0
$\begingroup$

I have a signal that contains the speaking of three vowels. The signal is contaminated with dc offset and an interference at 60 Hz. I want to perform LPC for the synthesis of the spoken vowels. Do I have to remove the dc offset and the interference before performing LPC?

My initial signal looks like this:

the whole single sided fft

zoomed single sided fft

There is also another question of mine about LPC here https://stackoverflow.com/questions/28909899/weird-output-of-lpc-algorithm

Improve this question
$\endgroup$

2 Answers 2

Reset to default
1
$\begingroup$

Usually before processing the speech signals, they are passed through a high pass filter with cutoff frequency of around 150 Hz to remove such artefacts.

Improve this answer
$\endgroup$
5
  • $\begingroup$ which will remove DC. $\endgroup$
    – robert bristow-johnson
    Commented Jul 31, 2016 at 17:42
  • $\begingroup$ Yes, this will ensure the unnecessary Low frequency and DC components coming from hardware/external source are suppressed/removed. $\endgroup$
    – Arpit Jain
    Commented Aug 1, 2016 at 4:37
  • $\begingroup$ DC is more than suppressed. it be gone with most HPFs having a gain of $-\infty$ dB at DC. this is more than just an artefact. turns out that because the autocorrelation $R_x[m]$ is the inverse Fourier transform of the power spectrum $S_x[k]$, which is also the square of the magnitude of the spectrum $|X[k]|^2$, removing DC before the autocorrelation insures that $R_x[m]$ is DC free. that means it must go negative for some values (and we know that $R_x[m]>0$ for some values of $m$ (like $m=0$) and gives you a solid threshold for separating peaks in the autocorrelation. $\endgroup$
    – robert bristow-johnson
    Commented Aug 1, 2016 at 4:57
  • $\begingroup$ and autocorrelation is definitely used in LPC, but the issue of separating peaks in the autocorrelation is more about pitch detection which might also find some use in LPC. $\endgroup$
    – robert bristow-johnson
    Commented Aug 1, 2016 at 5:01
  • $\begingroup$ Yes, LPC coefficient calculation does need auto-correlation. $\endgroup$
    – Arpit Jain
    Commented Aug 1, 2016 at 5:26
0
$\begingroup$

Are you sure that the 60Hz is interference? When looking at the signal, I am not so sure.

Regarding your question about weird output, it looks to me that your LPC may be too small. The LPC size should be at least twice the number of formants.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.