Speech:Spring 2018 Isaac Marsh Log


 * Home
 * Semesters
 * Spring 2018
 * Proposal
 * Report
 * Information - General Project Information
 * Experiments - List of speech experiments

Week Ending February 5th, 2013
1/30/18 test edit. learned about the logs in class today, wanted to make sure I can record my events without any problems.
 * Task:

2/2/18  Joined the Discord chat for our class set up by Dan B. Also have started narrowing down times to meet in person with my group members Rose, Arias, and Tri to start researching. I have started reading about Sphinx and looking at last year's logs created by the graduated students.

2/3/18  Checking in, read other student's logs and progress. Still reading and learning about Sphinx.

2/4/18  Continued to read more information about Sphinx. Checked in with my group to see if anyone had done anything differently. Tri said he attempted to do a train but the process failed. This is something I definitely want to learn how to do in this week's upcoming class. We are deciding as a group to also meet outside of class on either Wednesdays or Thursdays.


 * Results:

2/2/18 Do more reading and learning about Sphinx. Plan to meet weekly outside of class with my group to continue research and start doing our data.
 * Plan:

2/3/18 I want to meet with my group outside of class this week coming up and figure out how to do tests, then start our first one.
 * Concerns:

Week Ending February 12, 2013
2/6/18 Our group met early before class to go over what we had done this week. Most of us had done research and reading other student's logs from past years. Rose and Arias both figured out how to do trains and will help show Tri and I how to do them.
 * Task:

2/8/18 Our group met at the UNH library to start reviewing the data we got when we all ran our first trains. I somehow had my train data in my student 2018 folder and had to delete it. I was able to re-run my train correctly in my Exp/0303/013/001 folder. We now went over each of our trains as a group and started learning how to do the Language Model.

2/9/18 Checking in and reading other logs. Starting the proposal with my group this weekend to figure out what we are going to focus on with our project.

2/11/18 My group worked on the proposal by different ways we can look at the data, running the trains as well as the language models and how to decode the data. All of the groups collectively put their proposals together so we can submit one document for everything. I also came onto the Wiki page to read other logs for the week.

2/6/18 steps we took to run a train: log into Caesar, then /mnt/main/Exp/0303/013(my student folder)/(any experiment folder). You can ssh into any of the drones at any time after longing into Caesar.
 * Results:

2/8/18 I now have the correct train data in my 001 folder, and not in my student folder.

2/6/18 Our group plans on making a proposal by the end of the week to help guide and direct us. We also all plan on knowing how to run a train by the end of the week too. We plan to meet this week on Thursday at 2pm to continue working on our trains. We also want to continue reading spring 2017 logs as well to find out where they left off.
 * Plan:

2/8/18 We are planning on continuing to read and research the Language Model and how to decode all of the data we have from our first train. We are going to come up with and design our project proposal over the weekend. We also want to read and look more into the SC light software that describes the scoring and alignment of our data (reference and truth text files, need to learn more about this for next week).

2/9/18 Continue working on our first draft of the proposal to have it ready for Tuesday's class. 2/6/18 The only concerns I have so far are learning how to interpret/ digest the data that comes from running trains and what exactly our group has to do overall to help support the other groups.
 * Concerns:

2/8/18 The first decode ran fine but the scoring didn't work because the hyp.trans file was empty. This is something else we have to look into this coming week.

Week Ending February 19, 2013
2/13/18 After meeting in class today, Professor Jonas told not to focus too heavily into getting the trains, language models and decoding to work properly by professor Jonas. He said we should focus more on the data itself and report back for next week on how it works, what it is, its relationship to the other groups, etc. and then do a short presentation to class on what we find.
 * Task:

2/15/18 We came together as a group on Thursday to edit our portion of the proposal. I had help and worked with Camden to fix my permissions on my 013 folder since I couldn't add the DELETE directory I needed for some parts of the data group. This folder is also affecting my trains even though it's not top priority this week. My group and I also did some reading on what specific parts of the data we need to look at and understand for Tuesday.

