VoxForge
VoxForge
Hi everyone.
Sorry for the dumb question, but how can one use 3-gram (built with LBuild, ARPA/MIT-LL format) for recognition with HVite/HDecode?
I found the paper describing HTK LVCSR system using a class-based interpolated 4-grams, but there is no recipe how to do this.
As far as I know, HVite can be supplied just with a word lattice. I cannot build it with HBuild, because it works just with bigrams.
In the absence of a word lattice, HDecode is supplied with a language model (sic!), in this case it performs a full decoding.
However, the current version supports just bigram full decoding...
I read everything from HTK book concerning language modeling and word networks, but found no cue how to incorporate developed 3-grams and 4-grams for speech recognition with HTK. I did not find anything in the HTK users mailing lists archive as well.
Thanks in advance.
--- (Edited on 11/20/2011 9:06 am [GMT-0600] by kometa_triatlon) ---
There's nothing to stop you writing code to convert a trigram into a lattice format, but unless you've got very low perplexity HVite will be too slow to be useful.
For HDecode TFM on page 267 says to use -w arpa.lm.
It is not correct to say HDecode "supports just bigram full decoding".
Tony
--
Dr Tony Robinson
Founder and owner: Cantab Research Ltd
--- (Edited on 20-November-2011 3:31 pm [GMT+0000] by TonyR) ---
Well, that is what I found in the HTK book v.3.4:
"The search space of the recognition process is defined by a model based network, produced from
expanding a supplied language model or a word level lattice using the dictionary. In the absence of
a word lattice, a language model must be supplied to perform a full decoding. The current version
of HDecode only supports bigram full decoding." (page 248, HDecode reference).
Actually I don't know how to convert n-gram into a lattice...
And what is TFM?
Best,
Dima.
--- (Edited on 11/20/2011 11:15 am [GMT-0600] by kometa_triatlon) ---
Dima,
Trust me - HDecode works with trigrams. I have it running on a large number of cores right now using trigrams.
Make sure you have HTK 3.4.1.
Alternatively, run HDecode with bigrams, generate lattices, and rescore using higher order n-grams. I do this for 5-grams for some of my work.
Tony
--
Dr Tony Robinson
Founder and Owner: Cantab Research Ltd
--- (Edited on 20-November-2011 5:53 pm [GMT+0000] by TonyR) ---
Thanks, I will try!
hmm, I did not think about rescoring, this is also an option, thanks for the advice!
--- (Edited on 11/20/2011 12:06 pm [GMT-0600] by kometa_triatlon) ---
Hi.
Is HDecode for creating network for any n-gram language model or a recognition command like HVite?
--- (Edited on 10/28/2014 4:40 am [GMT-0500] by mariya celin) ---
HDecode is like HVite, is used for estimation.
To build N-gram Language model you can check htkbook 203
--- (Edited on 3/2/2016 10:01 pm [GMT-0600] by Visitor) ---