| At 07:47 28/02/97 -0500, you wrote:
>> Can anyone point me in the direction of technical books or papers on
>> the subject of morphing audio signals?
>
>
>Not to reopen this can of worms, but I have seen very little serious
>litterature on audio morphing. There was considerable debate a while
>back about this, on this list and elsewhere. One can can effectively
>argue that morphing two audio signals is very similar to simply mixing
>them. Interpolating two fft analysis files on a bin-to-bin basis over a
>period of time would, in theory, be close to the well-known analogous
>process in computer graphics/animation, but in my experience, the
>results of this process in audio is almost identical to conventional
>linear mixing. Which is to say that other more complex operations
>(stretching, compressing, transposing etc.) on a pair of
>frequency-domain data blocks have much more interesting sonic potential.
>The term "morphing" is the problem here.
I use the Composer's Desktop Project software which includes tools that do
this, and in my experience the idea that you can simply interpolate between
FFTs and get a morph to sound convincing is naive. And I disagree that a
morph is the same as a crossfade. A true morph would create a half-way
stage that is neither one sound or the other but includes characteristics
of both. A cross-fade doesn't do this.
The problem with the FFT is that it's looking at this problem the wrong
way. When you identify a sound you don't hear it as an FFT, but as a
combination of pitch, amplitude envelope, short-term timbral detail, and
medium-term timbral envelope. To create an acceptable morph you have to be
able to isolate these factors and control them individually. This is
moderately easy for monophonic sounds but spectacularly difficult for
polyphonic sounds.
In short, to create a convincing morph between sounds that are dissimilar
you have to model the sounds and interpolate the coefficients of the model
from one sound to the other. Which model you use - LPC, physical modelling,
some variant of the FFT that adds extra information about pitch and timbral
envelope, or whatever - doesn't matter as much as the ability to describe
the sounds convincingly at each end of the morph and change the model's
coefficients smoothly.
If you do this you can get some very interesting results, particularly with
short morphs that mutate a single monophonic sound quite quickly. But it's
a long and tedious process, and requires a lot of trial and error before
you find something that works well. For some sounds, especially those with
a very variable amplitude envelope, you can get very good results by simply
copying the envelope variation from one sound to another and doing a
crossfade 'under' the new envelope. For other sounds the process is a lot
more complicated...
R. |