About 30 msec on-off-on switching at 1.00 seconds, starting at about 14 kHz. A shorter one (about 10 msec) is visible at about 1.06 seconds.
About 60 msec switching at 1.14 sec, starting at about 11 kHz.
MP3 ME plugin, CRC disabled)
-h --nspsytune
Notes by ff123
In the Test for Ringing, DualIP says that the artifacts he hears are caused by the "on/off switching of higher subbands." This comment plus the fact that these artifacts are inaudible to me suggests two things: 1) they are temporal distortions, and 2) they must occur at frequencies out of my hearing range (probably > 10 kHz). The graph type most suited to viewing such artifacts would be a spectrogram, which plots frequency on the vertical axis and time on the horizontal.
Hans Heijden sent me spectrograms of a portion of eb_andul_shortXX.mp3 in which he heard ringing, which I show at the bottom of the page.
I have decided to also show the portion of the sample in which the first "s" is heard. The on/off switching at higher frequencies appear much more clearcut to me (although they are probably less audible as "ringing"). The first "s" in the plots below comes at about the 1 second mark. I have chosen to show only the left channel. I have annotated the Lame 3.87RH graph, to point out the places where I think "ringing" could be occurring.
NOTE: I do not advocate qualitative comparison by the analysis of graphs! I am only trying to correlate visible distortions with audible artifacts, with the goal of understanding what could be done to improve codec quality.
| Description | Spectrogram | Comment |
| Original | |
|
| Lame 3.87RH -h | |
About 30 millisecond on-off-on switching at 0.84 and
0.92 seconds, starting at about 10 kHz. About 30 msec on-off-on switching at 1.00 seconds, starting at about 14 kHz. A shorter one (about 10 msec) is visible at about 1.06 seconds. About 60 msec switching at 1.14 sec, starting at about 11 kHz. |
| FastEnc (Cool Edit Pro with MP3 ME plugin, CRC disabled) |
|
|
| Lame 3.88CVS, 010104, -h --nspsytune |
|
The region from 1.14 sec to 1.20 sec is more uniformly covered by Lame 3.88 -- nspsytune and by mp3enc31 when compared with Lame 3.87RH and FastEnc. |
| mp3enc31 -qual 9 | |
The region from 1.14 sec to 1.20 sec is more uniformly covered by Lame 3.88 -- nspsytune and by mp3enc31 when compared with Lame 3.87 and FastEnc. |
Here is the graph sent by Hans Heijden, along with his comments:
An example of a bad spot (for me) is towards the end, let's call the word 'prohmin'. 'oh' is halfway, and is where it rings'. Note the on/off behaviour of sample 16 (3.87RH) above about 9 kHz. Both FhG encoders look better. mp3enc [sample 45] looks best I think, but fastenc [sample 24] goes higher in freq.
