Cepstral coeff's same from LPC and FFT?
Tony Robinson
> Are there any differences between cepstral coefficients obtained from
> LPC coefficients and those obtained from a DFT spectrum?
Yes:
a) The FFT coefficients assume that each component frequency is
independent, whilst LPC analysis assumes that they are related by
finding the maximum likelihood autoregressive model parameters.
b) The FFT coefficients are often warped onto a mel (or other
psychoacoustically justified) non-linear frequency scale.
> Is there any computational advantage to using the spectrum instead of
> linear predictors to derive the cepstrum?
No, in fact the LPC derived coefficients are computationally cheaper.
However, the mel scaled cepstum coefficients may well be more suited to
your application. For example, a large vocabulary speech recognition
system is likely to run faster using FFT based mel scaled coefficients.
The increased cost to compute the cepstra is minimal compared with the
performance increases through increased pruning of the recognition
search that the more accurate models will provide.
Tony Robinson
James Salsman
Are there any differences between cepstral coefficients obtained from
LPC coefficients and those obtained from a DFT spectrum?
Is there any computational advantage to using the spectrum instead of
linear predictors to derive the cepstrum?
:James Salsman
::Bovik Research
James Salsman
Tony Robinson <a...@eng.cam.ac.uk> wrote:
>
> However, the mel scaled cepstum coefficients may well be more suited to
> your application. For example, a large vocabulary speech recognition
> system is likely to run faster using FFT based mel scaled coefficients.
> The increased cost to compute the cepstra is minimal compared with the
> performance increases through increased pruning of the recognition
> search that the more accurate models will provide.
Thanks, Tony.
Does a signal preemphasis and a bilinear transform (to simulate a mel
scale) make up enough accuracy for the assumptions of non-independence
in the LPC model?
I would like to get a quantitative grasp of the value of doing a FFT
because the only real search in the application that I am studying
(reading literacy tutoring software) is at the phone level, and
considerably smaller than the syntax models of speech-to-text.
There are so many tradeoffs, such as the use of a smaller VQ window,
that make sense outside the speech-to-text domain. I guess there is
no better way to decide than to simulate all the options beforehand.
:James Salsman
Olivier Siohan
|>
|> Are there any differences between cepstral coefficients obtained from
|> LPC coefficients and those obtained from a DFT spectrum?
|>
Yes.
Cepstrum coefficients obtained from LPC are derived from the smooth spectrum, due
to the autoregressive smoothing.
Moreover, in such a case, the log function is an approximation, obtained by
truncation of the Taylor's development of the AR filter.
Olivier
--
==========================================================
Olivier SIOHAN Phone :(33) 83.59.20.95
CRIN - CNRS & INRIA Lorraine Fax :(33) 83.41.30.79
B.P. 239 e-mail : sio...@loria.fr
F-54506 Vandoeuvre-Les-Nancy
FRANCE
==========================================================