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: AACEnc 2.5.9.7, SF (Read 10007 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

AACEnc 2.5.9.7, SF

Apart from regular quality improvements (tonality estimation introduced in 2.5.8.2 had been tweaked further-more,  some ATH problems were fixed, better M/S allocation, new PNS algorithm, new bit rate/distortion control additional algorithms, etc..), AACEnc v2.5.9.7 (included in Nero 6.0.0.28)  also has a brand-new algorihm for AAC encoding.

It is currently available under "fast" encoding mode, it is still very experimental and under development (however it is way better quality than old "fast" mode) - on very fast machines (P4) it won't be much faster, though ))

Anyway,  this method has completely different quantization and coding kernel and some other changes, like bit allocation / bitres control  -  currently it is not yet fully finished, but it is capable of delivering very good results.

AACEnc 2.5.9.7, SF

Reply #1 –
How does it compare to high quality mode?

AACEnc 2.5.9.7, SF

Reply #2 –
Since it is using "scale factor estimation" and compltely new rate control, it still might have issues (namely with ATH and over-estimation of the noise in low-mid freq area),  so I wouldn't compare it yet directly to HQ mode.

Anyway,  it is interesting since both modes now share the same psychoacousitc model and other AAC tools 

AACEnc 2.5.9.7, SF

Reply #3 –
Let me guess: you have a single loop algorithm instead of the dual "imbricated" loop?

AACEnc 2.5.9.7, SF

Reply #4 –
Yes,  2-pass loops have been replaced with outer-loop free quantization

AACEnc 2.5.9.7, SF

Reply #5 –
That's desesperating...
Lame's speed is improved, but it will still seem to be very slow because every other encoder will now be using a single loop algorithm...

But that's very nice for Ahead.

AACEnc 2.5.9.7, SF

Reply #6 –
Hmm - well, it is just there for "fast mode" (by the way, the versio of SFE used in 2.5.9.7  is old already, since SFE was implemented last week and it is being tweaked furthermore)

AACEnc 2.5.9.7, SF

Reply #7 –
And, yes - on SSE2 machines (like P4) the difference between SF estimation and 2-pass iteration is very very small,  SFE is only much faster (up to 30%)  if "hard to encode" clip is used where there is constant shortage of bits for each frame, in which chase the 2-pass will "choke" with that    SFE always require same amount of loops.

AACEnc 2.5.9.7, SF

Reply #8 –
Sorry for going kinda off-topic

i dond consider encoding speed a issues
however is there a way to increas its Quality efficiency by doing 3pass outer lopp (total noob qusitison as i have NO programming skill)

i woudl mind 3x the encoding speed for "10% better audio" quality
Sven Bent - Denmark

AACEnc 2.5.9.7, SF

Reply #9 –
Quote
Sorry for going kinda off-topic

i dond consider encoding speed a issues
however is there a way to increas its Quality efficiency by doing 3pass outer lopp (total noob qusitison as i have NO programming skill)

i woudl mind 3x the encoding speed for "10% better audio" quality

i agree,
quality is 999999999% more important than speed

AACEnc 2.5.9.7, SF

Reply #10 –
Quote
i agree,
quality is 999999999% more important than speed

Well, quality is 999999999% more important within these walls of HA

People elsewhere won't tolerate an encoder that encodes slower than realtime on a 2gHz+ computer. They will get annoyed and switch back to xing or something like that.

AACEnc 2.5.9.7, SF

Reply #11 –
Quote
People elsewhere won't tolerate an encoder that encodes slower than realtime on a 2gHz+ computer. They will get annoyed and switch back to xing or something like that.

Uh ... don't mention this 4-letter mp3 encoder again ... it causes allergic reaction all over my face 
The name was Plex The Ripper, not Jack The Ripper

AACEnc 2.5.9.7, SF

Reply #12 –
Ok, people - hold your horses

First of all - SF estimation doesn't have to do anything with 4-letter encoder (not LAME) :-)

Second,  it seems that, if implemented correctly, SF estimation might even get better results than 2-loop quantization - encoder that won 128 kbps listening test uses SF estimation, and not 2-loop quantization (QT: SF estimation,  FhG and Nero - 2-loop quantization)

Reason for this is low-precision of the psymodels in audio codecs and big quantization step size (1.5 dB in AAC and 3.0 dB in MP3)

AACEnc 2.5.9.7, SF

Reply #13 –
I say, 2 loop quantization module is too demanding in computational requirement..
It is much faster to just estimate the SF and G_Gain first and then do the quantization module.. I know that ISO recommend a 2 loop quantization module but it's efficiency is just marginal.. too small to justify the extra computational complexity..

