Hello

For which task do you need that language model? Language models are very dependent on the topic of the speech you are going to decode. A language model for broadcast news is different from a language model for a personal assistant.

For some topics it's very easy to create a language model, it should take you less than to ask the question. For others it's more complex.

I'd like to use a model for dictation software. The dialogue would be taken from broadcast news clips.

Sorry, broadcast news are not a dialogue. It's when a single dictor is reading the news. Did you mean something else? Can you please clarify?

It would be broadcast news with dialogue from interviews that would occur between two speakers.

I've actually found a model I think would work well, a Spanish broadcast news model, located here. However, it doesn't seem to be working well with the Voxforge acoustic model. When running the speech recognizer (using the Lattice demo) I keep receiving errors like these:

12:27:52.901 SEVERE lexTreeLinguist Bad HMM Unit: V

12:27:52.909 SEVERE HMMTree Missing HMM for unit E with lc=V rc=R

12:27:52.909 SEVERE lexTreeLinguist Bad HMM Unit: V

12:27:52.909 SEVERE HMMTree Missing HMM for unit N with lc=O rc=V

12:27:52.909 SEVERE lexTreeLinguist Bad HMM Unit: V

12:27:52.909 SEVERE HMMTree Missing HMM for unit V with lc=N rc=E

I receive hundreds of these same types of errors. Could their be a problem with my config.xml file? I've attached it for viewing.

configVox.xml

The dictionaries h4.dict and voxforge_es_sphinx.MX.dic are for Mexican Spanish acoustic model and is not compatible with Voxforge Spanish acoustic model. The phone sets are different.

Use the dictionary from Voxforge Spanish acoustic model, it is compatible. The name of the file is voxforge_es_sphinx.dic

For better assesement of the recognitoin accuracy provide the audio file you are trying to recognize, not just the configuration file. Often issue is in the audio file itself.

Please do not ask the same question twice, it's not fun to copy-paste answers.

Thanks, that seems to have fixed it!

Hmmm, though the recognizer is now working, the recognition results seem to be extremely poor, almost completely off. Could it be the audio file I'm using? It's an 8kHz signed 16-bit PCM little-endian. I've uploaded it below.

Sample.1minute.ToSeparateSpeakers_new.wav

Please provide more information. What sphinx4 version are you using? What config file are you using? What are the results you get exactly?

Your audio is pretty tight to recognize and contain music. Sphinx4 doesn't have a way to strip the music from the audio, try to remove music manually first.

For more information see the FAQ:

http://cmusphinx.sourceforge.net/wiki/faq#qwhy_my_accuracy_is_poor

I'm using the latest version of Sphinx4 with the Lattice demo. I've attached the config file as well.

config.xml