i built up my own acoustic model and dic etc.in the end of training process result for my test wave files the accuracy is up to 96 to 97%.but now in live speech recognition on mobile app is not accurate.for some words are recognized perfectly while some are confused for other words.i also noticed it if i put less number of words say 10 to 15 words in digits.gram file then recognition really improves even it recognize each word correctly.can some body tell me what is the real problem.i use semi continuous model in training process.and sampling rate is 16000.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
i think there is some problem with my acoustic model.as my acoustic model is semi context dependent model with 200 senones. where as the model en-us-ptm used in demo android pocketsphinx project is context independent and ptm model.what you suggest i should also make context independent ptm acoustic model?the data is given in the attachments
200 senones is too small for anything serious. Our models are built from 100 hours of data and use 4000 senones. Numbers on required data size are provided in acoustic model training tutorial.
To recognize correctly from the start you need to add the following line:
-cmninit 40,3,-1
to the feat.params file of your model.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
i built up my own acoustic model and dic etc.in the end of training process result for my test wave files the accuracy is up to 96 to 97%.but now in live speech recognition on mobile app is not accurate.for some words are recognized perfectly while some are confused for other words.i also noticed it if i put less number of words say 10 to 15 words in digits.gram file then recognition really improves even it recognize each word correctly.can some body tell me what is the real problem.i use semi continuous model in training process.and sampling rate is 16000.
You need to provide the data to get help on the accuracy.
i think there is some problem with my acoustic model.as my acoustic model is semi context dependent model with 200 senones. where as the model en-us-ptm used in demo android pocketsphinx project is context independent and ptm model.what you suggest i should also make context independent ptm acoustic model?the data is given in the attachments
200 senones is too small for anything serious. Our models are built from 100 hours of data and use 4000 senones. Numbers on required data size are provided in acoustic model training tutorial.
To recognize correctly from the start you need to add the following line:
to the feat.params file of your model.