[vorbis] Transcoding listening test

[Gian-Carlo Pascutto]

As far as I can see, transcoding could be usefull
for people who do not primarly care about quality
but about filesizes. 

One could assume that such a user would have a
collection of mp3's at 128kbps or higher bitrates,
and uses an encoder like BladeEnc or Xing. He wants
to take uses of ogg's supposed quality and transcode
his 128-or-higher files into 96 or 112kbps oggs to
save diskspace. Even if he realizes it will only
lower quality, it will gain him/her diskspace.

We know that transcoding will introduce errors, but
it is interesting to check how bad they are. If the
errors are small due to ogg's encoding quality,
they may not be all that bad on mediocre equipment,
and the diskspace is a real gain.

Therefore, I decided to do a listening test. I took
several mp3 encoded files, decoded them and reencoded
them in vorbis. I used different mp3 encoders and
encoded in beta4 @ 128 kbps. It would be reasonable
to estimate that 1.0 will give comparable quality at
96kbps, which is what such a user would use.

I intentionally used very bad listening conditions,
to weed out all the small errors and make sure that
if I could percieve anything, it would also be perceivable
on mediocre equipment to an untrained listener. I used
a cyrix150 with noisy cpu and power fans, an old SB16 ISA 
soundcard (headphones connected to speaker out, line out 
was unusable as it sounded like the sound was highpassed. 
I had to drop volume to lowest level because the speaker 
out was so loud. There was very strong hiss/noise even 
when not playing anything) I was in a room a few meters 
from a street with fast traffic and with a portable 
airconditioner running at full power 2 meters from me. 
I used open headphones so I could still hear all this.

This was about the worst listening environment I could
imagine, while it still being plausible that a user would
want to listen to music there.

I only tested 3 samples, because doing the tests was
rather slow given the equipment and hardly pleasant. 

Each time I ABX'ed the decoded mp3 versus the decoded
mp3->encoded ogg->decoded ogg. I was effectively listening
to the transcoding artifacts only.

First sample, part of remix of 'Stardust', l3enc(FhG) at
128kbps original, only drum/bass with some effects, resulting
ogg was 125kbps (very easy, mp3 is probably transparent
already)

I could not reliably hear any differences between the mp3 
and the transcoded ogg. (I could when I turned off the
airconditioning)

Second sample, part of 'Fire on Babylon' (Sinead O'Connor)
256kbps Blade original, voice and instruments, resulting
ogg was 124kbps

The ogg sounded much flatter, especially in the voice of
the singer. It was not all that annoying, but it was obvious
that quality was harmed, even in these conditions.

Third sample, part of 'Bongo bong' (Manu Chao), BladeEnc
at 160kbps, voice with instruments in background, resulting
ogg was 118kbps

The sound was horrible. There were strong distortions
in the voice of the singer which were extremely annoying.

Given these results, I would conclude transcoding will
give very annoying artifacts on many files which will even 
be hearable on bad equipment. The effects are not as bad but 
still easily hearable on files which were encoded with
higher quality. But as we are considering users which are
more concerned about diskspace than quality those will be
a lot less present than lower quality encodings, which will
sound simply horrible.

It seems reasonable to assume even a non-quality-oriented
listener will not want to transcode his files to save space
if this makes his music sound like his speakers are badly 
broken.

Since would have been about the only reasonable use of 
transcoding, I do not think transcoding should _ever_ be 
done.

(I am talking about mp3->ogg. For mp3->mp3 it could
be expected that the results would be better)


-- 
GCP

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to 'vorbis-request at xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is needed.
Unsubscribe messages sent to the list will be ignored/filtered.