but it doesn’t hit as much of a quality ceiling.
- File size. This is the whole point, isn’t it? Constant bitrate files at 320k aren’t very compressed in terms of file size.
On the topic of file sizes, a 40MB song in WAV format compressed about 4x with 320k MP3, about 5x with the default iTunes AAC format, and a bit more than that with LAME’s V0 setting.
A lot of audiophiles use lossless compression (which doesn’t reduce file size much), and produce high-quality MP3 or AAC files from them. For my purposes, I’d like to have
- Only one compressed copy of my files (I still have the CDs, after all).
- Good quality sound even if it’s not perfect.
- Minimal file size.
- Music available everywhere without horrendous reencoding and other problems.
To achieve this, it seems that if I want to use services such as Google Play Music, my best option is to encode to MP3 with the LAME encoder and a high-quality variable bitrate such as v0 or v1. (Constant bitrates bloat files needlessly; lossy compression is designed to achieve a desired perceived quality, which doesn’t always require a lot of bits). However, I have no desire to reencode tens of thousands of songs in this way.
The other option I think I need to consider is using Apple’s music service instead of Google’s. It supports both AAC and MP3 files as I’d expect, and now seems to be a pretty direct competitor to Spotify and Google’s offering. Last time I gave it a trial, though, the music matching feature didn’t match anything, so it was potentially going to upload all of my music, which would have exceeded the limits they had at that point and would have cost me a pretty fair amount of money.
If I used Apple’s services instead, then my choice would be AAC files, not MP3. Although MP3 is more broadly compatible, AAC produces clearly better quality audio in my opinion, without the penalty of larger file sizes. (Update: I now use Apple Music with the Music Library feature, so I get a combination of my personal library with the Apple-hosted cloud-based music catalog. I’ve found this to work really, really well; and I haven’t had an audio quality problem yet.)
In closing, I’ll leave you with the opening lyrics of Fugitive:
Is the answer none of the above…
Image credits: Cymatics
I'm Baron Schwartz, the founder and CEO of VividCortex. I am the author of High Performance MySQL and lots of open-source software for performance analysis, monitoring, and system administration. I contribute to various database communities such as Oracle, PostgreSQL, Redis and MongoDB. More about me.