Re: cepstrum: computing epsilon (Paul Boersma )

Subject: Re: cepstrum: computing epsilon
From:    Paul Boersma  <boersma(at)FON.HUM.UVA.NL>
Date:    Wed, 23 Jun 1999 21:42:27 -0600

On Jun 23,  7:46pm, Laszlo Toth wrote:

> FFT (or, rather, DFT) is a formula that
> converts a series of numbers into another series of numbers. For me.
> (...) Now I see that by
> "Praat fft" you meant "the spectrum analysis facilities (and the belonging
> objects) in the Praat software package".

Exactly. That is the advantage of having a program that models the
physics (like Praat) instead of the mathematics (like Matlab).

Someone complained about the problem of choosing the "epsilon" in
a Matlab implementation of the cepstrum. In a Praat script, you
can express this epsilon in physical units. So if you want to add
white noise with a density of 10^-8 Pascal/Hertz, your cepstrum
routine would be

To Spectrum
Formula... if row=1
   ... then log10 (self[1,col]^2 + self[2,col]^2 + (1e-8)^2)
   ... else 0 fi
To Sound


(row one is the real part, row two the imaginary part;
 Matlab does have the advantage of doing complex arithmetic less verbosely)

--
Paul Boersma
Institute of Phonetic Sciences, University of Amsterdam
Herengracht 338, 1016CG Amsterdam, The Netherlands
tel:  +31-20-5252385 / 5252183 (fax: 5252197)
http://www.fon.hum.uva.nl/paul/