2/15/18 I was able to fix my folder issue with Camden today because he found out that Greta somehow had the editing permissions of my folder. He changed the permissions back to me and I was able to successfully add the DELETE folder I needed. This also fixed my issue with not being able to run trains and I was FINALLY able to run my trains correctly as well!! I will also update my experiments Wiki as well. when
 * Results:

2/19/18 I did research on which open source application would be able to filter out the background noise from the audio files, and I have decided that Audacity would be the best option for us to use as the free open source software to use in order to remove unwanted background noise. This is because it gives users visuals of the audio file, and we can see exactly when someone/something makes a noise we don't want. The application allows the user to see when the noise was created, and remove it by physical snipping it out of the original sound file so that it sounds cleaner. 2/13/18 My group plans to meet this Thursday in order to analyze as much of the data as we can in order to give the entire class depth on what kind of data we are all working on in each of our groups, since each group will eventually use the data because it pertains to running the trains.
 * Plan:

2/18/18 My group came back together over Friday and Saturday and was able to hammer out most of the revisions for our part of the proposal while asking clarifying questions with Professor Jonas. We are planning to finish up discussing the data of the entire project so that we know exactly what it is and how we want to present our understanding of it to the rest of the groups during class. We are trying to meet one more time before class time, around 10am Tuesday morning. One of the tasks I have before class is doing some research to identify a safe, open-source software program for UNIX-based systems that we can use to clean up audio files to get rid of background noise. I will be able to present some ideas on Tuesday to see what everyone in class will want to decide on (or just my group will decide, not sure on that yet.)
 * Concerns:

Week Ending February 26, 2013
2/25/18 My group has decided that we also need to listen to about 50 of the audio files a week and edit out any background noise or random noises in the audio files to that they can be understood better by the system better and hopefully give us a lower score overall in our percent to error.
 * Task:

2/20/18 I concluded from my research that Audacity will be the easiest as well as the opensource and free option for an audio editing we will need for our group so that we can edit out anything that isn't speech. Audacity has many different tool that I have read through and started playing with to get a better understanding of it, as well as actually testing it on some copies of the audio SPH files I found on the server that the trains use in order to give us our percent to correctness of the trains after they have been run. I came in early Today before class to meet with my group partners and show them how this works and what I have concluded from testing out Audacity. In order to edit the SPH files the steps I took were: logging into Caesar and copying a few audio files from /mnt/main/corpus/switchboard/5hr(can be any of the hrs folders)/train/audio/utt. One in the Utt folder there are tons of SPH files I can copy to listen to and edit. then I went to Convertios website to change the SPH files into a WAV file so that Audacity can read, play and open it correctly (it can use tons of audio files but SPH isn't a normal file type for it) Then I take the newly converted WAV file and open it in Audacity where I can see the exact speech waves and can listen to and pinpoint where in the file the background noise, speech, laughter, etc. is and can then either use a Silence Audio Selection button which takes a section of the audio you highlight and removes all sound from it (like a random noise that happens when no one is speaking) or use the Noise Reduction option which you can give it a sample of the noise you want to reduce, then can change the filtering, decibels, frequency, and sensitivity of the sound(s) you are trying to get rid of. I did two files just to see how both options worked and if it would make any difference in the audio file, and I found out that the Silence Audio Selection works amazingly on noises that are isolated by themselves, (I was able to remove a loud pop at the end of someone talking with this option) and I was able to edit another audio file that had some background buzzing almost like a radio, and get rid of most of it (can just faintly hear it with headphones). These are findings I will be able to present to all of the Groups in class tonight.
 * Results:

2/22/18 My group met again today to go over and edit more of the parseTrans and genTrans files. Currently we have to learn which Regular expressions will positively affect the parseTrans file so we can go in and make edits to the code in that file in order to have better read outs on the trains and experiments everyone can run, that way we know that we are improving the Language Model and the process of improving the word error rate. We would like to get the error rate down to 25% by the end of the capstone.

2/25/18 out of the audio files I listened to today, I wasn't able to find any problems with most of them except for 2. I was able to edit out the abnormalities and saved the edited audio files into a new folder I will consolidate with my other group mates each week.

