[Tremor] Tremor Result vs Vorbis Floatpoint decoder results

[xiphmont at xiph.org]

On Thu, Feb 3, 2011 at 1:07 AM, AMIT PANDEY <amit.kp at samsung.com> wrote:

> Thanks for your reply. I checked vector with huge difference. I found they are not as different in listening as well as in spectral testing.

Yes, large apparent differences when decoding files that made use of
floor0 is due to approximations in the LSP curve synthesis. These are
most noticeable at low frequencies where there's substantial energy.
The fact that LSP encoding is sensitive to quantization and roundoff
is part of the reason it was eventually deprecated.  Clipping can also
make the differences appear larger than they really are.

May I ask what bit depth you were using for testing?  The magnitudes
of differences you're reporting are somewhat higher than I'd expect
for some of the samples (and have just confirmed by testing here)
unless you're decoding to some bit depth other than 16bit, or perhaps
if you're using _LOW_ACCURACY_ mode.  For example, the peak magnitude
difference you report in chain-test3 is implausibly large; the signal
would also have to be exhibiting a substantial phase error in addition
to amplitude error, which could not be caused by LSP approximation.

> Actually , i want to know one thing. As i understand these test vectors [http://people.xiph.org/~xiphmont/test-vectors/vorbis/ ] are used for
> Vorbis decoder. Has anybody from tremor's developer community checked these tes tvectors ? or did they
> test some other vectors to ensure their fixed point implementation ?

These same vectors are used for both the Vorbis reference decoder
(libvorbis) as well as testing of Tremor.  Regression testing is
performed via spectrogram and physically listening to the output.

Monty