I'm reading a lot of topics that meantion usage of -lw 1.... and even I use it with succes.
I can't find any explanation of this option/feature!
So, its a mistery if it needs to be used in PocketsSphinx mobile apps or not..... and how/of the value 1 can be tested and improved.
So what does LW 1 mean, and what is used for?
( I can't find deeper sources than fsg_model_init and fsg->lw = lw.... and what happens afterwards ???)
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Langauge weight controls balance between language model score and acoustic score. With values like 10-20 you give more preference to langauge model in decision of the best recognition result, with 1 you give almost no preference to language model and the best path is determined simpy by acoustic model. If your model is a simple choice between words, then lw 1 makes sense, if it is something different, you can experiment with different language weights and get best one. Usually the more accurate acoustic model is, the smaller langauge weight is required.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I'm reading a lot of topics that meantion usage of -lw 1.... and even I use it with succes.
I can't find any explanation of this option/feature!
So, its a mistery if it needs to be used in PocketsSphinx mobile apps or not..... and how/of the value 1 can be tested and improved.
So what does LW 1 mean, and what is used for?
( I can't find deeper sources than fsg_model_init and fsg->lw = lw.... and what happens afterwards ???)
Langauge weight controls balance between language model score and acoustic score. With values like 10-20 you give more preference to langauge model in decision of the best recognition result, with 1 you give almost no preference to language model and the best path is determined simpy by acoustic model. If your model is a simple choice between words, then lw 1 makes sense, if it is something different, you can experiment with different language weights and get best one. Usually the more accurate acoustic model is, the smaller langauge weight is required.
Great answer, thanks!