[Speex-dev] Distances in codebook search

[Christopher Bader]

Björn,

I would be very interested in looking at your data and the R file.
Actually, I think a lot of people would.  You might want to post it on a
website...

CB 

-----Original Message-----
From: speex-dev-bounces at xiph.org [mailto:speex-dev-bounces at xiph.org] On
Behalf Of Björn Thalheim
Sent: Monday, September 11, 2006 4:43 AM
To: speex-dev at xiph.org
Subject: [Speex-dev] Distances in codebook search

Hello,

I recently measured the distances ndist[0][nb_subvect-1] which are used to
make the decicion, which codebook entries are to be used in
libspeex/cb_search.c (function split_cb_search_shape_sign).

For that I took a couple of audio books (all german speech), downsampled the
tracks to 8kHz and encoded them with "speexenc -n --bitrate 15000 ..." (I
used svn revision 11825).

Usually, the distances are in average 7.544420 and the deviation of the
distances is around 2.139503.

But the encoding of the data of one of these CDs shows for some tracks
average distances of around 12163.722796 and a deviation of 24312.252308
(that's just one example, there are more of them). I find this quite
interesting. More interestingly, it still sounds alright.

Is somebody of you interested in the raw audio files producing these weird
distances? I think these weird distances are a problem in the codec. Maybe
it is possible to find out where the problem is more exactly than based on
tracks (maybe frames).

Regards,

Björn


PS: if anybody of you uses R, I could send the R workspace file, too.

--
You will triumph over your enemy.

--
Important! Please recognize my new GPG Public Key!
                 Björn Thalheim
gpg fingerprint: 2F22 AAEB 1818 1548 EC78  1AE8 9D2E FCB4 0980 28CC
   download key: wget http://www.ifsr.de/~bjoern/gpg/public_key.asc
       See also: http://www.ifsr.de/~bjoern/gpg/key.html