Musical Instrument rcecognition

I'm interested in recording a session where 2 or more different musical 
instruments (a violin, a recorder, etc) are playing simultaneously (not 
necessarily the same "notes" at the same time - but, say, in counterpoint).

I know how to display the time-frequency representation of the sampled 
audio signal, but now am interested in perhaps looking into filtering 
the separate instruments, so that I can display the T-F representation 
of the separate instrumnets, as much as is possible.

Could you please direct me to some information or techniques to enable 
to to try to filter out the various instruments?

"Ted Lechman" <eastwood132@yahoo.com> wrote in message
news:UsqdnR-Xm9LxExnYnZ2dnUVZ_vqpnZ2d@adelphia.com...
> I'm interested in recording a session where 2 or more different musical
> instruments (a violin, a recorder, etc) are playing simultaneously (not
> necessarily the same "notes" at the same time - but, say, in
counterpoint).
>
> I know how to display the time-frequency representation of the sampled
> audio signal, but now am interested in perhaps looking into filtering
> the separate instruments, so that I can display the T-F representation
> of the separate instrumnets, as much as is possible.
>
> Could you please direct me to some information or techniques to enable
> to to try to filter out the various instruments?


You'll be lucky! With a single Piano it's just about possible (I say
optimistically!) but when you add all the other instruments it's a big ask.
You get overlapping spectra for the instruments for one thing and the
instrument is never a pure tone. It's a research  topic at present.


Tam



-- 
Posted via a free Usenet account from http://www.teranews.com

Ted Lechman wrote:
> I'm interested in recording a session where 2 or more different musical 
> instruments (a violin, a recorder, etc) are playing simultaneously (not 
> necessarily the same "notes" at the same time - but, say, in counterpoint).
> 
> I know how to display the time-frequency representation of the sampled 
> audio signal, but now am interested in perhaps looking into filtering 
> the separate instruments, so that I can display the T-F representation 
> of the separate instrumnets, as much as is possible.
> 
> Could you please direct me to some information or techniques to enable 
> to to try to filter out the various instruments?

Filters separate frequencies, not instruments. You might have some 
success by using separate microphones near each instrument, and 
recording each on time-synchronized (to within 12 microseconds) separate 
tracks. Then software might be able to isolate the individual 
instruments better than the microphones alone. It's still a difficult 
task, worthy of a thesis.

Jerry
-- 
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������

Ted Lechman wrote:
> I'm interested in recording a session where 2 or more different musical
> instruments (a violin, a recorder, etc) are playing simultaneously (not
> necessarily the same "notes" at the same time - but, say, in counterpoint).
>
> I know how to display the time-frequency representation of the sampled
> audio signal, but now am interested in perhaps looking into filtering
> the separate instruments, so that I can display the T-F representation
> of the separate instrumnets, as much as is possible.
>
> Could you please direct me to some information or techniques to enable
> to to try to filter out the various instruments?

You should probably start by really studying the spectra of various
instruments and getting an idea of what (usually the relative strenght
of overtones) makes each unique.  Unfortunatley, many instruments won't
be consistent across their playing range.

To really complicate matters though, the overtones of each instrument
will be approximately harmonically related, but quite often so will the
fundamental notes being played by the various instruments.  And then
you get composers who exploit similarities in tonal spectra - using the
bassons as a second pair of horns, or ask players to do things that
radically alter the tone quality, all sort of tricks.

To sort out the puzzle, you will need to be good at DSP - but you will
probably also need to develop a knowledge of the instruments and the
ability to solve the problem yourself, by both ear and eye.

Ted Lechman wrote:

> I'm interested in recording a session where 2 or more different
> musical instruments (a violin, a recorder, etc) are playing
> simultaneously (not necessarily the same "notes" at the same
> time - but, say, in counterpoint). 
> 
> I know how to display the time-frequency representation of the
> sampled audio signal, but now am interested in perhaps looking
> into filtering the separate instruments, so that I can display
> the T-F representation of the separate instrumnets, as much as
> is possible. 
> 
> Could you please direct me to some information or techniques to
> enable to to try to filter out the various instruments?

