Re: Quality of CoreAudio SRC
Re: Quality of CoreAudio SRC
- Subject: Re: Quality of CoreAudio SRC
- From: Brian Willoughby <email@hidden>
- Date: Fri, 1 Feb 2008 00:28:41 -0800
Executive Summary: I don't think we have seen any published test
results for the CoreAudio SRC, but I really would like to.
Hi Herbie,
I've seen this White Paper a few times, and you're right about the
color assignments: the gradation seems to stop at a certain point in
cyan and then jump to a steady blue which does not change through the
remainder of the range. This color scheme acts as a visual noise
gate to hide the flaws in the BIAS product. However, if the same
color assignment is used for all products, then it is at least a fair
indication that the others they tested all have higher noise.
I'm not too shy to say that I am unimpressed with the Mac OS X
programming abilities of BIAS Inc. But I am thankful that their
tests are basically open source, by publishing the Matlab scripts,
source code, and test signal files. I plan on running the tests
against CoreAudio at my first opportunity, although I would need help
from someone with Matlab to produce the graphs.
The link that Bias cites as "matching our results" is much more
accurate in showing the flaws of the Bias SRC.
<http://src.infinitewave.ca/>
This shows clear aliasing in Bias Peak Pro 5.2, and I'm not sure how
the White Paper was able to hide this aliasing, even with their color
assignments. I like the way the Infinite Wave graphs look, better
than the prior Matlab ones.
I will disagree that we can assume Logic is using the CoreAudio
AudioConverter SRC, but I suppose someone may easily clear up this
assumption soon. My impression, having licensed Logic before Apple
bought the company, is that large parts of the code were obviously
completely independent of Apple at one point, and have only been
rewritten where absolutely necessary. I really doubt that the
internal pre-Apple Logic SRC used the same calling scheme as
AudioConverter, so I don't see why they would spend time re-factoring
it. But you've not made a bad guess. For all we know,
AudioConverter is based on the SRC obtained from the Logic buyout.
Testing of the CoreAudio SRC on its own would confirm this, and would
make a fine addition to the Infinite Wave SRC Comparison results
browser. I have written to Infinite Wave with a request to add such
results from CoreAudio, and I have offered to support them in
obtaining the measurements. We developers want to compare SRC
against commercial SRC, even if the latter is not available in our
programs. There are already several flavors of results from each
company represented, so I would like to see samples of various
AudioConverter "quality" levels, hopefully including the new
"mastering quality." Thank you for mentioning the specifics of the
AltiVec versus SSE3 implementations, Herbie, because I also requested
that the results for Intel and PowerPC be shown separately. I should
have asked them to test Logic Studio 8, because they updated their
page on Wednesday, and Logic 8 belongs on the list by now.
P.S. Look at how well iZotope 64-bit SRC performs! I think I know
what my next purchase is...
Brian Willoughby
Sound Consulting
On Jan 31, 2008, at 13:29, Herbie Robinson wrote:
Bias published a paper on SRC about a year ago that included FFT
analysis of frequency sweeps:
http://www.bias-inc.com/products/peakPro5/resampling/
peakResamplingWhitePaper.pdf
This information is 2 years old; so, it might be out of date.
Undoubtedly, the thresholds for the color assignments are chosen to
make the Bias SRC look perfectly clean, but that doesn't make the
information inaccurate. This has been widely publicized on some of
the audio engineer e-mail reflectors and nobody has ever challenged
it as being inaccurate. Unless somebody from Apple tells us
otherwise, we can probably assume Logic is using the Core Audio SRC.
Note that distortion components in that plot are getting into the
green range, or -120db. There is also aliasing around -140db. My
guess about this is that the Wavelab SRC is using single precision
floating point for the MAC loop, the Logic SRC is using the Altivec
MAC (which is a little better than single precision) and the really
clean SRCs are using double precision for everything.
This information is two years old. Also, note that if one compares
Intel vs PPC using Altivec, the result must be different, because the
Altivec floating point MAC instruction is a unique beast that does 4
multiplies and 4 adds in double precision, but truncates the result
to single precision -- Theoretically, that would be 12dB less
roundoff error than straight single precision (but much much more
than using full double precision).
Bias has published the Matlab scripts for generating the plots at:
http://www.bias-inc.com/products/peakPro5/resampling/
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Coreaudio-api mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden