Speech Research, Decide Speaker Threshold

I want to make a "Speaker Recognition". Currently, I have finished MFCC
feature Extraction and Vector Quantization parts. I want to ask about
pattern matching.

I have tried to do pattern matching by looking for the distance of the
nearest clustered input vectors and vector in the database. The problem
is, how can I decide the speaker threshold? I found that the distance is
changed time by time, although the change is not significant. Can anyone
tell me how the pattern matching should be done?

Thank you very much

> I want to make a "Speaker Recognition". Currently, I have finished MFCC

I'm not sure what a speaker threshold is, but perhaps you would benefit
from reading these papers from the creators of currently the most
accurate speaker recognition system ("Save as..." PDF files):

http://www.speech.sri.com/cgi-bin/run-distill?papers/icassp2005-spkr-system.ps.gz
http://www.speech.sri.com/cgi-bin/run-distill?papers/icassp2005-spkr-phonelats.ps.gz
http://www.speech.sri.com/cgi-bin/run-distill?papers/icslp2004-snerfs.ps.gz
http://www.speech.sri.com/cgi-bin/run-distill?papers/odyssey2004-nerfs.ps.gz
http://www.speech.sri.com/cgi-bin/run-distill?papers/asru2003-spkrid.ps.gz
http://www.speech.sri.com/cgi-bin/run-distill?papers/eurospeech2003-spkrdur.ps.gz

Sincerely,
James Salsman
--
www.readsay.com - maker of the ReadSay PROnounce English literacy system
400 MHz PDA included: $499 -- http://www.readsay.com/PROnounce.html

[speaking of SRI]

On which task?

OG.

: is, how can I decide the speaker threshold? I found that the distance is
: changed time by time, although the change is not significant. Can anyone
: tell me how the pattern matching should be done?

A good way to tackle the variability of the absolute match score is to
normalize it. There are at least two simple options:

1. Zero mean, unit=variance normalization:
2. Normalization relative to other speaker models/
For this you can Google for 'world modeling'
'cohort modelg' 'cohort normalatn' etc.

Simple example from 2:

D_i' = D_i - max D_j

Where D_i' is the normalized score and maximum is over all different
speakers other than i.

Score normalization transforms the scores into a common range and you can
experimentally find good threshold.

Hope th helps.

Tomi