Speech:Exps 0249

From Openitware
Jump to: navigation, search

Description

Author: Colby Johnson

Date: 3/31/14

Purpose: Attempting to use the Sphinx3 decoder as it was meant to be used.

Details: Corpus: mini/train Dictionary: switchboard.dic GenTrans: genTrans8.pl


This Experiment utilizes the true functionality of the Sphinx3 decoder. With this new process running a decode will require a more simple call (called from the Base exp dir)

sudo perl scripts_pl/decode/slave.pl

This file will utilize the sphinx_decode.cfg file for parameters and LM references. It also contains the functionality for using the Npart option just like the training. After a bit of debugging everything works seamlessly. to set up for a decode simply run

/mnt/main/root/sphinx3/setup_decode.pl -task <exp#>

this is very similar to how the training is set up. The LM will be created the Same way. A few minor changes must be made to the sphinx_decode.cfg file such as the DB references (same way as the sphinx_train.cfg setup. Also you must specify a LM to use. We have learned that for large LM's we should use:

tmp.lm.DMP 

and for small LM's use:

tmp.arpa



The Npart Option worked great! decode took less than half the time using Npart = 2

Results: Results were Identical to the past exps run with the mini/train

SYSTEM SUMMARY PERCENTAGES by SPEAKER

     ,-----------------------------------------------------------------.
     |                            hyp.trans                            |
     |-----------------------------------------------------------------|
     | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err |
     |=================================================================|
     | Sum/Avg |  549  10774 | 85.1    7.0    7.9    5.0   19.9   86.0 |
     |=================================================================|
     |  Mean   |  2.9   57.0 | 86.1    6.8    7.1    9.0   22.9   88.3 |
     |  S.D.   |  1.9   44.5 | 12.3    7.6    9.3   17.5   20.1   21.5 |
     | Median  |  3.0   47.0 | 88.9    5.3    5.0    4.5   18.0  100.0 |
     `-----------------------------------------------------------------'

Successful Completion