Newsgroups: comp.speech,comp.dsp
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!news.sprintlink.net!news.indirect.com!s203.phxslip2.indirect.com!user
From: emeb@indirect.com (Eric M. Brombaugh)
Subject: Re: Comparing DFTs
Message-ID: <emeb-1502952300200001@s203.phxslip2.indirect.com>
Sender: usenet@indirect.com (Internet Direct Admin)
Organization: Internet Direct, indirect.com
Date: Thu, 16 Feb 1995 06:00:20 GMT
References: <3hsgrn$9ae@mercury.dur.ac.uk>
Lines: 28
Xref: glinda.oz.cs.cmu.edu comp.speech:4660 comp.dsp:16850

In article <3hsgrn$9ae@mercury.dur.ac.uk>, "I.H Morgan"
<I.H.Morgan@durham.ac.uk> wrote:

>SNIP<

>  This returns a value of 0 iff the two DFTs, (and hence, presumably, the
> original speech signal, and the synthesised version) are the same, as is
> required. Unfortunately, apart from that, it is not very `discriminating'; the
> score doesn't very adequately indicate how `close' a synthesised version 
> sounds to the original, and the function will even give something that sounds
> ghastly a better score than something which sounds fairly reasonable.
>  So, does anyone have any ideas or references for a better way of quantatively
> comparing two DFTs ? (Or, simply, of comparing two speech signals ?)
>  I realise that the above method does not take into account the amount the
> signal changes between frames, (which I believe has quite an important affect
> on the way we perceive speech) but is there any way of assessing such change ?
>  Thanks in advance,
> Ivor Morgan

Of course, a DFT has no clues about what "sounds ghastly".  Perhaps if
you weighted the DFT bins so that frequencies within the normal speech
range were given preference over those outside you might get a more
accurate assesment of reasonable vs. ghastly.

-- 
Eric Brombaugh     KC7GXA
emeb@indirect.com (private)
ericb@sicom.com   (work)
