In this homework we are going to practice the decoding process. Please follow the steps below:

1. Note that in this homework, we redefine the development and test sets. Specifically, each set contains five speakers: dev: CH09[0-4]* , test: CH09[5-9]* . The reference TRNs are data/CH/trn.utf8.set/trn.utf8.dev5 and trn.utf8.test5, respectively.
2. Follow Exercise-9 to decode the development set using the following setup:
• Four acoutic models (AMs): context independent, LDAd24, LDAd32, and LDAd40.
• The trigram word based LM from Homework-8.
• Three lzs: 24, 32, 40
• Configure the svmap as
  svmap configure -phonePen 0.0 -wordPen 0.0 -filPen 30 -lz $lz
• Do spass run . The script should write the hypotheses to files in the TRN format, which is
  word1 word2 word3 ...(key)
• With this setup, the script should generate 12 sets of hypotheses (4 AMs * 3 lzs).
• Use sclite to score these 12 hypotheses and report the setup with the best word error rate.
3. Decode and rescore the test set using lattices (spass.glat). Using the best setup in the previous task, do spass.glat configure -topN 100 -alphaBeam 100 followed by a spass run decoding. Then do spass.glat rescore using the following setup:
• 7 lzs: $bestLz , $bestLz +- 2, $bestLz +- 4, $bestLz +- 6, where bestLz is the best lz value on the development set in the previous step.
• 4 lps (wordPen): 0, 2, 4, 6
• 3 fps (filPen): 24, 32, 40
• Configure the svmap as
  svmap configure -phonePen 0.0 -wordPen $lp -filPen $fp -lz $lz
• Do spass.glat rescore . The script should write the hypotheses to files in the TRN format.
• With this setup, the script should generate 84 sets of hypotheses (7 lzs * 4 lps * 3fps).
• Use sclite to score these 84 hypotheses and report word error rate table(s) to show the word error rates of all the lz/lp/fp combinations.

• Please send the NFS paths of your work to Stan.

Last modified: Fri Feb 17 18:18:38 EST 2006
Maintainer: scjou@cs.cmu.edu.