Click here to register.

Speech Recognition Engines

Flat
HTK LM for use in Julius
User: Visitor
Date: 5/21/2016 4:22 am
Views: 2764
Rating: 0

I have gone through the tutorial in the HTK Book chapter 15 and have created a couple of language models. The one I am using is a 5k trigram model created using

LBuild -T 1 -c 3 1 -n 3 -l lm_5k.2gram lm_5k/5k.wmap lm_5k.3gram holmes.1/data.* lm_5k/data.*

I then converted this into a Julius binary file using mkbingram. No errors.

I am now trying to get them working with Julius. Here's where I am up to annd having problems:

I downloaded the latest nightly acoustic model from VoxForge and set my test.jconf to use:

-d lm/lm_5k.bingram 
-v acoustic_model_files/dict
-h acoustic_model_files/hmmdefs
-hlist acoustic_model_files/tiedlist

I run Julius and get the following errors:

...

Error: ngram_lookup: "ZOOELOGY" not exist in N-gram

Error: ngram_lookup: "ZUE" not exist in N-gram

Error: ngram_lookup: "ZURICH" not exist in N-gram

Error: --- Failed to map 10425 words in dictionary to N-gram

Error: --- Specify the word to which those words are mapped with "-mapunk" (default: "<unk>" or "<UNK>"

ERROR: m_fusion: failed to init

I feel as though I am so close to acheiving an actual positive result here but can't figure this one out on my own. Please help :)

--- (Edited on 5/21/2016 4:22 am [GMT-0500] by Visitor) ---

Re: HTK LM for use in Julius
User: wildcard
Date: 5/21/2016 4:24 am
Views: 38
Rating: 0

[This was my post, just forgot to log in]

--- (Edited on 5/21/2016 4:24 am [GMT-0500] by wildcard) ---

Re: HTK LM for use in Julius
User: colbec
Date: 5/22/2016 5:28 am
Views: 26
Rating: 0

Ok so you have found out something about LMs.

Julius is having a problem matching your audio to the language model.

In an earlier thread you were offered some advice about going back to something that works and using that as a launching point, a suggestion which it seems has yet to sink in. The point is that if you pull up a successful working grammar that Julius is happy with, then you know that the audio and language sides are balanced. How would you turn a grammar into a LM?

--- (Edited on 2016-05-22 6:28 am [GMT-0400] by colbec) ---

Re: HTK LM for use in Julius
User: wildcard
Date: 5/22/2016 6:16 am
Views: 25
Rating: 0

Right, OK. I'm assuming that the advise to go back to something that worked was that I should go back through the tutorial on VoxForge?

If not, that is the only thing I have got working. 

The tutorial created a grammar file and then the acoustic model. You think I now need to create a new acoustic model? If so should I use the text from the prompts files I used to create the language model?

I know these questions may sound like I'm being an idiot but I am finding it hard finding resources that I can understand from an entry level. All I have at the moment is to ask stupid questions and hope you don't lose patience! So I am very appreciative of you help.

--- (Edited on 5/22/2016 6:16 am [GMT-0500] by wildcard) ---

Re: HTK LM for use in Julius
User: colbec
Date: 5/22/2016 7:08 am
Views: 1031
Rating: 0

"should I use the text from the prompts files I used to create the language model?" Aha - obviously the LM will only be a toy thing, but manageable from a learning perspective. When you get error messages they will refer to something tangible, after all the audio matches the grammar and you have everything right there in small files. Above all avoid the temptation to stick in extra words or a huge dictionary until you have the known set of words working.

--- (Edited on 2016-05-22 8:08 am [GMT-0400] by colbec) ---

PreviousNext