Re: [Kaldi-users] Optimum segment length for training

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

I recently came across kaldi-trunk/egs/wsj/s5/local/run_segmentation.sh
> which was introduced in r4728 and which limits the maximum segment
> length during training to 15s.
>


>
> This seems very short to me.    The Cantab code uses segments between
> 17s and 44s.  Don't ask why these strange numbers, they are ancient and
> I'm looking to clean this up now.
>

That script was just added as an example script to show how in principle
you could segment long training audio.  It's not being actually used in the
main-line WSJ recipe, and the 15 sec is not intended as some kind of global
optimum.


In my mind the segment length should be:
>
> a) long enough to do CMS without distorting the audio.   I have no
> number for this but I assume over 10s.   Of course CMS can be pooled
> over speakers if you have speaker ID.
>

I think in this setup we were probably assuming the CMS is done per speaker.


b) be long enough to find good pauses between words.   Again, I think
> this is probably about 10s.
>
> c) be short enough to align even when you have poor initial models.
> Bigger beams can always be used and provided it's only a few minutes
> long anything can be made to align.
>
> d) be short enough to contain only one speaker.   Exactly how short this
> is depends on the nature of the audio.
>
> So I'd like to ask the question as to what everyone thinks is a good
> minimum audio length when speaker information isn't present (a) and (b)
> and what is a good maximum audio length (c) and (d)?
>
> I don't suppose anyone has run a grid search, but that would be
> interesting.
>


I don't think we ever did an experiment on this.

Dan




> ** Cantab is hiring: www.cantabResearch.com/openings **
> Dr A J Robinson, Founder, Cantab Research Ltd
> Phone direct: 01223 778240 office: 01223 794497
> Company reg no GB 05697423, VAT reg no 925606030
> 51 Canterbury Street, Cambridge, CB4 3QG, UK
>
>
> ------------------------------------------------------------------------------
> Dive into the World of Parallel Programming. The Go Parallel Website,
> sponsored by Intel and developed in partnership with Slashdot Media, is
> your
> hub for all things parallel software development, from weekly thought
> leadership blogs to news, videos, case studies, tutorials and more. Take a
> look and join the conversation now. http://goparallel.sourceforge.net/
> _______________________________________________
> Kaldi-users mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-users
>