Re: [Kaldi-developers] align reverse

[Tony R. <to...@ca...>]

On 10/14/2014 07:00 AM, Saman Mousazadeh wrote:
> I have trained two models for alignment one mono and the other is tri. 
> now I want to use these models to align data in reverse (i.e. from the 
> end of utterance to the beginning). I have changed the L fst and the 
> mono models works pretty well but the tri model does not work at all 
> (as expected!! ). Is there any way (except new training ) to use these 
> tri model for reverse alignment?

This is an interesting theoretical problem.   There are two things to 
reverse, the WFSTs and the features.

The WFSTs either all need to be reversed (i.e. each of H C L G), or you 
need to reverse the composition.

You may also need to reverse the features, that is if you view your 
frames as t+1, t, t-1 then the first order differences will have the 
opposite sign to the normal window of frames, t-1, t, t+1.     Here by 
far the easiest is to compute all the higher order features (e.g. to 
third order differences) in the forward time order then reverse these.   
Perhaps it works to flip the sign of even differences, but I wouldn't 
trust this.

I'm finding it hard to resist the temptation to ask why you want to do this!


Tony
-- 
** 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