f



Speech recognition - but not speech-to-text

I have a problem which should be fairly simple, but can't find anything
around on the web due to the large amounts of stuff to do with essentially
dictation systems.

The problem is basically this:
Two people speak (or possibly one person twice in a different mood, with
different background noise etc).
I want to decide (give a probability of) whether they said the same word.
I want to get a kind of 'similarity' value, without actually converting the
two inputs to text.

Can anybody give me any pointers?

Thanks,

Alun Harford


0
Alun
5/23/2005 8:58:38 PM
comp.speech.research 1026 articles. 0 followers. Post Follow

2 Replies
840 Views

Similar Articles

[PageSpeed] 31

Alun Harford wrote:

> I have a problem which should be fairly simple, but can't find anything
> around on the web due to the large amounts of stuff to do with essentially
> dictation systems.
> 
> The problem is basically this:
> Two people speak (or possibly one person twice in a different mood, with
> different background noise etc).
> I want to decide (give a probability of) whether they said the same word.
> I want to get a kind of 'similarity' value, without actually converting the
> two inputs to text.
> 
> Can anybody give me any pointers?

Search for LPC. That is one way to model what word was spoken, and the algorithm 
with source code can be found in many speech codecs.

> Thanks,
> 
> Alun Harford

-- 
Phil Frisbie, Jr.
Hawk Software
http://www.hawksoft.com

0
Phil
5/23/2005 9:41:30 PM
On 2005-05-23, Alun Harford <alunharford@yahoo.com> wrote:
> The problem is basically this:
[compare two spoken words, compute a similarity index]

Compute cepstrum (mfcc or plp) from the signals, align both results
with Dynamic Time Warping (DTW).  You'll get a useful similarity score
from that.

  OG.
0
Olivier
5/24/2005 9:20:52 AM
Reply: