VQF vs. MP3
Comparison
 June 15, 1998.
by Dejan Milic.
Well, what can I say? VQF sounds better even with the lower bit rate than MP3.

But, the subjective test was not good enough for me. I surfed the Web in quest for some objective stuff, but the only thing I've found is the test image at MP3bench. Of course, ALL of the VQF pages have links to some kind of a "comparison", but it's always the same useless table.

So I decided to put some real data on the Web. I've done some testing myself and here are the results:

I took a sequence of wide spectrum audio data (including the frequencies up to 22 kHz) and I coded it with MPEG Layer3 Producer (128 kbps Joined Stereo) and Yamaha SoundVQ Encoder 2.50b1 (48 kbps/ch. Stereo), with the best quality offered by the coders. Then I wrote a 3D spectrum analyzer in MATCAD, and made contour plots of the spectra before and after coding. And here it is! Colors vary from red (peaks in power spectra) to blue and violet (the lowest signal power). The horizontal axes are normalized time axes (0-174 ms), and the vertical axes are normalized frequency axes (0-22050 Hz). Well, judge by yourselves!
 

 
  1. You can see that the MP3 psychoacoustic model excludes completely some high frequencies (colored blue) when it decides that they are irrelevant. Clearly, VQF designers have decided not to exclude any part of the spectrum.
  2. MP3 preserves power spectra peaks (colored red) very good, but it has its problems with the "green" and "yellow" parts; this can be heard by a careful listener. VQF does not preserve the peaks at the highest frequencies that good, but it beats MP3 at everything else (especially at mid-frequencies).
Bellow is the coding loss of the same sequence for both, VQF and MP3 compression formats. As a reference, lossless coding has 0 dB flat loss, without the "ups" and "downs".
 

 

More examples:

This audio sequence looks pretty good in both types of coding. Keep in mind that the VQF bit rate is 25% lower than MP3.

 

But, VQF is not better at everything! Here is an example of one problem with VQF which I discovered accidentally, only in spectrum analysis (I guess I didn't listen very well). It's the pre-echoes problem, which has been solved in MP3 by a technique called "window switching".

The pre-echoes appear when a quiet sequence is followed by a strong percussive sound (with short attack time) like castanets or a triangle. Such input causes relatively high instantaneous quantization errors, and transform-based coding spreads these errors over the time domain. The solution is to use short coding blocks (windows) to limit the echoes spreading, and then switch back to longer blocks after.

 
Looking the pictures above, it seems that the VQF team did not develop very efficient method to battle pre-echoes. Of course, this may be a drastic audio sequence, but again, it's the real music. On the other hand, MP3 is O.K. from this point of view (except in the highest frequencies).
 
 
Conclusion?
It seems that MP3 has a better psychoacoustic model. VQF sounds (and looks) more natural. Certainly, both formats need more development...
 
 

 


Mirror support: vorehov@paranoia.ru