Speech:Train Archive All


 * Home
 * Semesters - Project Work by Semester
 * [Information]
 * System Description
 * Experiments - List of speech experiments

Project Notes

 * Unix Notes
 * Speech Corpus Setup - Switchboard,  NOAA
 * Speech Recognition Related Readings
 * Experiment Setup
 * Scripts Page
 * Model Building - more info on data prep,  language models, &  building models
 * Step 1: Run a Train
 * Archive
 * Wizard Style
 * [All In One]
 * First Edition
 * Step 2: Create the Language Model
 * Step 3: Run a Decode

All In One
Creating an Experiment and Running a Train


 * Step 1: Create a new exp directory inside /mnt/main/Exp/0252/. Example: mkdir 023


 * Step 2: cd 

Example: /mnt/main/scripts/user/prepareExperiment2.pl first_5hr/train
 * Step 3: run /mnt/main/scripts/user/prepareExperiment2.pl


 * Step 4: cd /etc.

Change: Density on line #107 and Senone on line #120. (Or any other parameter in the file) To edit a file press i on the keyboard for INSERT. After you're done editing press Esc on the keyboard. Then Save changes by pressing ZZ (uppercase) on keyboard.
 * Step 5: vi sphinx_train.cfg

Change line #43 from test to train Change line #44 from test to train Change line #51 from 3 digit exp number to tmp. Example: $DEC_CFG_LANGUAGEMODEL = "$DEC_CFG_LANGUAGEMODEL_DIR/tmp.lm.DMP"; Save changes
 * Step 6: vi sphinx_decode.cfg


 * Step 7: Back to base experiment directory. cd ..

If an error pops during the process run the script again.
 * Step 8: perl /mnt/main/scripts/user/generateFeats.pl


 * Step 9: nohup scripts_pl/RunAll.pl . &


 * Step 10: Wait few hours for the train to finish. You can check if the process is running by using top -u

Creating a language model

 * Step 1: Inside your Exp base directory create a LM folder: mkdir LM


 * Step 2: cd LM

Example: cp -i /mnt/main/corpus/switchboard/first_5hr/train/trans/train.trans trans_unedited
 * Step 3: cp -i /mnt/main/corpus/switchboard/ /trans/train.trans trans_unedited
 * Step 4: /mnt/main/corpus/dist/Switchboard/transcripts/ICSI_Transcriptions/trans/icsi/ParseTranscript.perl trans_unedited trans_parsed


 * Step 5: cp -i /mnt/main/scripts/user/lm_create.pl.


 * Step 6: ./lm_create.pl trans_parsed

Start Decoding

 * Step 1: Get back to your experiments base directory. If your inside LM directory then just do cd ..


 * Step 2: nohup scripts_pl/decode/slave.pl &


 * Step 3: Wait several hours for the decode to finish You can check if the process is running by using top -u


 * Step 4: From base directory cd result


 * Step 5: vi .align

Example: TOTAL Words: 9762 Correct: 6086 Errors: 3703 TOTAL Percent correct = 62.34% Error = 37.93% Accuracy = 62.07% TOTAL Insertions: 27 Deletions: 2283 Substitutions: 1393
 * Step 6: Press Shift+G to take you to the bottom of the file to see your experiments results