Speech:Spring 2013 April 3rd Group C
- Harry Dodson
- Josh MacPherson (team leader)
- Michael Mailloux
- Marc Southard
- Michael Tormos
- Sam Workman
Week 1 (3/27/13)
JM: The first task for Group C is creating new experiments and running a first training session for each. Each member of the group will have an experiment of their own and will run a train for that experiment. We had an online meeting on Friday in which I demonstrated the general process for running a train. Our goal for next week is to move everyone on to decoding and scoring their experiments.
HD: I was able to run through the first part of the experiment ok, training. There were some places were I was unclear on the directions but I made it through them. The only other issue was adding words from the transcript to the dictionary.
MM: Couldn't make the meeting last Friday, but the updated instructions on the wiki as well as help from Josh was enough. Today was my first time trying to create an experiment, was pain free and went well. Will try actually running the train at home due to time constraints. Creating the transcript was time consuming compared to the other steps of the process.
-Running the train today 3/31
-Tried to start running train multiple times. Added list of words not it initial dictionary to the dictionary, got no errors. Try to run the train again, still have the same set of words stopping the train? Not sure how to resolve this one. Will update further. EDIT: Saw this line, "Please note: Each time you run the Sphinx trainer, the output will be added to the end of this document. So in other words, to get the list of words preventing the last executed train from running: Scroll all the way down!" Saw this, had a chuckle, moved on to next problem. Had to update <experiment #>.phone with vi to update the added pronunciation to the phonelist. Train running as we speak, errors have been found but it is still running.
MT: Started Exp 0036, 0049, 0064. Exp 0036 and 0049 ended in failure. I made contact with several fellow students to troubleshoot the errors I was receiving, which I posted about in my weekly updates. Exp 0064 is up and running and the errors I was receiving on Exp 0036 and 0049 have been corrected. I got up to Scoring the train right after the DECODE. The following is what I'm receiving when running the following command:
sclite -r 0064_train.trans -h hyp.trans -i swb >> scoring.log
I will continue to push forward and get past this problem, if it is even a problem to begin with.
SW: Getting everything set up for running trans and then having everything break on me.
March 27, 2013 Miraculix wasn't able to setup an experiment so I changed over to Obelix. Created folder for experiment 0042. Configured the sphinx_train.cfg through vi. Ran the getTrans to generate the transcript to be used.
April 1, 2013 Generated the Phone List without much of a problem. I might mention how it would be worthwhile saying which folder you should be in for that step as it was where I last left off. It's currently only implied based on the previous step. Generated the feats data without any problem.
April 2, 2013 Have started debugging the train. There are a lot of words missing from the dictionary so I'll begin the process of removing them to get the train up and running hopefully.
April 3, 2013 Spent the later half of April 2 and this morning of April 3 getting the 50+ errors in my phone/dictionary running. Now the train is actually running properly with a few errors getting picked up. Finished the language model, decode and scoring.
Week 2 (4/3/13)
JM: Group C has moved on to the decode and scoring portion of the modeling work. Every team member should now be comfortable getting through the entire modeling process from experiment setup through final scoring. The issues that some people were having the first week have largely been resolved and everyone now has at least one scored experiment to their name. Moving forward our group will be working closely with Group B, as we move onto larger portions of the corpus. The increased size of the data set should yield improvements in our acoustic model.
HD: Ran through a successful experiment, 0066. No major issues while running through, every error I had there was a section to resolve them.
MM: Created and ran through a full experiment in class today from running the train to scoring ([]). Had expected errors throughout, was able to resolve them with little difficulty.
Completed Exp 0064 with the aid and help from my fellow students. Here is the following score to Exp 0064:
'-----------------------------------------------------------------' | Sum/Avg | 518 11051 | 82.0 10.5 7.5 7.5 25.5 91.5 | |=================================================================| | Mean | 2.7 58.2 | 81.4 11.7 6.9 11.7 30.4 93.0 | | S.D. | 1.7 43.0 | 12.6 9.3 8.0 16.3 19.9 17.8 | | Median | 2.0 52.5 | 82.7 10.3 5.7 6.9 26.6 100.0 | '-----------------------------------------------------------------'
04/03/13 Set up mini train and got it running after adding a few words to the dictionary.
MS: Ran through experiment 0042 successfully. The problematic issues I ran into were described well on the wiki. Refer to http://foss.unh.edu/projects/index.php/Speech:Exps_0042#Description to see more.