2/21/18 Both my group and the modeling group meet in a group chat on Dischord in order to go over what we need to learn from each other about working on the data and how to edit the Pars Trans files. My groups is planning to meet up tomorrow at 1pm in order to go over more of the files we need to edit, and then provide backup for the modeling group on firday when they will try using our edited Pars files.
 * Plan:
 * Concerns:

Week Ending March 5, 2013
3/2/18 continuing to assess and edit 50 audio files a week with each group member and store them in a separate place from the main audio files that the trains use (for now until we have enough changed so we can try running trains with just the edited audio files.).
 * Task:

3/3/18 I went over all of the newer script edits that Tri had made, with Rose over Dischord chat (since we couldn't meet in person), as well as the email we received from Professor Jonas this morning explaining more of Tri's questions about the "-" character being removed from the final readout. We discussed that we had a list of things we want to tackle before Tuesday, as well as continue working on next week. Our list included making more regular expressions to make and add to our own versions of Gen trans, and learning Perl to modify more of the train building files.
 * Results:

3/4/18 completed some more of the audio files I wanted to listen to and watched Camden's video. Made plans to meet with Rose tomorrow to modify and edit more expressions and keep working on different versions of Gen Trans.

3/2/18 I have planned to meet with my group member Rose tomorrow (Saturday). I am planning on watching a video Camden, another classmate recommended at  Youtube this link. 3/2/18 Came down with the flu Tuesday and had to leave our class meeting an hour before it ended. Was bed ridden for three days and tried to keep up on most of the things going on. I want to also catch up on anything I may have missed with my group members as well.
 * Plan:
 * Concerns:

3/5/18 Trying to run a 5HR train on Obelix and running into an issue with the genFeats.pl -t command. Getting an error that says:

[iam1002@obelix 003]$ genFeats.pl -t

-t        (had an extra -t show up here even when I hand typed the code)

003_train.fileids

Fatal Error: Can not open etc/003_train.fileids!

Complete!

Run "nohup scripts_pl/RunAll.pl &" to begin training.

Rose found out that I didn't have file permissions to change anything in the 003 folder so I could have changed the folder permissions to allow me to run the train in that folder, but I added a folder with my own permissions to save myself some extra time.

Week Ending March 12, 2013
3/6/18 Class meeting today. Jonas helped us discover that part of the reason why Rose, Tri and I were all getting different word percent error rates is because there is an issue in one of the Trans codes that we need to discover and fix. He said we might want to work with the modeling group this week to see if we can find it and fix it. We met with Brian, Hannah and Steve to see if we could come up with a way, as two groups, to find and fix the error.
 * Task:

3/12/18 Jonas has assigned the 2 new larger groups today so we can focus more on the modeling aspect of the project. My new group members include Steve, Josh, Dan R., Yashna, Lamia, Tri, Jaden, and Faruk. We have started another server on Dischord for this group, appropriately named the Avengers by Jonas. We all talked for about a half an hour tonight about what each group has been doing and what we can do to help each other out. Since we need to focus mainly on the modeling stuff, Steve gave a great update and made a note of what he is stuck on and what we need to help with. I wanted to save what he said and mention it here for future referencing:

"These .sph files are showing up in the wav sub-directory when doing a Seen Decode and Scoring, but not when doing an Unseen. Somehow this was NOT the case with the previous semester, whether or not they altered the makeTest.pl or did so manually, we have yet to figure out. I have moved the .sph files manually into the wav sub-directory of the Unseen Decode and Scoring sub-experiment, but that has not fixed the problem when running SClite, which errors out with files not being matched or found. The wiki addresses this particular issue, but the remedy that is proposed there has not worked for me in fixing it."

