In this homework you are going to train two Gaussian mixture models (GMMs) for speech-silence classification. Please follow the steps below:

1. Set up the feature description file (featDesc)
   • Apply what you have learned in Exercise-3 to set up a feature description file, featDesc, so that in the tcl script you can do something like: fs setDesc @some_path/featDesc, where @ means the feature description is contained in a seperate file.
   • Apply the FeatureSet methods in Exercise-3 to generate the features: MCEP, DELTA, and FEAT. In the featDesc file, use comments to describe the meaning of these three features in one sentence each.
   • Also generate and describe the POWER4 feature. Find out the usage of the FeatureSet method thresh, which can be applied on POWER4 to generate a new binary feature SPEECH as a speech-silence indicator. Note that the idea is to set SPEECH to 1 if POWER4 >= 0 and to 0 if POWER4 < 0, in the frame level.
   • Test your feature description file against some utterance in the database you created in Homework-2.
2. GMM Training
   • Firstly, follow Exercise-3 to create two GMMs for speech and silence, named SPEECH and SIL, respectively. Each GMM should contain one Gaussian with diagonal covariance.
   • As described in Task 5 of Exercise-2, use all the utterances of speakers "0[36]*" as the training set. Iterate over these utterances to collect speech and silence feature statistics. Please make use of the database created in Homework-2 to get utterance information.
   • For each training utterance, evaluate the FeatureSet fs and accumulate the feature statistics to the GMMs SPEECH or SIL, depending on the value of the binary feature SPEECH (if feature SPEECH == 1, accumulate to GMM SPEECH; if 0, to GMM SIL. Note that don't be confused by the GMM SPEECH vs. the feature SPEECH.
   • After iterate over all the training utterances, update the GMMs and store the GMM descriptions and parameters, as you have learned in Exercise-3.

• This homework should be managed to done in one tcl script. Please send the NFS path of the script to Stan.
• Note that the behavior of the featDesc file might be confusing. If so, feel free to ask Stan for help.

Last modified: Wed Jan 4 01:32:27 EST 2006
Maintainer: scjou@cs.cmu.edu.