Re: my tests of MP3 performance

DAB sounds worse than FM wrote:
Doug McDonald wrote:
Richard Evans wrote:

Oh I still however think there is little point in subtracting the
original from the encoded audio, except perhaps to teach you what to
listen for when judging the quality of the encoded files. Subtracting
the original takes no account of the masking properties that are very
important in codecs such as mp3.


Oh but it DOES tell you something,


No, it doesn't.

Basically, your post seems to be the first step towards acceptance that all you've said in the past about measures such as the mean squared error don't actually apply to perceptual audio coded material.


because you would *very quickly* see that the time domain errors signals, such as the difference signal or the mean squared error are irrelevant - it's all to do with where the errors are in the *frequency* domain that counts, NOT in the time domain.


On more observation: I went back and looked at the old tests. I compared
the errors of the older tests and the newer tests in the frequency domain,
side by side. I looked at both long sections (seconds long) and short
ones (1/50 sec at various places) and in general there was little
difference in general nature between the old encoder and the new.
They both, always, looked like noise, with the intensity falling
as the frequency rose, up to the filter cutoff. In the time domain I could
tell some difference, with the newer encoder residual looking more like
Gaussian noise and the older residual showing some patterns. So instead
of the spectrum I looked at the autocorrelation function, and there
there was a clear difference in old and new results: the newer encoder
autocorrelation looked like noise, the older one, in the short snippets,
had quite a bit of oscillation, representing the general pitch of the
chirping canaries. Clearly the algorithms in the newer LAME
encoder could tell the difference too. The new encoder is using the
noise shaped more or less Gaussian noise.


Hmmmmm .....

Exactly what am I supposed to see in the frequency domain? What do
YOU see? You have actually done the looking I presume. (These are
rhetorical questions. What I'm supposed to see is that the noise is
"shaped" so it follows the spectrum of the signal. And, indeed,
looking at MP3s of multitone tests, or a low amplitude solo flute,
that's what I see: the "psychoacoustic" algorithms have shaped the
noise so that it is "hidden" by the signal. But the signal of
normal classical music, more complex than that of low amplitude solo
wind instruments, basically has a simple-appearing nature at
all time scales except in millisecond snippets: it's essentially
filled in completely on the width scale of the noise one sees in the
MP3s of test tones. This, of course, means that all the "psychoacoustic
algorithms" can do is decide on a general plan of whether to try to make
the added noise relative louder at low or high frequencies. I do believe that
you are thinking about very, very simple music when you think that it,
optimally, does much else, which of course it does, as I well know
and discuss above.)

Doug McDonald
.

Relevant Pages

  • Re: Rotational Encoders
    ... looking signals, and you watch the phase relation between ... the two and count the number of pulses to figure out ... For the encoder, ... LEDs you can see the change in pulses really clearly. ...
    (sci.electronics.basics)
  • Re: DNA carries information
    ... signals that require some 'interpretation', ... communicating with its members. ... where the encoder was in this model. ...
    (talk.origins)
  • Re: Diffing LPCM vs. BRR Formats
    ... cancellation/nulling using two identical audio signals, ... A Google search on "lossy audio coding" turned up the following: ... The encoder attempts to ... common example of this is that higher audio frequencies can usually ...
    (rec.audio.tech)
  • Re: Diffing LPCM vs. BRR Formats
    ... when going from linear PCM to bit-rate reduced formats like AAC, ... cancellation/nulling using two identical audio signals, ... The encoder attempts to ... common example of this is that higher audio frequencies can usually ...
    (rec.audio.tech)
  • Re: quadrature to voltage converter
    ... I have a rotary encoder outputting a quadrature signal that I would like ... to convert to a voltage that indicates the encoders rotation direction ... Run the four signals into some combinatorial ...
    (sci.electronics.design)
Loading