Speech:Spring 2016 Ryan O'Neal Log


 * Home
 * Semesters
 * Spring 2016
 * Proposal
 * Report
 * Information - General Project Information
 * Experiments - List of speech experiments

Week Ending February 9, 2016

 * 2/3: Read the entire wiki to include past logs from previous semesters. Research the various groups assigned in Capstone. Ponder which groups to sign up for. Pick top 3 groups. Get familiar with Linux scripting commands to be used in Caesar.  Log into Caesar.  Look up directories for experimentation.  Look at scripts that run a train. Start communicating with peers over proposal due the end of February.


 * 2/4: Our group started a chat session via group messaging over text.  Was no able to physically meet up with group due to my work hours.  Read some logs concerning running a train, creating a language model and running the decode.


 * 2/8: Received an email from professor Jonas concerning new log in ID's.  Our group continued to go through the scripts and learning commands.  I was not able to meet with anyone and the school shut down at 3:00 pm due to inclement weather.


 * 2/9: Our group started a group chat session again.  Everyone is preparing for tomorrows seminar.   I continued to read through logs to get a better picture about what it takes to be in the modeling group.  Seems the experiment group and the modeling group work closely together.  Not sure what the other groups need from the modeling group.  I would like to set up a channel with the other groups.  Perhaps Slack might be the best alternative if nobody can do a google hangout.

[root@caesar Exp]# cd 0283 [root@caesar 0283]# chmod g+w 002
 * 2/9:  8:00pm: Learned about SSH into cisunix.unh.edu first before SSH into caesar.  Logged in with username and changed password using passwd command.  Successfully Logged in as root and changed permission using:

Copied over the transcript from the corpus directory and prepared the transcript and copied the transcript that creates the language model and executed the script. Will decode tomorrow before seminar.

Week Ending February 16, 2016

 * 2/10:
 * Task: Duplicate last semesters processes in a 125_hr train.  Get the word error rate down.  Get the Real time factor formula.  Find a way to see how long the train took to complete.  Look at convergence ratio, variance normalization, and density in the logs and sphinx wiki.  Update the deadlink for SPhinx guide and replace with new link.


 * Results: Started 125_hr train at apx 4:50pm on 2/10/16.  Looked up the command for seeing how long a train actually took to run.  Used ps -o stime,time,cmd.  Looked up the formula for real time factor and came out with P/I where P= time it took to train and I = time in real time.  Updated sphinx guide to   Used nano_sphinxtrain.cfg for configuration of attributes.


 * Plan: Continue going through sphinx guide and wiki logs to find out about convergence ratio, variance normalization, and density.


 * Concerns: Thinking the 125_hr train will take 5 times longer than a 5_hr train.

Task: Check on decode. Still waiting for decode to complete. Need to run SCLite. Results: N/A.
 * 2/14:

Results: Still waiting for decode to complete. Received twitter update about not using any other products for proposal other than wiki. Working on proposal with group. Worked on the proposal and took the online document from Access 365 and put it on wiki per Prof. Jonas request. Emailed the creator of the document and had them delete the file from Access 365.
 * 2/16: Task: Check on decode.  Still waiting for decode to complete.  Need to run SCLite.

Week Ending February 23, 2016

 * 2/17: The decode finished last night and the scoring was started at 12:00pm today.  James and I prepared the hypothesis transcript.  The SCLite scoring was done.  The results were posted on the 0283 003 experiment log.  here is the score:

SYSTEM SUMMARY PERCENTAGES by SPEAKER

,-.     |                            hyp.trans                            | |-|     | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err | |-+-+-|     |=================================================================|      | Sum/Avg |10000  143752| 79.7   15.0    5.3   15.0   35.4   94.6 | |=================================================================|     |  Mean   | 57.5  826.2 | 79.8   15.2    5.0   16.0   36.3   94.8 | | S.D.   | 25.2  341.8 |  6.2    4.8    2.3    7.2   10.5    5.7 | | Median | 55.0  785.0 | 80.7   14.6    4.7   14.3   34.5   96.0 | `-'

Worked more on the proposal. Created a {| class="wikitable"} for the Introduction. This includes all the groups and the names of the group members. James and I thought this would be better to use the wiki markup instead of HTML. This is because James and I believe the Prof. Jonas told the class not to use HTML tags in the wiki.

