🚀 go-pugleaf

mikael@terapin.com (Mikael Andersson) writes:

>Was ther any results on the Pitch-shifting discussion in this conference?
>Are there any shortcuts or do i have to use FFT to change it? (And still
>keep the quality?)

You'll have problems keeping the quality with any method, since pitch-shift
is an awkward, very non-linear operation which can only really be
approximated to. There are three methods that I know of, the first of which
is the most common:

 - circular buffer, with different read-in and read-out rates. Try a buffer
   of 10 to 50ms in length.
   Main problems: discontinuities when the input and output pointers cross.

 - FFT, scale frequency, IFFT.
   Main problems: discontinuities at block boundaries, computational load.

 - Signal Modelling approach. Model the signal (eg linear prediction) and
   adjust the model parameters so that when the signal is reconstructed, it
   is at the changed pitch.
   Main problems: model choice, block boundaries, computational load.

If you want to try this, I imagine you will get best results with the first
method, though the second might be worth a go. I have no personal experience
with the third, but I believe it has been used successfully on less than 
telephone-quality speech. I doubt it would be practical for music.

> 
>Mikael

Christopher Hicks
--
 ==============================================================================
 Christopher Hicks    |    If it doesn't fit...
 cmh@uk.ac.cam.eng    |                        ...you need a bigger hammer.
 ==============================================================================

cmh@eng.cam.ac.uk (C.M. Hicks) writes:

>mikael@terapin.com (Mikael Andersson) writes:

*** Stuff deleted ***

> - FFT, scale frequency, IFFT.
>   Main problems: discontinuities at block boundaries, computational load.

Windowing the input function with a Hann window, shifting only half the
FFT length, and adding the appropriate outputs of the IFFT will produce
fairly smooth output time series.  But likely not in real time.

*** Stuff deleted ***

cmh@eng.cam.ac.uk (C.M. Hicks) writes:
> - Signal Modelling approach. Model the signal (eg linear prediction) and
>   adjust the model parameters so that when the signal is reconstructed, it
>   is at the changed pitch.
>   Main problems: model choice, block boundaries, computational load.

Xavier Serra did some really excellent work on this approach when he was
at Stanford.  His thesis talk included some great examples of just the
pitchy (deterministic) and noisy components of a musical signal.  He also
played some examples that were pitch shifted.  They sounded perfect to
my ears.

Here's the complete reference from INSPEC.

							Malcolm Slaney
							Apple Perception Group
							and Stanford CCRMA

TITLE:         Spectral modeling synthesis: a sound analysis/synthesis system
                 based on a deterministic plus stochastic decomposition.
AUTHOR:        Serra, X. (Dept. of Music, Stanford Univ., CA, USA); Smith, J.,
                 III
PUBLICATION:   Computer Music Journal (Winter 1990) vol.14, no.4, p. 12-24. 21
                 refs.
DOCUMENT TYPE: Journal article
LANGUAGE:      English
ABSTRACT:  When generating musical sound on a digital computer, it is important
   to have a good model whose parameters provide a rich source of meaningful
   sound transformations. This paper addresses the synthesis technique of
   spectrum modeling. It describes a technique called spectral modeling
   synthesis (SMS) that models time-varying spectra as (1) a collection of
   sinusoids controlled through time by piecewise linear amplitude and
   frequency envelopes (the deterministic part), and (2) a time-varying
   filtered noise component (the stochastic part). The analysis procedure first
   extracts the sinusoidal trajectories by tracking peaks in a sequence of
   short-time Fourier transforms. These peaks are then removed by spectral
   subtraction. The remaining 'noise floor' is then modeled as white noise
   through a time-varying filter. A piecewise linear approximation to the upper
   spectral envelope of the noise is computed for each successive spectrum, and
   the stochastic part is synthesized by means of the overlap-add technique.
   The SMS technique has proved to give general, high quality transformations
   for a wide variety of musical signals.

mikael@terapin.com (Mikael Andersson) writes:

>Was ther any results on the Pitch-shifting discussion in this conference?
>Are there any shortcuts or do i have to use FFT to change it? (And still
>keep the quality?)

At least two programmes have come out of the discussion (one newly written, one
was previous research that was made available). Mine is real-time, and based
on a ring buffer as suggested by someone in this group. It requires a
Silicon Graphics Indigo to run (and X/Motif to compile). Ben Cox also
made a programme of his available, which works file->file rather than 
realtime (and hence is not restricted to one particular set of sound hardware).

Both sound good with small pitch-shifts, but start sounding strange with
larger ones.

I'm thinking of using some of Ben's technology to improve mine, but it's 
still on my "list of things to do".

Cheers,

Ross-c

herman@nosc.mil (John W. Herman) writes:

>cmh@eng.cam.ac.uk (C.M. Hicks) writes:

>>mikael@terapin.com (Mikael Andersson) writes:

>> - FFT, scale frequency, IFFT.
>>   Main problems: discontinuities at block boundaries, computational load.

>Windowing the input function with a Hann window, shifting only half the
>FFT length, and adding the appropriate outputs of the IFFT will produce
>fairly smooth output time series.  But likely not in real time.

Yeah, this is basically what I had in mind, with a block length of around
10 to 50 milliseconds (maybe 1024 CD samples, roughly 20ms). My fear is that
it may introduce a modulation at either one or two times the block frequency.
Real-time should be possible with a DSP with some clout.

Christopher Hicks

--
 ==============================================================================
 Christopher Hicks    |    If it doesn't fit...
 cmh@uk.ac.cam.eng    |                        ...you need a bigger hammer.
 ==============================================================================