Acoustic Model Discussions

Flat
Multiword 'words'
User: TwoNotes
Date: 1/5/2015 3:55 pm
Views: 5698
Rating: 1

I need to recognize a bunch of proper names.  Luckily, most of the individual names (first and last) are already in the voxforge lexicon for Julius.  For example, "Doris" is "d ow r ih s" and "Day" is "d ey".  But if I try to create a compound word DorisDay, concatenating the phonemes for "Doris" and "Day" to be "d ow r ih s d ey", I get an error from Julius at startup, complaining that there are no triphones to bridge the gaps such as "ih s d".  I guess that sequence never appears within a 'word' in English.

I tried putting a 'sil' between the 's' and 'd' but that did not help either.

Is there a way to indicate a break in the triphone sequences to allow combining words in this way?

The trouble I run into is that my toal vocabulary is over 400 full names, and if I put all the individual name words into one class, and say that a "name" is one or more members of this class, than it gets a lot of false hits.

For example, "The Andrews Sisters" are in the list.  "The" and "Day" are close enough that Julius sometimes reports I said "Doris The" instead of "Doris Day".  Obviously, Julius has no knowledge of what is a valid name.  I am looking for ways to increase the odds of a correct recognition without entering a grammar rule for each of the 400+ names.

 

--- (Edited on 1/5/2015 3:55 pm [GMT-0600] by TwoNotes) ---

Re: Multiword 'words'
User: colbec
Date: 1/7/2015 8:18 am
Views: 2730
Rating: 1

There are a couple of ways of looking at this. Either you consider a specific language model where the probability of Day after Doris is very high and taken into account by Julius, or provide more relevant audio examples. Consider if you are looking for  ? s + d ? then if you collect a bunch of example words such as misdeed, misdemeanour, disdain and add these to your samples with appropriate phonemes then you are thoroughly practicing your triphones.

Sometimes you can make a sample word do double duty and exercise more than one triphone.

If you think that the lexicon already has a triphone that should fit then examine carefully the phoneme representation and see if you are attempting to establish a different pattern.

Do a Google search on "HTK triphone missing" and you should find a number of hits with suggestions what to do.

--- (Edited on 2015-01-07 9:18 am [GMT-0500] by colbec) ---

PreviousNext