I am working on a library for generating LPC for speech synthesis. I am currently using a hamming window for the spectral analysis which goes in 200ms blocks over the signal, and does the a-to-k conversion.

I have read some stuff online about a technique of overlapping these windows so it would process the signal in groups like [1-200, 150-350, 300-500, etc].

My question is: How exactly is this done, and more important-- is there a benefit to overlapping the windows? Will I get better results with my analysis?

  • $\begingroup$ What you are talking about here is called Welch's Method of spectral estimation. If I remember correctly, 50% overlap is the most you would ever want to overlap (1-200, 100-300...), beyond 50% there is no additional benefit. $\endgroup$ – CMDoolittle Sep 10 '15 at 20:02
  • $\begingroup$ can you explain where the benefit comes from? How does it work exactly, and I'm a little tripped up by the fact that frames 1-200 are for one frame of LPC data (ks, rms, pitch), so if I am generating my a-to-k based off of 100-300, how does that fit into one frame? Am I taking an average? $\endgroup$ – patrick Sep 11 '15 at 8:39
  • $\begingroup$ I don't know anything about LPC, just spectral estimation. $\endgroup$ – CMDoolittle Sep 11 '15 at 16:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.