Speech:Exps 0150

From Openitware
Jump to: navigation, search

Description

Author: David Meehan

Date: 2/6/0214

Purpose: Attempt to build a language model off the 0148 acoustic model.

Details: Build the LM and then decode.

Results The language model was built without a problem. I tried to run a decode after the LM was built but it failed. This was because I used a separate directory for my acoustic model than I did my LM. When building the LM I created symlinks to the folders in 0148. This created a problem as the decode script tried to access the file /mnt/main/Exp/0150/model_parameters/0150.cd, but because model_parameters is a symlink to 0148, the file is actually called 0148.cd. To remedy this problem, I am creating a symlink in 0148 called LM which points to the language model generated in 0150. There may be a more elegant way to remedy this problem but this appears to have worked for now. The results of the decode are as follows:

                     SYSTEM SUMMARY PERCENTAGES by SPEAKER                      

      ,-----------------------------------------------------------------.
      |                            hyp.trans                            |
      |-----------------------------------------------------------------|
      | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err |
      |=================================================================|
      | Sum/Avg | 3099  47946 | 69.1   23.6    7.4   16.8   47.7   96.7 |
      |=================================================================|
      |  Mean   | 53.4  826.7 | 69.3   23.7    7.0   18.4   49.1   97.2 |
      |  S.D.   | 17.9  292.8 |  7.9    6.2    2.7    9.4   12.8    4.8 |
      | Median  | 51.5  789.0 | 69.6   23.2    6.6   17.4   47.0  100.0 |
      `-----------------------------------------------------------------'

Successful Completion