Initializing HMM for audio classification using the Encog Java library

I have to train an HMM to classify audio data. Especially to recognize a doorbell sound. And I have to do it in Android with the Encog library.

I am working with a generic Framework that takes care of the whole flow and I just need to "throw in" the implementations for the single steps. The features are extracted with an MFCC that outputs 13 MFCC features.

I have found these Examples: HMMSimpleKMeans.java, HMMSimpleDiscrete.java

They seem helpful, but both assume a known HMM. Which is a problem, because I don't know the HMM I need.

I have read that you can initialize an HMM with creating a random one and then applying an iteration of k-means with the training data. With this rough estimation you should be able to perform baum-welch training.

Is this possible? And if yes, could you give me a hint on where to start? I feel pretty much lost in the moment and any help is appreciated.