Martin
--
Quidquid latine scriptum sit, altum viditur.
Reply by robert bristow-johnson●June 22, 20062006-06-22
in article 1150942879.139304.232360@g10g2000cwb.googlegroups.com, Ron N. at
rhnlogic@yahoo.com wrote on 06/21/2006 22:21:
> robert bristow-johnson wrote:
>>
>> not all FIR coefficient sets will have precisely the same magnitude response
>> nor necessarily have a perfectly phase-linear (constant delay) spectrum for
>> all polyphases or fractional-delays. a higher order FIR filter can be used
>> to make the sliding modulation of frequency response less in effect.
>
> A smaller sample rate change might require more "phases" or
> fractional delays if the number in the denominator is bigger.
> Thus a small difference might well be more likely to expose
> a bad tap in a big table.
that is an interesting way to think about it.
i've usually thunk that, at least for audio, that we would never need to
have more 512 phases if you could linearly interpolate between them for
fractional delays that are not a multiple of 1/512. but then you want the
phase or group delay to be as close to the constant delay for this phase,
and you want the magnitude to be a minimally varying gain.
--
r b-j rbj@audioimagination.com
"Imagination is more important than knowledge."
Reply by Ron N.●June 21, 20062006-06-21
robert bristow-johnson wrote:
> in article UaydnY_QoscMMwTZnZ2dnUVZ_oidnZ2d@giganews.com, somenoob at
> sserpy@hotmail.com wrote on 06/21/2006 16:39:
>
> > Thanks to all for your responses. While I'm trying to learn more about
> the
> > theory -- I do believe my original question can be answered
> independently of
> > my test implementation as those details appear to be
> misleading people on what
> > I'm really after. All I want to know is if
> there is any correlation between
> > the order of filter required to make
> aliasing imperceptible when resampling
> > and the magnitude of sampling rate
> change.
>
> not all FIR coefficient sets will have precisely the same magnitude response
> nor necessarily have a perfectly phase-linear (constant delay) spectrum for
> all polyphases or fractional-delays. a higher order FIR filter can be used
> to make the sliding modulation of frequency response less in effect.
A smaller sample rate change might require more "phases" or
fractional delays if the number in the denominator is bigger.
Thus a small difference might well be more likely to expose
a bad tap in a big table.
IMHO. YMMV.
--
rhn A.T nicholson d.0.t C-o-M
Reply by Martin Eisenberg●June 21, 20062006-06-21
somenoob wrote:
> Thanks to all for your responses. While I�m trying to learn
> more about the theory -- I do believe my original question can
> be answered independently of my test implementation as those
> details appear to be misleading people on what I�m really after.
> All I want to know is if there is any correlation between the
> order of filter required to make aliasing imperceptible when
> resampling and the magnitude of sampling rate change.
Please be aware that those slanted quote marks below don't display
correctly everywhere as they are beyond the ASCII range and your post
does not specify a character set.
> It sounds to me from Martin�s comment: �There is no simple
> connection to the decimation ratio� that the answer is �no�.
That's not the same thing. If you consider that the transition band
of a windowed-sinc filter comes from convolving the ideal sinc
response with the window spectrum, you see that in this case its
width is indeed independent of the corner frequency, but I'd like to
emphasize that this is a property of that particular design
procedure. I haven't done much with FIRs at all, but there is
certainly a dependence in Parks-McClellan IIR design.
Martin
--
Quidquid latine scriptum sit, altum viditur.
Reply by robert bristow-johnson●June 21, 20062006-06-21
in article UaydnY_QoscMMwTZnZ2dnUVZ_oidnZ2d@giganews.com, somenoob at
sserpy@hotmail.com wrote on 06/21/2006 16:39:
> Thanks to all for your responses. While I�m trying to learn more about
the
> theory -- I do believe my original question can be answered
independently of
> my test implementation as those details appear to be
misleading people on what
> I�m really after. All I want to know is if
there is any correlation between
> the order of filter required to make
aliasing imperceptible when resampling
> and the magnitude of sampling rate
change.
not all FIR coefficient sets will have precisely the same magnitude response
nor necessarily have a perfectly phase-linear (constant delay) spectrum for
all polyphases or fractional-delays. a higher order FIR filter can be used
to make the sliding modulation of frequency response less in effect.
--
r b-j rbj@audioimagination.com
"Imagination is more important than knowledge."