For instance,
http://www.google.com/search?q=source+separation+dafx+filetype:pdf


Martin

-- 
Liberty: One of Imagination's most precious possessions.
--Ambrose Bierce

cs_posting@hotmail.com wrote:
> Ted Lechman wrote:
> 
>>I'm interested in recording a session where 2 or more different musical
>>instruments (a violin, a recorder, etc) are playing simultaneously (not
>>necessarily the same "notes" at the same time - but, say, in counterpoint).
>>
>>I know how to display the time-frequency representation of the sampled
>>audio signal, but now am interested in perhaps looking into filtering
>>the separate instruments, so that I can display the T-F representation
>>of the separate instrumnets, as much as is possible.
>>
>>Could you please direct me to some information or techniques to enable
>>to to try to filter out the various instruments?
> 
> 
> You should probably start by really studying the spectra of various
> instruments and getting an idea of what (usually the relative strenght
> of overtones) makes each unique.  Unfortunatley, many instruments won't
> be consistent across their playing range.
> 
> To really complicate matters though, the overtones of each instrument
> will be approximately harmonically related, but quite often so will the
> fundamental notes being played by the various instruments.  And then
> you get composers who exploit similarities in tonal spectra - using the
> bassons as a second pair of horns, or ask players to do things that
> radically alter the tone quality, all sort of tricks.
> 
> To sort out the puzzle, you will need to be good at DSP - but you will
> probably also need to develop a knowledge of the instruments and the
> ability to solve the problem yourself, by both ear and eye.
> 
It turns out that
1. I can do DSP - in either matlab or VHDL (FPGA); and have some 
experience with time-Frequency spectral analysis
2. I will also be participating in the live musical performances, such 
that I can have full control over the circumstances. I expect that the 
music samples will (initially ) three different single voiced 
instruments (rather than chordal instruments)
3. My question concerns which direction might be most fruitfull to start 
out in: matched filters, spectral analysis with constrained 
optimization, etc.

Ted Lechman skrev:
> >>Could you please direct me to some information or techniques to enable
> >>to to try to filter out the various instruments?
...
> It turns out that
> 1. I can do DSP - in either matlab or VHDL (FPGA); and have some
> experience with time-Frequency spectral analysis

Stick with matlab on this one.

> 2. I will also be participating in the live musical performances, such
> that I can have full control over the circumstances. I expect that the
> music samples will (initially ) three different single voiced
> instruments (rather than chordal instruments)

Three is too many. Start out with two instruments, I would suggest
one brass and one woodwind (?), say, a trombone and a clarinet.
First, get references of each, say, they play a scale, and some
long notes. Next,have one play a long note while the other plays
the scale. Last, have the two play various long notes in different
combinations.


> 3. My question concerns which direction might be most fruitfull to start
> out in: matched filters, spectral analysis with constrained
> optimization, etc.

If you get some basic data to start out with as above, you get
something to start out with. I would suggest spectrograms as
analysis tools.

A lot of other people have already suggested such projects
are challenging. I fully agree. The exercise above is as
simple as it gets, and I think you will have your work
cut out for you, trying to separate these two instruments.
Not theat I have deliberately suggested two instruments with
very different percieved tones. If you can't get it to work
with them, you can't get it to work with instruments that
sound similar. 

Rune

Ted Lechman wrote:
> 2. I will also be participating in the live musical performances, such
> that I can have full control over the circumstances. I expect that the
> music samples will (initially ) three different single voiced
> instruments (rather than chordal instruments)

Can you record the instruments on separate tracks ?

jwes wrote:
> Ted Lechman wrote:
> 
>>2. I will also be participating in the live musical performances, such
>>that I can have full control over the circumstances. I expect that the
>>music samples will (initially ) three different single voiced
>>instruments (rather than chordal instruments)
> 
> 
> Can you record the instruments on separate tracks ?
> 
I don;t have the capability to record separate tracks, but can record 
each instrument separately as well as together - which is not exactly 
the same thing.

I've stumbled upon something called the "Gaussian Mixture model" - does 
anyone have any experience with this and knows its range of capability 
(incapability)?

Ted Lechman wrote:

> I'm interested in recording a session where 2 or more different musical 
> instruments (a violin, a recorder, etc) are playing simultaneously (not 
> necessarily the same "notes" at the same time - but, say, in counterpoint).
> 
> I know how to display the time-frequency representation of the sampled 
> audio signal, but now am interested in perhaps looking into filtering 
> the separate instruments, so that I can display the T-F representation 
> of the separate instrumnets, as much as is possible.
> 
> Could you please direct me to some information or techniques to enable 
> to to try to filter out the various instruments?

I've just reread *YOUR* posts to this thread.

You have never said "why do it".

In general the answer will be "not possible".
However, for a particular goal, the answer could be anything from 
"impossible" thru "difficult" all the way to "trivial".
Where are you wishing to go?