naturerecordists
[Top] [All Lists]

Filter Techniques -- Revealed At Last.

Subject: Filter Techniques -- Revealed At Last.
From: Marty Michener <>
Date: Tue, 22 Jun 2004 15:38:18 -0400
At 01:20 AM 6/22/2004 -0400, Scott wrote:

>I am going to try this one more time. I went through all 13000+ messages
>in the archives and could not find anything on basic filtering. I have
>seen comments where one contributor stated that every recording he does
>goes through an FFT filter. I know that not everyone here uses CoolEdit
>but I am sure that the principles are the same. When one makes
>recordings that catch background car noises, wind buffeting, operational
>machinery, or water, what is the approach that you would use in
>filtering the recording? Are there settings that one would consider
>standard in editing any outdoor recording? If there is a good reference
>somewhere on this, I would really appreciate knowing it. I don't have a
>manual for my version of Pro 2.1 and I don't have an acoustic background.

Dear Scott:

I have written explanations of how I use Cool Edit's FFT filters over the
years (see copy of message #376 below).  The details were in a followup,
Msg. # 394 et seq., e.g. #489, #497, #867 for tips.  I have written and
offered a short computer course, with examples files, exact settings and
key-by-key instructions for free to anyone who wants to try it, just to get=

to know how to edit any natural recording to produce the best sound of a
single source from what you happened to get.  It included morphing filter
designs. I had one taker who took the sound files, spectrum examples and
course text and never got back to me, so I have no idea if interest was
lost or it was all just too detailed or incomprehensible for him.  It was
two years and two computers ago, and I would have to hunt for it, now.  It=

is about 20 megs in size, if I recall.

I suggest you not try to assimilate this post all now, but print it out,
and use it with actual field recordings as the need occurs.

First, I must say, your question I think cannot be answered as such,
because I do not know to what purpose your final recordings will be
put.  For example, Lang Elliott and Bernie Krause usually want to recreate=

sound-scapes for the listener.  In contrast, I want to give the listener
brief sound-bytes that quickly allow them to learn species identity for my=

bird-ID product, EnjoyBirds - Q.V.
http://www.enjoybirds.com/index_files/page0006.htm  So I deliberately cut
out other ambient sounds and brutally chop out quiet places to present the=

most sound for the least disk space.

Where the purpose is to surround the listener with realistic sounds, the
use of any FFT filters is usually avoided.  Any notch or band filtered out=

makes the overall recording sound harsh, muffled or hollow, depending on
the band you have eliminated. But where three birds are singing
simultaneously, and their frequency usage bands are exclusive, the FFT
filters produce three separate files, one of each species, and it is in
that mode I have some experience.

Usually there are these steps in reducing a raw recording to a usable
species example, IF THAT IS YOUR PURPOSE.  I usually record from two angled=

shotgun mics at once, so initially I decide which was pointed at the singer=

most accurately, or which had the LEAST annoying extraneous sound in the
channel, and convert that single channel to a mono file for editing.  Often=

low level recordings are also normalized to 85% of max at this time. DO
THIS FIRST, before editing. If you thought the b.g. noises were nearly
acceptable initially, this step usually makes them so loud you just feel
you need to do some editing.  This is the order:

1. Scan the whole recording for the clearest sound examples, with least
overlap from other noise sounds.
2. Using click-drag select, and the <delete> key, remove the loudest
interruptions and trim the file to begin to produce a final cut.  Save it
as a new file name at each stage.  If I have a recording of an American
Robin, the original raw cuts would be saved in a project-specific directory=

named \Pittsburgh2001 or some such, and would be named as
TURMIG00.wav,  TURMIG10.wav, TURMIG20.wav, etc.and I would save the first
one edited as TURMIG01.wav.  Say it produced three songs, with some car
noise and roosters in the b.g.
3. I remove any impulse noise BEFORE band filtering, which would spread out=

and ruin any sharp clicks or pops.
4. Set in spectral mode, I try to get a handle on the noise frequency
structure, and on that of my desired species.  This typically involves
looking more closely at the lower bands, using Alt-UPARROW to expand the
vertical spectral scale.

