Newsgroups: comp.speech
Path: lyra.csx.cam.ac.uk!warwick!pipex!sunic!EU.net!sun4nl!news.nic.surfnet.nl!rlmix1.rulimburg.nl!BERRY.mi.rulimburg.nl!berry.debruin
From: berry.debruin@mi.rulimburg.nl (Berry de Bruyn)
Subject: Re: Tomorrows World Speech Recogniser
Message-ID: <berry.debruin.65.2DBBB2BC@mi.rulimburg.nl>
Lines: 44
Sender: news@rlmix1.rulimburg.nl (USENET News System)
Organization: University of Limburg, Maastricht
X-Newsreader: Trumpet for Windows [Version 1.0 Rev Final Beta #9]
References: <2otn0a$dcm@bambam.cs.city.ac.uk> <1994Apr19.111807.8191@leeds.ac.uk> <jimn8CoIv7D.JJ@netcom.com>
Date: Mon, 25 Apr 1994 12:08:28 GMT

jimn8@netcom.com (Jim Nitchals) writes:
>een6njb@sun.leeds.ac.uk (N J Bailey) writes:
>>I didn't see the programme, but I bet it detected wow & flutter to decide
>>whether the source was tape or voice.  So if you want to break in, use
>>a DAT Walkman!
>Detecting a regular analog tape is possible, given enough source
>material.  Just look for dropouts.

>That's too easy, though.  Let's look at ways to defeat DAT recorders.

>If the system remembers every instance of the subject making the sound,
>or passphrase, it might be able to reasonably detect a recording based
>on word durations and pitch information.

>Spectral characteristics of real voice are likely different too.
>You'd need good, flat frequency response, including speakers with
>good bass, for starters.

>You'd then need to alter word durations and pitching.  This isn't as
>easy as it sounds, although if you're a spy trying to break into the
>enemy's HQ it'd be worth the effort to do properly.

>You could also trick the target into speaking the passphrase, or
>parts of it, in a situation away from the security system.  The
>users would undoubtedly be warned to look out for attempts of this
>sort.

>>I'd be interested to hear from anybody on this subject who knows what they
>>are talking about!

>Apologies for offering mere conjecture, although I do sound and speech
>processing software as part of my job.

I haven't seen the program either.. but it may have something to do with 
background noise. A security system might check if the background noise in the 
signal-to-be-verified is the same as a pure measure of background noise. A 
prerecorded 'signal-to-be-verified' could never contain exactly identical 
background noise of the time of playback. Especially if the security system 
generated this background noise itself (in order to verify if a message is 
spoken real time at that exact location).

Just a guess

Berry