AACEnc 2.5.9.7, SF

Reply #14 –
make a slow mode for people having REAL computers and/or patience :-)

I relly can complain on the speed. its fast.
but i would realy like increas qulity in streaming mode


btw why are the norminal bitrates så mcuh off ?
-streaming say 90-110  its more like 120-140
i seldom see a -streaming fiel get under 120 there are mostly around 13x
the lowest i got was 96kbits and higher was 160 with spahm and 166 with some intro to a children CD (hmm)
Sven Bent - Denmark

AACEnc 2.5.9.7, SF

Reply #15 –
Quote
People elsewhere won't tolerate an encoder that encodes slower than realtime on a 2gHz+ computer. They will get annoyed and switch back to xing or something like that.

This is precisely why I switched from LAME to QuickTime AAC - except that I don't have a 2GHz computer, but just 400MHz. It may be that LAME alt-preset standard is more reliable quality-wise than 192kbps AAC, but to me, it doesn't justify the 0.6-0.7x speeds versus QuickTime's 2.5-3.0x realtime.

AACEnc 2.5.9.7, SF

Reply #16 –
Quote
I say, 2 loop quantization module is too demanding in computational requirement..
It is much faster to just estimate the SF and G_Gain first and then do the quantization module.. I know that ISO recommend a 2 loop quantization module but it's efficiency is just marginal.. too small to justify the extra computational complexity..

Actually, due to some other things,  SF estimation might give better results than 2-loop quantizaton because the psychoacostic model itself does not have enough precision. Sounds strange, but some data collected indicates so.

Then again,  2-loop quantization can be done very efficiently - for example, in current Nero - under normal bit/sample load (around 1.4)  SFE is not at all faster than 2-loop quantization (sometimes 2-loop is faster, and I can't explain why .. yet - maybe because of SSE architecture). 

Of course, if peak bit demand is much higher than available bits, like on some hard-to-encode clips,  the SFE is much faster, since 2-loop quantizer would perform 40-50 loops.

AACEnc 2.5.9.7, SF

Reply #17 –
Quote
make a slow mode for people having REAL computers and/or patience :-)


The point is - slow mode is not always better

AACEnc 2.5.9.7, SF

Reply #18 –
Quote
Quote

make a slow mode for people having REAL computers and/or patience :-)


The point is - slow mode is not always better 

Why not do it even slower and better?

AACEnc 2.5.9.7, SF

Reply #19 –
Well, if slower was always better, we would probably be using ISO demonstration code...

AACEnc 2.5.9.7, SF

Reply #20 –
Quote
Why not do it even slower and better?


Ok, allright - I give up

I'll try to explain one more time - SF estimation is indeed faster than 2-loop quantization, at least on paper.

Now, also it might occur that this method gives better results if implemented and calibrated properly.

SFE method in Nero AACEnc 2.5.9.7 is in very early phase (I implemented it in one day)  and it is being optimized as much as current priority list allows so - which means that if, some day, it gives statistically better results on majority of samples than normal 2-loop quantization (used in normal 'high quality' mode)  it will definitely replace it -  because it would be proven to be better.

I hope I was clear enough  So, no - nobody here will sacrifice speed for quality, and I am the first person who would opose that.

We use carefully organized double-blind listening tests for every new algorithm/improvement - and if something degrades quality, it just doesn't get into new version.

AACEnc 2.5.9.7, SF

Reply #21 –
Anyway, for those interested - here are some demos of SFE + new PNS on real-world samples.

si02.aac - 100.3 kbps
41_30sec.aac - 108.1 lbps
sm03.aac - 111.7 kbps
castanets.aac - 104.6 kbps

Files are in ADIF AAC format,  encoded @44.1 kHz, Stereo, VBR

SF estimation + PNS is turned ON

AACEnc 2.5.9.7, SF

Reply #22 –
And, yes - Nero AACEnc32.dll v2.5.9.7 won't give same results since PNS bit allocation issue for SF estimation mode was fixed minutes ago  Bit rates might be 10-20% higher for same set of samples...

AACEnc 2.5.9.7, SF

Reply #23 –
And what about sleep() to improve quality?

AACEnc 2.5.9.7, SF

Reply #24 –
...sorry, couldn't resist