3/10/18 I have been working on editing more of the audio files that are used for the trainings and experiments. Continuing to listen and edit (if need be) the audio files to help clean up any noise or issues with the sound files so we can have clear and concise audio across all of the different audio training hours. Also staying current with all of the emails that have been happening back and forth between all of the groups and professor Jonas about how we as a class haven't been letting the decode happen as long as it should have, and having a problem with a piece of the code when we run our trains. My group has been talking and researching through the different versions of ParsLM.trans, Hyp.trans, and Gen.trans to try and find the issue. Jonas thinks it's our trans_parsed file in our Language Model that is corrupted, so I will be taking a look at that between tonight and tomorrow. Going forward to help prevent this issue for future students, it is important to check the output files and not rely on just the scripts that run to cath errors in the code.
 * Results:

3/11/18 Started to go through the different files as listed in my 3/10 Plan to make sure they are formatted correctly. There is a lot of information and code in each file and almost overwhelming, but with the ability to use Google to decipher the things I don't understand helps a lot. I did some more researching and trying different kinds of regular expressions as well to help with the decode and language model today too. I also talked with Tri and Rose about the latest emails from Jonas, meeting up with the modeling group sometime soon, and being able to generating our own dictionaries without scripts right in the Terminal with the sed command.

3/10/18 We have been tasked to clean up our experiment logs in the 0305 folder, get rid of duplicate experiments/ consolidate them into one, checking the tmp.arpa file, dictionary, and train.trans are formatted correctly. A lot on our plates to do, but thankfully it is spring break this week and I can devote a couple days to just going over everything we need to make sure everything comes out fixed and positive for the next class meeting.
 * Plan:
 * Concerns:

Week Ending March 26, 2013
3/19/18 Over spring break I worked more on listening and editing any of the .sph files that needed cleaning up. I also worked on trying to edit and make new regular expressions in order to have some for the permanent corpus we are designing. Rose and I are going to create our own 5hr corpus first and as long as it works correctly, then we can work towards completing the permanent 300hr corpus as well.
 * Task:

Rose and I met up together today to start working on our version of the corpus. We started by researching information about the sph files because we need to come up with a way we can use random sph files (like taking every Nth file out of the thousands of sph files we have) when running trains from our corpus, but we don't want the same speakers (who speak in the sph files) in the train data set and the testing data sets, which are the dev.trans and the eval.trans. We also looked at some of the pdf whitepapers that rose found, Comparing Human and Machine Errors in Conversational Speech Transcription and Telephone Speech Recognition which talk about technical modeling and different ways the switchboard has been used by other companies.

3/23/18 I talked both with my Avengers group and Data group to see what we need to do over the weekend and where we need to be in place for Tuesday's class. Our Avengers group was given a written set of instructions from our team member Steve on how to run unseen data. All of the members of our group are going to read this and make sure we know how to do it by Tuesday. I have also started and had issues trying to download the 300hr and Full folders I need if Rose and I are going to build our own Corpa. I started by trying to download the Full folder directly from the Caesar machine with Camden from the systems group (He had access to the server room for the whole capstone) but the machine didn't like the format of my hard drive and reformatting it didn't help either. The other issue it gave us was "omitting directory" every time we tried transferring the folders, which means that it couldn't use the link created by Linux in order to transfer the files (we assumed that the files were located in a different folder somewhere that the 300hr folder calls upon, but we couldn't confirm it.) I then tried to use Filezilla even though we shouldn't rely on it all of the time, and was able to transfer most of the full folder in one session (filezilla can pick up where it left off if your computer goes to sleep) I am in the process of finishing the download of the full folder and starting the 300hr folder as well. I have set up my computer to not shut down or go to sleep through the night so that it can download all of the files in one go (hopefully- will update tomorrow).

3/24/18 My transfer of the entire 300hr folder didn't work. it could see all 8300 something audio files but error-ed out after that sometime during the night when i left my computer on all night and refused to transfer anything. I tried transferring it again and it still wont work. the good news is that the full folder doesn't have that problem and is currently downloading the rest of that folder into my flash drive. I'm going to have to figure out why the 300hr doesn't want to transfer, and because there is so many audio files in the 300hr it still freezes when trying to open the Utt folder so I am going to see if I can transfer just that whole folder without opening it. I also went through and read the awesome document Steve had made for all of us in the Avengers group about training unseen data and I will try and run a test 5hr experiment tomorrow based on his instructions.
 * Results:

3/25/18 Just a quick update tonight, the full folder has somewhere around 1,950,000 files in it that I started downloading yesterday to have a copy to use for designing a corpa. Some files aren't transferring and I can't figure out why at the moment because I still have 1,472,900 files downloading (I have had my computer on for almost 2 days straight downloading the whole folder, thankfully the files are small and I only have 17GB downloaded so far.)

