At guruboolez's suggestion, I did some bitrate comparisons of the effects of the --athaa-sensitivity 1 switch. I felt that athaa-sensitivity (abbreviated A-S from now on) should remain as it was in the presets but be exposed as a valid option in betas and release versions. Guruboolez disagreed because of his experience with ABX tests on this issue. So I needed some hard data to know exactly how things change with the A-S command.
KBPS result for VBR Presets with and without --athaa-sensitivity 1
CODE
Antistar-30sec Antistar-full
V no a1 diff
9* 32 32 0
8* 32 32 0
7* 32 32 0
6 39 40 1
5 63 83 20 99 113
4 141 150 9
3 165 168 3
2 176 177 1
Reservations-30sec Reservations-full
V no a1 diff
9* 33 35 2
8* 57 62 5
7* 76 88 12
6 100 107 7
5 114 118 4 125 130
4 136 136 0
3 154 154 0
2 168 170 2
LvBQuart-30sec LvBQuart-full
V no a1 diff
9* 30 32 2
8* 50 52 2
7* 62 66 4
6 78 85 7
5 99 122 23 115 135
4 157 164 7
3 176 179 3
2 188 188 0
DvorakNo9-30sec DvorakNo9-full
V no a1 diff
9* 30 33 3
8* 50 52 2
7* 63 68 5
6 80 87 7
5 98 118 20 105 123
4 144 157 13
3 167 172 5
2 178 179 1
Reverie-30sec Reverie-full
V no a1 diff
9* 23 25 2
8* 38 38 0
7* 46 47 1
6 69 78 9
5 104 126 22 111 126
4 152 160 8
3 172 176 4
2 187 188 1
* marks the presets which automatically resample.
Antistar = Massive Attack - 100th Window - Antistar (low frequency tones)V no a1 diff
9* 32 32 0
8* 32 32 0
7* 32 32 0
6 39 40 1
5 63 83 20 99 113
4 141 150 9
3 165 168 3
2 176 177 1
Reservations-30sec Reservations-full
V no a1 diff
9* 33 35 2
8* 57 62 5
7* 76 88 12
6 100 107 7
5 114 118 4 125 130
4 136 136 0
3 154 154 0
2 168 170 2
LvBQuart-30sec LvBQuart-full
V no a1 diff
9* 30 32 2
8* 50 52 2
7* 62 66 4
6 78 85 7
5 99 122 23 115 135
4 157 164 7
3 176 179 3
2 188 188 0
DvorakNo9-30sec DvorakNo9-full
V no a1 diff
9* 30 33 3
8* 50 52 2
7* 63 68 5
6 80 87 7
5 98 118 20 105 123
4 144 157 13
3 167 172 5
2 178 179 1
Reverie-30sec Reverie-full
V no a1 diff
9* 23 25 2
8* 38 38 0
7* 46 47 1
6 69 78 9
5 104 126 22 111 126
4 152 160 8
3 172 176 4
2 187 188 1
* marks the presets which automatically resample.
Reservations = Wilco - Yankee Hotel Foxtrot - Reservations (piano with background clicks and hum)
LvBQuart = Beethoven - String Quartet in F major (op 135) - Movement 3 (strings)
DvorakNo9 = Dvorak - Symphony No 9 ("From the New World") - Movement 2 (strings and wind)
Reverie = Jacques Loussier Trio - Debussy - Rêverie (piano)
The 30 second samples taken from the most quiet sections of each track using Audacity. Other than that they were unmodified from the CD. Flac samples will be available tomorrow.
I also tested on a range of more "normal", ie louder, music. Bitrate changes were minimal across the board, never more than 5 kbps and averaging much less. Particularly reassuring was that songs with long silent or near-silent area, like for hidden track cds, had little change as well. In comparison, I think almost any classical music should be expected to gain 10 kbps or more at V5 with --athaa-sensitivity 1.
One of the things that leaps out is the more or less bell curved shape of the diff results. It is obvious that V 5 has the biggest problem. But is --athaa-sensitivity 1 the best solution? One thing I looked at was this old thread about how the A-S function is actually set in the presets. The change from being set to 32db on V0-4 to 16db at V5 shows the reason that most of the time V5 has the biggest change. (Though the expected result of --athaa-sensitivity 1 is to raise by 1db, it always moves from 32. So at V5 the effect is to raise by 17db.) But in that case, why does V4 also have a signifigant, though smaller, bitrate increase? The A-S alteration there is much smaller.
Based on these results, I have to change my opinion about athaa-sensitivity. However, I am not sure that the best solution is to set V5's A-S to 32. While the bitrate of the raised A-S in my results is more similar to "expected" V5 bitrates, looking at other bitrate results shows that these are not difficult samples. 32 may be a little too high. A possible compromise position would be to keep athaa-sensitivity as it is on V0-4, set it to 24 at V5, and remain at 16 at V6-9. This would improve V5 but avoid any "overshoot", and allow for more testing time. This small and safe change could even be done for the 3.97 final build, if the lame devs felt it was a good idea.
