Hello, myself Ashok Singh from Hyderabad, India. I am trying to create a Hindi speech recognition system. I have 20 hours of speech data. And started creating acoustic model with 8000 senones and 32 gaussian. got testing accuracy around 78%. I am thinking to change number of senones to 16000. I want to know from which stage I can further run the training the model previously build or should I train the system entirely from the begining?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello, myself Ashok Singh from Hyderabad, India. I am trying to create a Hindi speech recognition system. I have 20 hours of speech data. And started creating acoustic model with 8000 senones and 32 gaussian. got testing accuracy around 78%. I am thinking to change number of senones to 16000. I want to know from which stage I can further run the training the model previously build or should I train the system entirely from the begining?
number of gaussian will be same I mean 32, only chaning senones from 8000 to 16000.
To improve the accuracy you need to increase the size of training data to something like 500 hours, number of senones does not affect anything.
To retrain existing model after reconfiguring number of gaussians use
Last edit: Nickolay V. Shmyrev 2015-09-23
Thank you .