Reply by robert bristow-johnson●June 21, 20062006-06-21
in article xuqdnTxw4f0N3ATZnZ2dnUVZ_tmdnZ2d@giganews.com, somenoob at
sserpy@hotmail.com wrote on 06/21/2006 08:54:
> I'm pretty new to the world of frequency domain as proven by this
question:
>
> Is there any correlation between the quality of resampler required to
> make
aliasing "imperceptible" and the magnitude of sampling rate change?
>
> I'm currently playing with various resampling algorithms, running some
44.1kHz
> content through them (with lots of high and low frequencies) and I
noticed
> increasing the sampling rate change made it easier to hear the
differences in
> quality between the various methods. Does this imply that
as the sampling
> rate change decreases one can get away with a lower
quality algorithm, or that
> I simply got lucky with my given content and
ratios I happened to pick?
it depends on how the reconstruction polyphase filter coefficients are
computed as a function of the instantaneous time that comes out the phase
accumulator (that increments very slowly). if it's asychronous (ASRC), it
also depends on how aware the algorithm is of precisely where it is relative
to the past and upcoming input and output clock signal edges.
--
r b-j rbj@audioimagination.com
"Imagination is more important than knowledge."
Reply by somenoob●June 21, 20062006-06-21
>Decimation quality presents a tradeoff between alias rejection and
>in-band treble conservation. You can always achieve a given stopband
>rejection by placing the filter's stopband edge at the new Nyquist
>frequency, instead of its passband edge as is commonly done. (Here,
>stopband edge means the least frequency where the filter attains the
>desired rejection, and passband edge is often taken to be the -3 dB
>or -6 dB frequency.)
>
>But doing so will increase damping of frequencies that are only just
>in band. To improve this measure (or both measures in the case of
>passband edge placement) you need a narrower transition band which
>requires higher filter order. There is no simple connection to the
>decimation ratio, though a given transition width may be harder to
>achieve around some frequencies than others.
>
>Martin
>
>--
Thanks to all for your responses. While I�m trying to learn more about
the theory -- I do believe my original question can be answered
independently of my test implementation as those details appear to be
misleading people on what I�m really after. All I want to know is if
there is any correlation between the order of filter required to make
aliasing imperceptible when resampling and the magnitude of sampling rate
change.
It sounds to me from Martin�s comment: �There is no simple connection to
the
decimation ratio� that the answer is �no�.
Nyquist isn't the whole story. It takes just as long to resolve Fs - .01
a s it does to resolve .01 Hz. Few notes are held for 10 seconds.
Jerry
--
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������
Reply by Martin Eisenberg●June 21, 20062006-06-21
somenoob wrote:
> Thanks! I�m trying to figure out if I can use a less
> computationally expensive algorithm if the sample rate change is
> small. Say the original 44100Hz (containing some tones near
> 22k) stream is not oversampled and I resample to 44000Hz. Do I
> really need to use a 90+ tap windowed-sinc or is something like
> a cheap 3rd order Lagrange or even linear adequate to mitigate
> perceptible aliasing for such a small change?
Decimation quality presents a tradeoff between alias rejection and
in-band treble conservation. You can always achieve a given stopband
rejection by placing the filter's stopband edge at the new Nyquist
frequency, instead of its passband edge as is commonly done. (Here,
stopband edge means the least frequency where the filter attains the
desired rejection, and passband edge is often taken to be the -3 dB
or -6 dB frequency.)
But doing so will increase damping of frequencies that are only just
in band. To improve this measure (or both measures in the case of
passband edge placement) you need a narrower transition band which
requires higher filter order. There is no simple connection to the
decimation ratio, though a given transition width may be harder to
achieve around some frequencies than others.
Martin
--
There are some ideas so wrong that only a very
intelligent person could believe in them.
--George Orwell
Reply by Richard Owlett●June 21, 20062006-06-21
somenoob wrote:
>>Unless the original signal is oversampled, resampling to a lower rate
>>costs high-frequency response. If inadequately filtered, if will also
>>create aliasing. A word about the methods and resampling ratios you use
>>would likely get you a more complete answer.
>>
>>Jerry
>>--
>
> Thanks! I�m trying to figure out if I can use a less computationally
> expensive algorithm if the sample rate change is small. Say the original
> 44100Hz (containing some tones near 22k) stream is not oversampled and I
> resample to 44000Hz. Do I really need to use a 90+ tap windowed-sinc or
> is something like a cheap 3rd order Lagrange or even linear adequate to
> mitigate perceptible aliasing for such a small change?