0

Hi i am working on cmu sphinx in windows xp . i have created language model as suggested by sphinx and when i tried to give a voice sample corresponding to the language model its working fine. but when i have more text in the language model and input the corresponding voice sample as input cmu is not able to detect words and it keeps using the words from top 100 lines of the reference text from which language model is created can anyone help me how to overcome this problem.

Thanks in advance

G Charan

Andremoniy
  • 34,031
  • 20
  • 135
  • 241
charan
  • 309
  • 2
  • 4
  • 14
  • When you ask the questions you always need to provide as much information as possible. That will enable people to give you a meaningful answer. For example once you shared the files here: https://sourceforge.net/p/cmusphinx/discussion/sphinx4/thread/e24efaef/ it become clear you didn't preprocess the text properly unlike tutorial recommends that. If you share the files earlier, you could get this answer faster. – Nikolay Shmyrev Jan 28 '13 at 16:28

1 Answers1

0

AFAIR the n-gram language model file stores the amount of lines for each N-gram type in the first line before given N. So you have to modify the number yourself to fit the amount of data in your model. It should look like this:

[100]
iaeug
asedrgkjh
3eirghb
aeirf
iajedf
...
Dariusz
  • 21,561
  • 9
  • 74
  • 114
  • Hi Dariusz Wawer can u explain a bit more please i am new to cmu sphinx so i am not able to completely understan what you have said. – charan Jan 28 '13 at 10:52
  • @charan And I can not really understand what your problem is. Try to explain what you have done in more details, show some of your code, show your generated model, etc – Dariusz Jan 28 '13 at 10:56
  • in cmu sphinx its mentioned that we can create language model through a reference text please look into ARPA model training in this link http://cmusphinx.sourceforge.net/wiki/tutoriallm according to the above link i have prepared the reference text consisting of 10 paragraphs where each paragraph corresponds to a particular audio file i have created .vocab, .arpa, .idngram and lm.DMP files now when i am giving audio sample corresponding to first three paragraphs sphinx is converting the voice into proper text but when i am giving audio sample corresponding to remaining it keeps selecting. – charan Jan 28 '13 at 11:31
  • words randomly from the first three paragraphs for the remaining audio samples also i hope now you are clear about it – charan Jan 28 '13 at 11:32
  • The best thing to do is to look at the resulting lm files and check if the words from other paragraphs are in it. – Dariusz Jan 28 '13 at 11:39
  • Then the model should be OK! Perhaps you started with simple sentences (these are recognized correctly) and the following become more complex and Sphinx simply fails to recognize them. What acoustic model are you using? Are the sentences in and is english your native language? – Dariusz Jan 28 '13 at 12:23
  • Hi Dariusz i am using wsj acoustic model and my project is medical transcription where i have audio samples of a U.S doc.all the sentences are of same level of complexity. – charan Jan 28 '13 at 12:46
  • a language model is one thing - but did you add the words to the dictionary? The one with phonetic representation of words? Or is it also somehow generated automatically? – Dariusz Jan 28 '13 at 12:51
  • i have used online tool to generate the dictionary and all the words are present. – charan Jan 28 '13 at 12:55
  • Post some sphinx recognizer output in the question, some first 10 sentences. – Dariusz Jan 28 '13 at 13:13