Re: [Kaldi-users] nnet2-online i-vector sensibility with short utterances

[Daniel P. <dp...@gm...>]

The lack of length normalization is actually on purpose.  It is the
only way to make it so the system can be in principle completely
invariant to data offsets.  It also enables more robust backoff to
when you have no adaptation data at all, because it smoothly
approaches the zero ivector (due to the prior term in the iVector
estimation objective function).
I think you should just not use the iVectors at all if your utterances
are very short.  For the CTS task, you can always use previous
utterances of the same speaker in the iVector estimation.  The setup
that's checked in does that unless you decode with --per-utt.
Dan


On Thu, Jun 18, 2015 at 9:29 AM, Nagendra Goel
<nag...@go...> wrote:
> I think it would make sense. Would you like to contribute that to the
> recipe.
>
> -----Original Message-----
> From: David van Leeuwen [mailto:dav...@gm...]
> Sent: Thursday, June 18, 2015 5:18 AM
> To: kal...@li...
> Subject: [Kaldi-users] nnet2-online i-vector sensibility with short
> utterances
>
> Hello,
>
> We're using the nnet2-online setup in a CTS task.  We have a good experience
> with the same setup for a BN task.  However, for the CTS task, where
> utterances can be very short ("yes", "mmm", etc), and we observe a very
> strong dependence of the ivector length on duration (which makes sense) a
> very strong dependence of ASR performance on ivector length (which also
> makes sense).
>
> It seems that in the nnet2-online setup the ivectors are not normalized to
> length as is customary in speaker recognition.  The nnet doesn't seem to
> like the duration dependence---what would be an approach to deal with this?
> Would it make sense to train the nnet with length-normalized ivectors?
>
> Cheers,
>
> ---david
>
>
> --
> David van Leeuwen
>
> ----------------------------------------------------------------------------
> --
> _______________________________________________
> Kaldi-users mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-users
>
>
> ------------------------------------------------------------------------------
> _______________________________________________
> Kaldi-users mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-users