In this example, the robin sound never falls below 800 Hz and the car
overlaps a bit from as low as the mic recorded - say 30 - 50 Hz, up through=

1200 or more Hz.  So I know this will be a compromise.  I also know that
any FFT filter curves I make need gradual edges, NEVER sharp domain
walls.  The latter cause distress among listeners of all kinds.  In fact,
the final sound file should have the background sounds as nearly flat
across the spectrum as possible for a listener to judge it as being
acceptably natural.  So my goal, given that the car and the birds overlap
in the 600 to 1200 Hz region, is not to cut OUT the car, but to reduce it
so it is the same loudness and other background sounds average in the range=

of the robin songs as well.  Put another way, if the robin was not singing,=

all the sounds should average out at about the same loudness from bass to
treble.  If the b.g. is structured on this plan, it will minimize user
questions such as:" What is that hiss?"  or "What caused that roar?"

5. I set and save the file to a new name, e.g. TURMIG11.wav, and begin the=

FFT trials.
I have a starting filter: at the top the Log box is checked; on the middle=

right the top box says 40 db, the lower on -40 db.
The line is in the middle stretching from 0 dB to 0 dB.  This flat filter
does nothing to the sound.  It is the canvas, if you will, on which the
best filter line will soon be drawn, by repeated trial and error.  My
starting filter looks like the picture saved to my web site:
http://www.enjoybirds.com/HowItWorks/exFFT01.jpg



I drag the lowest point down to about -30 and drop it. When you put the
cursor on the line anywhere and press and hold the left button, you set a
node on the curve.  You can move it up or down, to cause gain or loss at
that point in the curve, which becomes a "knee".  Where you have loud b.g.=

noise throughout the recording, you drag the curve down, making as few new=

nodes as you need.  Typically I drag 60 Hz down part way as low as the
bottom, and 400 Hz part of that way, making a gradual curve that increases=

the bass loss as the freq. point goes down.  But this will vary with each
recording.  The rule is make the b.g. in final recording sound flat, and
only drag a point as low as it needs, avoiding vertical parts for the
curve.  Also: Do it all with one complex curve.  doing successive filter
makes it sound lousy very soon.  To delete an extra knee, drag it off the
edge of the window (that one took two weeks to figure out).

Do not be tempted to raise any part of the line above 0 dB, as the ensuing=

sounds will very likely exceed the maximum signal level for the file
without warning and the result will be permanently ruined.  Keep clicking
the [Preview] button {in the middle right} to hear the result.  It will all=

sound almost the same until you do it a few hundred times.  When you like
it, click [Ok] and look at the spectrum that results.  Almost there?  Press=

Ctrl-Z to undo the filter, then [F2] to redo it, modify the position of a
few nodes again, and try [Ok] again.  Play with the file for as long as it=

takes to fulfill your design goals (see above).  It might look like the
example I posted:
http://www.enjoybirds.com/HowItWorks/exFFT02.jpg

Now you are ready to tackle noises WITHIN the main song band.  These fall
into two categories: those in "silences" between songs, and those within a=

song.  The first is easy, the second impossible, usually, but read on.
6. If you want to shorten the interval of "silence" between songs as with
EnjoyBirds audio files -- which offends many biologists who want to
preserve the rep-rate of the original singer -- just click and drag and
[delete] the unwanted among-band sounds in your recording.  I often do NOT=

wish to do this, so replace these "silent" portions with Copy-Paste
sections from the same recording, somewhere among other "silent"
places.  This is simple:
A. mark the offending sound to find its total length (see lower right of
screen)
B. find a quieter part elsewhere and click and drag a same-length portion,=

press Crtl-C to copy it.
C. return to your offending sound, click and drag the same part of it, hit=

Ctrl-V and the quieter part will paste-replace the offending sound.  Be
sure an listen to the whole thing several times to be sure it blends ok.

