On Jul 29, 2011, at 5:00 AM, Alexandros Tsilfidis wrote:
In the dereverberation context, room reverberation is regarded as
the combination of early reflections and late reverberation. It is
well known that early reflections produce a spectral degradation
which is perceived as coloration, while late reverberation produces
a decaying noise-like effect.
certainly the early reflections cause coloration in the same manner
that a filter does since all that's all they are and the spacing
between the early reflections is less than the period of any low to
mid frequency content. so it's just like a basic filter, not much of
any perceived transient but a definite effect on tone. but i wouldn't
always call that effect a degradation. some rooms (that some people
have liked) have some early reflections that play an important and
desirable role in defining the subjective character of the room.
that's why, in a simple Schroeder or Gerzon style of reverb alg, you
want some pre-delay that is user definable (or at least it's coupled
to some knob or knobs the user can twist) that precedes the delay-
feedback network (whatever is in your digital verb's "secret sauce").
and in the gap in the impulse response that is your pre-delay, you
want to possibly put in some explicit FIR taps that would straight-
forwardly emulate these early responses that may characterize some
room. the reverb tail (late-term reflections) doesn't care so much
where the noise-like impulses live, so emulating those little spikes
with a delay-feedback network (if your secret sauce is tasty) may very
well suffice in lieu of a convolutional reverb which we know is much
more expensive, even if fast convolution using the FFT is used.
and if fast convolution is used, the processing delay is enough that
it can't cover the early reflections in a real-time context (for
processing files, there perhaps is never a reason for not using a
convolutional reverb unless someone just likes the sound of some
algorithm or plate or spring or something). so, even with a
convolutional reverb using fast convolution, you would need to think
of it as if preceded by a big pre-delay and perhaps some explicit FIR
reflection taps have to be put it to really copy some room.
Most dereverberation techniques usually handle separately the early
and late reverberant signal components. Cepstral processing is used
in order to remove the spectral coloration produced by the early
reflections of the room response. I think in 1975, Stockham was
probably the first one who tried to dereverberate recorded signals
through cepstral blind deconvolution. The technique is well
described in:
A V Oppenheim, R W Schafer, and J R Buck. Discrete Time Signal
Processing. Prentice Hall, 1999.
say, i don't have the 1999 rev with Buck, but i have the '89 rev and i
think this stuff is largely unchanged from what might be in that
book. it's in chapter 12 in my copy and 12.2 describes Homomorphic
Deconvolution. i dunno if this section describes what Stockham did,
but i would be interested in knowing.
However, it seems that when no prior information is available on the
recorded signal or the room response such techniques may introduce
severe processing artifacts unless you have a "white" input.
i think, what you would want to do, is come up with a time-invariant
model for the room based on analysis of the entire recorded sound
file. i *don't* think you want a blind deconvolution alg that changes
its characteristics based on the sound that is more local in time.
i have to read this 12.2 section again (it's been a decade or two),
but i can sorta see how a blind fix can be done to the late
reflections, but the overall spectrum of the signal would be affected
("colored" if you prefer) by the early reflections. i dunno how you
would separate the spectrum due to the early reflections from the
spectrum of the music or sound content. except, maybe, the
contribution to the spectrum that is from the room, that should not
change throughout the recording. maybe that's how to separate the two.
thanks for the reference, Alexandros.
--
r b-j r...@audioimagination.com
"Imagination is more important than knowledge."
--
dupswapdrop -- the music-dsp mailing list and website:
subscription info, FAQ, source code archive, list archive, book reviews, dsp
links
http://music.columbia.edu/cmc/music-dsp
http://music.columbia.edu/mailman/listinfo/music-dsp