An Empirical Exploration of Hidden Markov Models:

From Spelling Recognition to Speech Recognition

1. Before you work on the following two steps, you should have read this paper and this power point slide set regarding spelling recognition in the context of a one-dimensional keyboard and the conceptual connection between spelling recognition and speech recognition.

2. Download and play with a spelling recognition demo program (open demo.exe in the zip file):

· Note that you can modify the vocabulary set in testVocabulary.txt in the folder as you want.

· The file realMessage.txt contains words randomly selected from the vocabulary set on the fly.

· The file corruptedMessage.txt contains the resulting outputs generated by simulating a person using the circular one-dimensional keyboard to key in the words in realMessage.txt.

· The file recoveredTestMessage.txt contains, for each corrupted output on each row of corruptedMessage.txt, the top 4 most likely words from the vocabulary set.

· Compare each realMessage.txt with recoveredTestMessage.txt row by row and you’ll see the accuracy of the spelling recognition in this process.

3. Download and unzip this zip file to browse the C++ source code in a Visual C++ project folder for the spelling recognition demo program. Examine the parameter settings of the spelling model and the keyboard model in model.cpp and see three sample results 1, 2, and 3 (with explanation) when a person with certain tendency of typing try to type the message in BiolaVision.txt.