In this assignment we will train phoneme HMMs and perform N-gram based recognition using approximate decoding strategies using a small trigram language model. The database can be downloaded from here.

Problem 1: Train Phoneme models from the "train" component of the AN4 data. The training data consist of approximately 1000 recordings of word sequences spoken by a number of people.

Recognize the test data in the "test" subdirectory. Use the dictionary and trigram language model provided in the etc/ subdirectory.

There is a file called "Assignment9.txt" included with the data. Read it carefully. It should explain what you will be required to do. If it is not clear enough, please let me know asap. There is also a file called "README" which briefly explains the layout of the data.

The most complex portion of this homework will be the implementation of the approximate decoding strategy. However, you already have most components required for it.

Do read note no. 5 in the "Assignment9.txt" file. It describes a magic term called the "language weight". The langauge weight is a multiplicative term factored in to LM probabilities. This is required for Ngram based recognition to work properly.

Due date: None. But if you complete this assignment by 8 May 2013, you can skip all of the other assignments required of you earlier in the course (if you have not completed them already).