3/20/18 First class meeting after spring break. My new group, the Avengers, will meet together for the first time after class. During class we went over what some people from each group did over the break, including all of the pdf whitepapers that Rose found (the ones I linked above in 3/19/18. She was able to find out we might not need to create that entirely new corpus we had originally set out to do. I will check in with Rose and Jonas to figure out what the outcome of that will be.
 * Plan:

I met with my other members of team Avengers and we have decided collectively as a group that we want to have the knowledge and understanding as far as Steve does in our group 3/19/18 one thing I learned today is that the entire switchboard (if it's the version we are using) is only 260hrs, how are we able to do a 300/311hr trains? I found this information thanks to Rose at Switchboard-1 Release 2 I found a small section that describes the audio files and an intro to the switchboard "consists of approximately 260 hours of speech and was originally collected by Texas Instruments in 1990-1, under DARPA sponsorship. Switchboard is a collection of about 2,400 two-sided telephone conversations among 543 speakers (302 male, 241 female) from all areas of the United States." But because of this information, we now know that all of the audio files we use are spoken by 543 different speakers.
 * Concerns:

Week Ending April 2, 2013
3/31/18 Came on to read other logs and check in. Going through the 5hr Train I started on Trubadix, and will check in with my Avengers group tomorrow. I will ask and see if we need to do a 30hr train next because I believe there has only been 5hr ones run on it so far.
 * Task:

3/30/18: Coming on to start a 5hr train on Trubadix, I need to remember that the decode will take a few hours too so I don't end it too soon. I also talked with Rose and Tri and we will be meeting Monday at 3pm in order to construct our URC poster. I also was able to successfully download the entire 30hr folder with all of the audio files to make the corpa with Rose. the 300hr one still won't download at all and I will check in with Camden on Tuesday during class to see if he knows any other ways to obtain it.
 * Results:

4/2/18 I met up with Rose today and we spent 3 hours making our poster for the URC conference. We also talked about the entire class poster that we have to do as well. We will report to our groups and talk tomorrow during class about the second poster we have to make as a class too.

3/27/18 Our Avengers group met after class today and went over our tasks for this week. Steve has sent us all a version 5 update to the directions he made to running unseen data trains. Dan R and the Systems group made Trubadix, a new server, and we will be using that one for our group to do trainings on as one of the 3 we can use. I will be running a 5hr unseen train on Trubadix to make sure that everything works the way it should now that it is up and running. 3/31/18 My download of the entire full folder finished on Tuesday night and I forgot to mention that I had a huge problem with it. out of the entire 1,900,00 something files, I was only able to download 600,000ish (about 27GB) of them because most of the files had error-ed out with a message of " unable to transfer file" which I couldn't figure out why it wasn't letting me copy them. I had seen online doing research that it could be either my flashdrive isn't a compatible format to work with Linux, or that the files weren't in that folder to begin with and are only called there when they need to be used (due to being a Linux based file system) so I need to take a look through the main server on Tuesday as well to get the rest of the files.
 * Plan:
 * Concerns:

Week Ending April 9, 2013
4/3/18 Our Avengers group met after after class to go over the stuff we should do for the week. I had trouble trying to run the 30hr train on Traubadix but was able to find my error and fix it (I was running step 16 as makeTest.pl -d switchboard/30hr 025/025 026/026 instead of running it with the main experiment numbers like: makeTest.pl -d switchboard/30hr 0305/025 0305/026 which is the correct way). At this time, the decode is running right now, and what we know from professor Jonas is that we need to let the decode happen for a few hours so it can complete properly. Our class as a whole also has to work on and complete a poster/ slide presentation for another conference aside from the URC happening in April too.
 * Task:

