Nowadays a lot of music is in digital format. One of the most popular digital encoding methods is MPEG-1 Audio Layer III, aka. “MP3“. Typically an MP3 file takes around 1/10 the size of the uncompressed original (depending on quality), thereby saving storage space.

MP3 uses a form of lossy compression. This means that the original sound can not be perfectly recovered. What is recovered (sound you hear) is some kind of approximation of the original.

MP3 does this lossy compression by removing the sounds which a human ear cannot distinguish, using so-called “perceptual coding“. The sounds which are removed are specified by a psychoacoustic model, which is created using human input, for example listening tests.

In other words, this means that many people are all the time listening to music with “something missing”. But what is this something, what does this something sound like?

I woke up this Sunday morning and thought to myself, now there’s an interesting question! So, I set about to find out the lost sound.

The procedure to do this is quite simple: take an uncompressed original sound (e.g. WAV format with linear PCM encoding); call this a, encode it to MP3, decode it back to linear PCM; call this b. Then, subtract b from a. The result is the signal which MP3 encoding loses forever.

My test track was a Cuban hip-hop track “300 Kilos (Featuring Yuri Buenaventura)” by Orishas, from the album Emigrante. I chose this track because it’s has a variety of different sounds: human speech, high-hats, bass, etc. and besides, it’s kind of dope.

I did the test with two quality settings: 128 kbits per second, and with a variable bit rate (VBR). I used the Lame encoder, using the command line options “-b 128 -h” and “-V 3″, respectively.

I have provided below short 20-second test clips. In order to save space, I’ve encoded the test clips back to MP3 format. The sound quality was not altered in a very perceivable way for the new MP3 encoding (as MP3 encoding was truly doing its job!), the point is to demonstrate what this missing audio sounds like.

• Original clip, high-quality VBR MP3 (lame -V 5)
• Sounds lost when MP3 encoding with 128 kbps (lame -b 128 -h)
• Sounds lost when MP3 encoding with VBR (lame -V 3)

As you can hear, the better the quality of the MP3 encoding, the more silent the lost sound is. This makes sense. But note, even with the best setting you can get (not included in examples), MP3 is lossy - at least some of the original audio is lost.

If you want to create a digital collection of your own CDs, using formats such as FLAC is the way to go. FLAC is lossless, and therefore you can recover the original sound exactly. By having your files as FLAC, you can later decode the FLAC to linear PCM, and re-encode to whatever format is handy.

My test audience of one person thought the missing audio sounded very “satanic”, something like in 1970-1980s Hollywood horror movies. From this observation we can present a follow-up question: does MP3 encoding remove the evil from your audio files? If this hypothesis holds, then encoding reversed tracks of Led Zeppelin should produce silence or near-silence. Testing this hypothesis shall be left for future research.