I am working for my Master thesis on Speech Recognition. I used sphinx4 to train a model to recognise the 10 digits in the Medumba Language, a local language in the West Region of the Country
But I have a problem in the HMM implementation of Sphinxtrain. And don't understand very well. How does Sphinxtrain initialize the first parameter for training?
As we know the training of HMM is based on the EM implementation for the BaumWelch algorithm, and the second step of the process is to choose the initialise parameters. Please can you help me to understand how does it work in Sphinxtrain?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
There are several types of models - continuous, semi-continuous and ptm. Continuous uses simple "flat start" procedure which calculates overall mean and variance of the data and then initializes parameters with that uniform value.
semi-continuous and ptm models have similar procedure but instead of single calculated mean they apply kmeans algorithm to select M gaussians (usually 256 or 128) which will be used in mixtures.
You can check the HTKbook for more detailed description of flat start.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Morning all,
I am working for my Master thesis on Speech Recognition. I used sphinx4 to train a model to recognise the 10 digits in the Medumba Language, a local language in the West Region of the Country
But I have a problem in the HMM implementation of Sphinxtrain. And don't understand very well. How does Sphinxtrain initialize the first parameter for training?
As we know the training of HMM is based on the EM implementation for the BaumWelch algorithm, and the second step of the process is to choose the initialise parameters. Please can you help me to understand how does it work in Sphinxtrain?
There are several types of models - continuous, semi-continuous and ptm. Continuous uses simple "flat start" procedure which calculates overall mean and variance of the data and then initializes parameters with that uniform value.
semi-continuous and ptm models have similar procedure but instead of single calculated mean they apply kmeans algorithm to select M gaussians (usually 256 or 128) which will be used in mixtures.
You can check the HTKbook for more detailed description of flat start.
Ok Well
Thanks. I trained a continuous model for better accurary as mentionned in the AM training Tutorial.
So I will check for "flat start" to understand how it work