Re: [Speechsc] Question on Verification Result Device type

[Nik Waldron <nik.waldron at kaz-group.com>]

Hi Christian,

I'm not an expert on MRCPv2 but as someone working on Verification
implementation and having some familiarity with the field I think that
this is due to the time at which MRCPv2 was originally drafted.  At that
time (a part of) the state of the art was an idea known as handset
normalisation or HNorm.  HNorm set about solving the problem that
different handsets had very different log likelihood ratio scores.  There
were several variants on the solution, but basically several scores for
different utterances against the same voiceprint, or scores for the same
utterance against the different voice prints (or a combination) were used
to normalise the score.

In particular the microphone type (carbon button, or electret) were major
categories for distorting scores, and compensating for the microphone type
found to be effective.  I guess that in order for the system to operate in
an online way (updating the sets used for normalisation), it might be
necessary for the results to be categorised by microphone type (if known).
Others more familiar with the motivations for these fields may wish to
correct me.

Things have moved on a bit since then, however if you do a search (HNorm,
TNorm or CNorm) you should be able to find the relevant papers.  If the
engine you are interested in does not support / use these types of
normalisation or identify the categories then I'm sure 'unknown' will
satisfy the specification.

Best regards,




NIK WALDRON



This is an email from Fujitsu Australia Limited, ABN 19 001 011 427. It is confidential to the ordinary user of the email address to which it was addressed and may contain copyright and/or legally privileged information. No one else may read, print, store, copy or forward all or any of it or its attachments. If you receive this email in error, please return to sender. Thank you.
______________________________________________________________________
This email has been scanned by the MessageLabs Email Security System.
For more information please visit http://www.messagelabs.com/email 
______________________________________________________________________