7. For offending sounds AMONG your wanted sounds we could likely write
several PhD dissertations, which would really be more useful than most
actual ones (including mine).  The key is: Structure is everything.  For
files intended only for educational species-ID purposes, with a repetitive=

sound (like many wren calls), I have actually (horrors!#%*^&@#) copied a
healthy part of the rep, and pasted it over the sick part - to make the
sound come out to the proper number of reps and the proper length.  But
this is artistry not science and I am sure I will be kicked out of the AOU=

for admitting it, and that everybody within the sound of this message will=

. . . .never, never, never again buy my EnjoyBirds product as a
result!  The very idea!  Heresy!  Might as well use a Moog and be done with=
 it!

Seriously, I also use morphing filters to remove sounds that slur up or
down, fade in and out from WITHIN songs and sounds I wanted.
Morphing filters work exactly like the FFT filter you just built, and all
start with making one with the "lock to constant filter" checked.
But if you want to remove a rooster crow, it rises then falls in frequency,=

so your curve has to follow that, or take too big a chunk out of the song
you wanted to save.  So after you nearly have it perfect as a single FFT
curve, and you have run it and Ctrl-Z undone it several times, you unclick=

the dot next to "lock to constant filter" , so it reads: Morph.  Now you
have two identical filters, and you can make one higher pitched and one
lower pitched to follow that bird (rooster)!

Inexplicably, Cool Edit chooses these two filter with two boxes at the very=

top right of the window!  It took months for me to discover these boxes,
since up until now, you work from the top downward on the screen. "View
initial filter graph" and "View final".  These are two set points, and I
discovered after much error that they only work well if the have the same
number of nodes in them, so the morphing process matches set points and can=

interpolate properly.

It is important to understand exactly what this does: It follows the
"Transition curve" over the course of whatever part of the sound you have
selected, and ranges between Initial and Final curves, depending on the
position of the Transition Curve (TC).  In the simple cas that the TC
starts at the bottom left, and rises along a straight line to the top
right, the program will begin with the Init. curve at the left of your
select interval, gradually apply interpolated setting to it as it makes it=

look more like the Final Curve, ending at the right side with the curve
exactly like the Final one.  If Final and Initial are exactly alike,
morphing is just like a constant filter.  If the initial "U"- shaped dip is=

dragged a few Hz lower in pitch, and the Final a few higher, and the TC is=

made to begin low, rise fairly quickly, the fall beck to the bottom at the=

end, the filter will remove the main freq. component of the rooster's
infamous wail.  Now, all you need is to save the morph filter, e.g. as
"rooster1" and duplicate it for the second harmonic, then the thirds, etc.

Nobody said it would be easy.

Although I predict that, like several of my previous posts, not a single
person will read through all of this, let alone use any of it, I STILL send=

all my hopes for good recordings and my best regards, ;^0

Marty Michener, MIST Software Assoc. Inc.,  P. O. Box 269, Hollis, NH 03049

"They that can give up essential liberty to obtain a little temporary
safety deserve neither liberty nor safety."  - Ben Franklin, 1775
"Pues cuando ardi=F3 la p=E9rdida reverdecieron sus maizales"


Nature Recordists Msg 376:

Doug:

I have DL your dove file and can tell you that I have lots of files
that are much worse, from which I can often extract enough
of a facsimile to produce ID teaching bird song examples.
I had a White-cheeked Pintail from USVI, overlapping kids
bleating for their mother's milk (uh, goats, that is), frogs in the
same pond, and Black-necked Stilts piping to beat hell.

