Notes by ff123
Up to now, I have taken it for granted that the auto-attenuation feature within the MAD decoder works as advertised to reduce the audible effects of clipping distortion. I decided to test it for myself.
I used Lame 3.88 beta at 128 kbit/s with the -h switch to encode to mp3. I compared the FhG decoder within Winamp 2.73 (which is fully ISO compliant and does not contain any bugs which affect decoding sound quality) with MAD 0.13.0b. I enabled the auto-attenuation feature with a sensitivity setting of 5 out of 7 (1 = least sensitivity, 7 = most sensitivity).
I found a sample in which I think I may be able to hear MAD improving the clipping distortion. It's a 7 to 8 second clip of Emma Bunton's "Free Up Your Mind" from the Pokemon Movie Soundtrack. There is a loud drum at the 3 second mark which is clearly distorted using both headphones I tried. When I don't enable auto-attenuation, MAD reports peak clipping of 2.0 dB in the sample.
Using my Grado SR325's, I believe I could detect the volume decreasing as the sample progressed in the MAD decode, but I could also hear worse distortion (I think) with the FhG decode at the loud drum. In any case, I was able to distinguish MAD from FhG 15 of 16 times in an ABX run.
Using my Beyerdynamic DT 990 Pro's, the change in volume as the sample progressed was harder to detect, but focusing in on only the drum, I could hear a difference -- an improvement in quality using the MAD decoder. I think the difference was in distortion, but there is a possibility that all I heard was a difference in volume which I interpreted as a difference in audible distortion. With two variables in play (volume and distortion), a listening test of this sort is somewhat problematic. Also, differences in distortion are fairly difficult for me to distinguish. Moreover, I can't definitively rule out the possibility that the headphones themselves are breaking up with the louder sample. I consider this last possibility to be remote, however, because 1) I didn't listen at a loud volume and 2) both headphones revealed a difference.
Using the Beyerdynamics, I scored 15 correct identifications of X in 16 trials.
The following show the time-domain and spectral views of the sample. In the time-domain view, one can see the reduction in amplitude and in the spectral view, one can see that the vertical lines which extend all the way up in frequency (the result of clipping harmonic products), disappear for the most part when MAD with auto-attenation is enabled. Click on the thumbnail images to enlarge.
| Decoder | Time-domain View | Spectral View |
| FhG decoder within Winamp 2.73 | |
|
| MAD decoder 0.13.0b | |
|
Conclusion: The MAD 0.13.0b decoder produced subjectively better sound to me than the FhG decoder within Winamp 2.73 on a sample which had clipping distortion.
Original wav file and decoded wavs were uploaded to my Audio Samples Page. The files have been losslessly compressed using the FLAC format as freeup.flac, freeup_mad.flac, and freeup_winamp.flac