Multi-pass MP3 encoding - listen for yourself

[A.S.]

There have been other 'hot topics' here for the past few weeks, but as requested - a nudge further up the pile for this one.

thanks Alan

Oops! Sorry - glad you reminded me and I feel a bit guilty now!

OK here is the outline of what I want to illustrate. Lossless files are ones where every single bit of data is captured and retained in the file. You can think of a lossless file as an Excel spreadsheet where every dot, every number, every equation must be saved and recovered without the slightest change. Imagine if you sent a rocket to Mars and the trajectory calculations were on a spreadsheet .... but someone decided to save disk space by saving the file in a lossy format ... even the change of one single bit of data could send the mission off in the wrong direction.

Audio and video files in their original bit-perfect (.WAV, .AVI) format are simply huge. So huge that they are very awkward to pass around, even now in with optical broadband and quite impossible in the dial-up days. So lossy audio/video compression was invented which made the reasonable assumption that a loud sound masked a quiet sound to the human ear, so no point coding-up and allocating precious data bits to the quiet, submerged sounds. So MP3 and WMV and similar formats were invented and opened up the entire concept of moving audio/video around the internet.

But as we've noted before, something has to give. If you discard 90% of the original audio (as MP3 typically does) you make a file that's only 1/10 as big. That's very handy. But what happens if that first generation lossy file is then re-compressed with the same lossy settings, and that second generation re-compressed with the same settings and so on. Easy: I've present the original wav file (I have to convert it to 320kb to play here), then the third, seventh and 12th generation re-compressed versions of the file.

Loading the player ...

The source file, first generation

Loading the player ...

Second generation, the first time through the MP3 encoder at 128kb

Loading the player ...

Third generation, the second time through the MP3 encoder at 128kb

Loading the player ...

Seventh generation, the sixth time through the MP3 encoder at 128kb

Loading the player ...

12th generation, the eleventh time through the MP3 encoder at 128kb

Note: the re-encode settings are identical for each generation. That makes this test unrealistically generous. In the real world, each time the file is re-encoded the new user may have increased or decreased volume fractionally (or a lot), the file may have been edited or eq applied or the re-encode settings may be even more severely lossy at say 96kb. In each case, the encoder will attack what remains of the audio afresh and discard even more audio data. So these examples below are definitely a best-case scenario.

What do you think of the sound? If the file had remained in .WAV format (or similar lossless formats I assume) we could have made 1000 generations of copy and the sound would be identical because the bits describing the sound would have been identical first generation to 1000th generation.

I thought I'd make a few more passes, so here is the delight of the 20th generation ...

Loading the player ...

20th generation, the nineteenth time through the MP3 encoder at 128kb

Incidentally, I've noticed again that after the third or fourth re-encode cycle, the file size actually starts to slowly increase again.

There are a couple of important points to make about this:

1. If there is any possibility that your source file will be opened, edited and then saved again DO NOT save it in a lossy format. Pluto can illuminate more, but in broadcasting this is a serious issue when audio files can pass through multiple hands before being transmitted.

2. If you must use lossy files as part of the work flow, keep a backup a first generation in LOSSLESS format.

3. Don't even consider ripping your lossless .WAV CD files to any form of lossy compression: you are sure to regret it. See you in Munich over the next few days.

================================================== ======
P.S. Follow-up to this post

May 2012: I had dinner with EveAnna of Manley Labs and mentioned that I'd used this recording for the above clip, knowing that it credited Manley microphones on the cover. She said that, years after this particular recording studio was decommissioned the microphones were found and re-checked and their technical measurements were far from the production standard. I commented that, technically perfect or not, the sound is lovely and lush and that the acoustics of the studio would have readily swamped any mismatch in the mics. The centre mics were stereo cardioids accompanied to L/R omnidirectional mics, which give that spacious sound.

If you are into tube amps, I have met EveAnna (see picture attached) several times over the years and have been hugely impressed by her pragmatic customer-focused engineering for tube products which are, inevitably, because they run hot, demanding of component choices. Some of the real-world issues that tube people have to deal with (fake tubes, sourcing problems generally) are seriously stressful but I'd feel confident that if tubes are your thing, you're unlikely to get a better designed tube amp with better components and after care anywhere. It was a real eye opener for both Pluto and myself just how much unseen effort is involved in top class tube amplifiers - far, far removed from the sort of kit you see on Ebay selling for almost nothing and with no back-up.

If you are going to invest in a tube (valve) amp, be sure it's from a supplier with real, credible back-up. They are extremely specialist and complex products to design, manufacture and service to close tolerances.

Picture of our gang in Munich setting off to dinner ....

>

[A.S.]

I'd be interested to know what you think of the MP3 examples.

[EricW]

I'd be interested to know what you think of the MP3 examples.

One of the things I found interesting is that the lead guitar seems to be the part of the recording least affected by the successive levels of compression. What I noticed most was the decrease in the sense of air or space with each re-encode. It was particularly evident in the percussion. By #20 the lead was still recognizable, but the percussion was a swampy, soupy mess, and sounded as though the phase was constantly shifting (don't know exactly what it was, but when I played electric guitar eons ago and ran it through a phase shifter, that's pretty much what it sounded like).

There also seem to be a successive loss of signal level; later generations sounded quieter to me than earlier ones.

[A.S.]

One of the things I found interesting is that the lead guitar seems to be the part of the recording least affected by the successive levels of compression. What I noticed most was the decrease in the sense of air or space with each re-encode. It was particularly evident in the percussion...

This could be because the guitar generates a strong fundamental note with just a few harmonics, more like a sine wave plus its overtones. But the drums generate more of an impulsive sound, and we know that if we technically decode an impulse we discover that it can contain many frequencies (fundamentals plus harmonics) of perhaps even equal strength right across the audio band.

In fact, it struck me that after several re-encodes that the high hat progressively sounded like modulated hiss than an actual instrument and had developed an electronic regularity with its 'shush.... shush.... shush' sound. If that's the case it should be a simple matter to take a random noise generator, connect it to a volume control, and rotate the control up and down on a cycle of a second or two with similar sonic outcome to the multi-encodes. I'll find the time to try it.

Later ....

Here it is. I took pink noise, modulated it at a period of about one second, chopped the HF off. There is some similarity to the sound of the multi-passed MP3 encode/decode ...

Loading the player ...

Modulated pink noise, not music

The implication of this must be that when music passes through a lossy data reduction system, by a process of entropy it will be progressively reduced to nothing more than random noise.