Let me emphasize that these files are only for helping
folks learn wild bird songs, they are not for any kind of
scientific comparisons, and many are labeled
"Warning, long silences have been shortened."
(or, as: Objects in the mirror appear . . . )

I have given up completely on COOLs "noise reduction"
filters. Just when you are getting some interesting
purification, you get all these distant bells ringing.
And the more you listen to them, the worse of an artifact
you decide they really are! I do it all now manually.
Having said that, I do use many of the tricks already
mentioned to produce a more normal-sounding final
example. My real goal is for a learner to just say:
"oh, that's what it sounds like." and not "What is wrong
with that bird?"

Usually, the first step is to remove clicks and crunches,
deleting (between selected zero crossing points)
in the time domain. Since I often walk as I record,
foot crunches are unavoidable. I might later say
something about how to hand hold shotguns while
minimizing the creaks and groans of sixty-year old
wrist joints <g>.

To keep the same pacing of a longer song, I measure
the times between each part, in milliseconds, and
when these "silences" are plagued with loud noises
(usually people, dogs or nearby common bird species)
I will copy a piece of quiet ambience of exactly the same
length, and paste-replace the noisy part. I know that there
is a great danger in doing this that you will overwrite
a subtle part of the bird's repertoire, so you've got to
be on your auditory toes to do this. But this doesn't help
when the "jaaayyyy!" overlaps with your Nightingale-thrush
song itself!

The second step is filtering, but I often use -40
dB, which usually produces garish sounding results,
as though coming through a culvert or from inside
a conch shell.

My favorite example was early near daybreak at
the Cockscomb Jaguar Sanctuary in Belize when
I had three species singing at once, all new to my library,
and they were mutually completely separable in their
frequency spectral bands. So, I finished with three
20 second digital files from the same 20 seconds of
DAT tape! That was less than 24 hours before the
DAT became useless, ostensibly from humidity.

I also make very specific (morphing) filters for
many of the common species of birds, esp.
the corvids, where I go through a recording and select
specific non-target sounds and morph filter them
into annihilation, like a frequency domain surgeon.
Watching in the spectral mode, the file will grow to
have black slices out of it where the jay or crow
previously had made the file useless for teaching.
I have literally hundreds of filters saved, named
for each species they exclude, e.g. de-LEPVER
for one that removes Leptotilla verreauxi. As we
all know, human voices are the toughest, since
our ears are exquisitely sensitive to murmurs of
other humans.

I then frequently use a file from which recognizable
specific sounds have been removed, as a source
of ambient hiss or whatever you wish to call it, and
perform a complimentary filtering on it, so it provides
the noise frequencies obviously missing from my brutally
freq-filtered original. Then I use a mix paste, as has
already been mentioned, to produce a more normal
sounding result. I will frequently use COOL
Ctrl-Z to undo, then F2 to re-try a filter or mix-paste,
trying different levels until it sounds best.

BTW I use a Bose Acoustimass system on my
computer, rather than headphones, which tend to rot
my ears in hot weather.

As we all know, the expensive part is the time spent in
Jamaica, Panama or Costa Rica, not that at the computer.

I am not at all proud to perform this analog of cosmetic
surgery, any more than a liposuction patient posts his operation
on his web site. But if it helps me or others learn the songs
better - it seems worth it.

my very best regards,

Marty Michener
MIST Software Associates, Inc.
75 Hannah Drive
Hollis, NH




  ----------


---
Outgoing mail is certified Virus Free.
Checked by AVG anti-virus system (http://www.grisoft.com).
Version: 6.0.708 / Virus Database: 464 - Release Date: 6/18/2004


[Non-text portions of this message have been removed]



________________________________________________________________________
________________________________________________________________________

<Prev in Thread] Current Thread [Next in Thread>
Admin

The University of NSW School of Computer and Engineering takes no responsibility for the contents of this archive. It is purely a compilation of material sent by many people to the naturerecordists mailing list. It has not been checked for accuracy nor its content verified in any way. If you wish to get material removed from the archive or have other queries about the archive e-mail Andrew Taylor at this address: andrewt@cse.unsw.EDU.AU