Skip to main content ↑ ↓

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: Fast Trellis Search for optimal ANMR (Read 2004 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Fast Trellis Search for optimal ANMR

I have a few question on trellis searches for ANMR optimal SF and HCB selection:

1) In Claus Bauer's paper The choice of Scale Factors and Huffman Code Books for MPEG-4 AAC as a direct function of the perceptual entropy of the audio signal (and longer form Joint Optimization of Scale Factors and and Huffman Code Books for MPEG-4 AAC) he writes:

Λ^R_{final} = c*ln(10)/(10M) * 10^(c(PE-R)/10M) (eq 16)

[blockquote]For a fixed value of R, and assuming that c is a positive
constant, we see from (16) that Λ^R_{final} increases with in-
creasing P E. This can be intuitively explained as follows:
For increasing P E, the minimum number of bits needed by
the encoder to encode an audio sample increases, such that
the transmission rate tends to increase and potentially vi-
olates the rate constraint. However, for an unconstrained
cost function defined as in (5), if the transmission rate is
high, the value of Λ must be large in order to strongly pe-
nalize a violation of the rate constraint.sectionApplications
and Simulation results[/blockquote]

Later in equation 17, he splits the constant c into 3 parts:

Λ^R_{final} = c1*ln(10)/(10M) * 10^((c2*PE-c3*R)/10M) (eq 17)

He goes on:

[blockquote]...we use the experimental data to
determine the constants c1 , c2 and c3 as a two-dimensional
least square problem. We obtain the values c1 = -9248.3,
c2 = 11.712, and c3 = 8.897. We see that the constants c1 -
c3 differ significantly from the value c = 6.02 in (10). This
can be expected as the relation (10) is only a very rough
approximation for non-uniform quantization. In order to
evaluate the formula (17) with the constant values chosen
as above, we compared the formula (17) with experimen-
tal values for Λ_{final} obtained by iterating over all initial Λ
values. Figure 2 shows that the actual values for Λ_{final} are
indeed very well approximated by the formula (17). Fig.[/blockquote]

However the strange fact that c1 is negative is never discussed. Does anyone have any thoughts on that?

2) Can anyone give any guidance on lambda selection for a full trellis search?

3) Both papers cite Anish Aggarwal's Ph.D. dissertation:
Aggarwal. A., Towards weighted mean-squared optimality of scalable audio coding. Dissertation, UCSB, Dec. 2002.

Any idea where I can find a copy of this?