4/6/18 I didn't have a chance to check the train I ran Tuesday until tonight and the results came back with an error somewhere in the decode process. I am re running the decode through the night and I will check it before work tomorrow morning. The error I got this time is: /usr/local/bin/sphinx3_decode: error while loading shared libraries: libs3decoder.so.0: cannot open shared object file: No such file or directory. this is inside my decode.log file itself and nothing else. The dictionary populated correctly so I asked my Avengers group and see what they say tomorrow as well as re running the decode now. I also checked the word count in the decode to see if that helped and came back with 0 0 0 0 instead of 3992. I will report back tomorrow about my findings and result with fixing this issue.
 * Results:

4/9/18 I was finally able to reach Steve and he told me that the error I'm having with the 30hr on Traubaidx is a lingering issue that was cloned over from Miraxulux (since traubadix was made from miraculix). The systems group had fixed the issue after the clone was already made, and I notified my entire group about it so that we can get it fixed this week and be able to run the trains / LDAs correctly. 4/7/18 I'm still having a problem running the newest train. I am having the same problem I did last night, even with deleting my decode folder and redoing that. My next step is to completely run an an entirely new train tomorrow from scratch. the error I was having is
 * Plan:
 * Concerns:

[iam1002@caesar etc]$ tail -f decode.log /usr/local/bin/sphinx3_decode: error while loading shared libraries: libs3decoder.so.0: cannot open shared object file: No such file or directory

I know this is related to the decode.log in the etc folder but I also believe it's related to my decode not running properly because when I also run the grep command to see how many words are in the decode.log this comes back:

[iam1002@traubadix etc]$ grep FWDVIT decode.log | wc

0      0       0

[1] + Done                          run_decode.pl 0305/025 0305/026 5000

which says there isn't anything in there (and I did read my decode log which only has the same error as above and then is completely blank through the rest of the log file.

I will be getting help tomorrow and Monday from my Avenger group to see if we can fix this problem, if I can't fix it by running the entire train over from the beginning.

Week Ending April 16, 2013
4/10/18 During our class meeting today the Avengers group met and we discussed how we want to approach our way of beating the other team by the end of the semester. Professor Jonas is giving us permission to keep logs "off the record" (paper copies that we can upload at the end of the contest/ give to him at the end of the semester) so that we don't have to expose our method of (hopefully) winning. We are starting our ideas right after class together and will keep a separate log of it.
 * Task:

4/12/18 Tri helped me figure out the problem I was having with the decode on Traubadix. We found out that drone cannot run decodes at the moment but can start trains and get all the way up to that point. He suggested I change to Obelix when doing the decode portion and that did the trick. I let the decode run and checked the final results from the Train. the results for the unseen are:
 * Results:

|===================================================================| | Sum/Avg | 3912 55254 | 50.3   39.2   10.5    9.6    59.3    91.9 | |===================================================================| | Mean   |  1.3   18.1 | 59.2   33.4    7.4   18.3    59.1    91.9 | | S.D.   |  0.5   16.2 | 22.5   19.3    9.4   38.8    40.4    25.5 | | Median |  1.0   13.0 | 56.7   33.3    3.6    6.3    58.3   100.0 | `---'

4/14/18 I did some secret Avengers group work with my groupmate Dan R. We have updated our logs that are off the record and will upload them when the contest is over against the other group.

4/15/18 I finished up and sent Dan R. my part of the secret avenger stuff we are doing towards beating the other team. I have updated my avengers private log as well.


 * Plan:


 * Concerns:

Week Ending April 23, 2013
4/20/18 I have been given another special avengers task to complete for next week. This one is longer and I will update my separate log with the details.
 * Task:

4/18/18 I had been given positive feedback for the secret avengers group work I have been doing. I have explained it in detail in my separate log.
 * Results:

4/22/18 Checked in with my group and talked about the new tasks we are completing for this coming week. I worked on some of it as well.

4/23/18 I stayed up really late to finish the special group work for the Avengers. I have made notes about it in my off record log.


 * Plan:


 * Concerns:

Week Ending April 30, 2013

 * Task:


 * Results:


 * Plan:


 * Concerns:

Week Ending May 7, 2013

 * Task:


 * Results:


 * Plan:


 * Concerns: