Recently I'd been reading some Sound & Vision magazines, and had seen Ken Pohlman assert several times that ATRAC drops everything over 16kHz when the going gets tough, thus being unsuitable for critical listening, though OK for portables and teenagers - or words to that effect.
This aroused my curiosity, and I wanted to have a closer look at what ATRAC did do in terms of simplification of the original signal.  Note that this has little to do with the audibility of the compression.  That's an entirely different kettle of fish, and a personal matter.  This is just an attempt to put a face to the names we hear bandied around about psychoacoustic compression.

These plots were produced with Spectrogram 5.1.7, a fine freeware FFT spectrum analyser.  So have a go yourself.  I wouldn't mind getting some standardised test CDs established, and having a generational / manufacturer comparison.  I'm also quite interested in seeing how the LP2 and LP4 formats go about their extra compression.  Yes it's all very inane and useless, but if you want to make some plots and post them somewhere, or send them to me, or ask me something else, go for it.

Unfortunately Spectrogram is no longer freeware. However you may still be able to download an older version from here. You might also find Avisoft Lite to be of use, although the spectrogram is not as detailed.
 
 

Full Range Spectral Analysis

Here are some spectragrams of a short audio sample that starts with a woman's voice, then a strange whooshing noise, then a pop song (it is in fact the beginning of "Bad Communication", a B'z Led Zeppelin ripoff).  If you've not seen a spectragram before, time is the X axis & frequency is the Y axis, so you get a picture of what is happening frequency wise wrt time.  The original wav file is shown along the top as amplitude vs time.  In these two plots, colour denotes the power level at a given frequency, where black is zero, going up through the blues, into the greens, yellows, and eventually would get to red, but not much here.

In the first two plots, the X axis is 8 seconds long & the Y axis stretches from 22Hz - 22kHz.  The Y axis is linear, not logarithmic, so a point halfway up the plot is 11kHz.

The CD

 

ATRAC 4? (Recorded on an R50)

 

Well they look pretty similar, I was expecting more difference to be honest.  But in the last 1.5 seconds, there's a fair bit of difference creeping in at the high end, and it does look like some stuff above 16kHz is being sacrificed to preserve the complexity at the lower frequencies, but by no means is it all dropping out.  The plots are not very high resolution, so I zoomed in to focus on the 1.5 seconds, and only the range from 16kHz up to 22Khz.  This gave more detail, but enormously increased the size of the plot, so this time I've tipped them over, leading to...
 

Limited Range Higher Detail Spectral Analysis
 

Now time progresses from top to bottom, and frequency from left to right, the leftmost frequency as stated about 16kHz, progressing up to 22kHz.  If you scroll over you'll see the amplitude time on the far right.  And I've gone into black & white as it seemed to highlight the banding better.

This is more revealing.  The CD has a lot of discrete frequency information above 16kHz.  Sometimes ATRAC is able to make a fair representation of it, sometimes there's more important (read audible) stuff going on down below, so ATRAC drops some, and makes broad smears of others.
 

The CD

ATRAC 4?

Get your own Free Home Page