Subject: Re: [linux-audio-user] Pitch shifting
From: Steve Harris (S.W.Harris_AT_ecs.soton.ac.uk)
Date: Mon Jan 28 2002 - 17:52:08 EET
On Mon, Jan 28, 2002 at 04:17:15 +0100, Guenther Sohler wrote:
> Whats the exact mathematical background of good pitch shifting ?
The hard part is maintaining the ratio between the partials.
> * the minimal frequency converteted depends on the number of samples
Not really, only the granularity. The 0th bin of an FFT is a
representation of all the components from zero to nyquist/no_of_bins.
> * I dont know whether the result would be good if you took too much
> samples(65536)
It improves the quality slightly, at the cost of CPU and latency. 4096
seems like a good tradeoff. You have to zero pad as well of course, as the
input time domain signal is not circular.
> * If you dont do a phase alignment, the result sounds very cracky (not
> monotonically)
Phase analysis is vital. It is also used to scale down the upper partials
to the correct ratios.
> so whats a good way to do pitch scaling ?
I think that the current best hardware uses LPC (Linear Predictice Coding).
I haven't tried this technique though, it is a bit of a specialisation.
- Steve
This archive was generated by hypermail 2b28 : Mon Jan 28 2002 - 17:43:12 EET