Start a new train: Create directory 0283/004. Set up directory structure. Configure the sphinx_train.cfg for variance normalization to yes, senomes to 8000, convergence ratio to 0.001 and change density 64.

Concerns: Making sure the 125 hour train will be configured correctly and produce the results we want.

7:30pm: Stated new train using script command nohup scripts_pl/RunAll.pl &


 * 2/20: Train was finished and created Language model. Made the lm directory than put a trans_unedited file in the lm directory. Prepared transcript. Started decode. Did not make the LM directory correctly so used mv command to change lm to LM. re-executed decode and started decode at 11:30pm.


 * 2/21: Decode took 5 hours and 24 minutes. Analyzing the results indicates that the Error rate is 33.9% which did not change very much from the last Error rate which is 39.4%. Results and scoring for 125hr_train:

SYSTEM SUMMARY PERCENTAGES by SPEAKER

,-.     |                            hyp.trans                            | |-|     | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err | |-+-+-|     | sw2001a |   32    541 | 79.9   15.9    4.3   22.9   43.1  100.0 | |-+-+-|     | sw2001b |   34    488 | 81.8   13.9    4.3   27.3   45.5  100.0 | |-+-+-|     | sw2005a |   53   1172 | 87.7    9.1    3.2    9.9   22.2   92.5 | |-+-+-|     | sw2005b |   77    817 | 74.5   18.0    7.5   26.4   51.9   96.1 | |-+-+-|     | sw2006a |   40    608 | 85.9   12.0    2.1   15.5   29.6   95.0 | |-+-+-|     | sw2006b |   43   1012 | 79.0   13.3    7.7    6.6   27.7   97.7 | |-+-+-|     | sw2007a |   86   1064 | 82.8   11.6    5.6   13.5   30.7   90.7 | |-+-+-|     | sw2007b |   80   1183 | 83.9   12.1    4.1   10.1   26.2   88.8 | |-+-+-|     | sw2008a |   28    369 | 86.4   10.3    3.3   16.8   30.4   96.4 | |-+-+-|     | sw2008b |   32    436 | 83.7   13.5    2.8   18.1   34.4   93.8 | |-+-+-|     | sw2009a |   37    605 | 74.7   17.9    7.4   11.2   36.5   97.3 | |-+-+-|     | sw2009b |   44    649 | 81.5   14.2    4.3   17.7   36.2   95.5 | |-+-+-|     | sw2010a |   38    528 | 87.9    9.3    2.8   17.6   29.7  100.0 | |-+-+-|     | sw2010b |   33    659 | 73.3   17.5    9.3   13.8   40.5   97.0 | |-+-+-|     | sw2012a |   67   1420 | 83.2   11.0    5.8   11.8   28.6   94.0 | |-+-+-|     | sw2012b |   43    846 | 82.9   12.8    4.4   12.3   29.4  100.0 | |-+-+-|     | sw2013a |   52    766 | 75.5   18.1    6.4   12.5   37.1   92.3 | |-+-+-|     | sw2013b |   88   1526 | 67.0   21.6   11.4    8.5   41.5   96.6 | |-+-+-|     | sw2014a |   23    311 | 75.9   20.3    3.9   38.9   63.0  100.0 | |-+-+-|     | sw2014b |   27    543 | 77.0   17.1    5.9   13.4   36.5   92.6 | |-+-+-|     | sw2015a |   29    611 | 80.0   11.8    8.2    3.3   23.2   93.1 | |-+-+-|     | sw2015b |   14    181 | 80.7   12.2    7.2   11.0   30.4   92.9 | |=================================================================|     | Sum/Avg | 1000  16335 | 79.9   14.2    5.9   13.8   33.9   95.0 | |=================================================================|     |  Mean   | 45.5  742.5 | 80.2   14.2    5.5   15.4   35.2   95.5 | | S.D.   | 21.2  354.6 |  5.2    3.5    2.4    7.9    9.6    3.3 | | Median | 39.0  630.0 | 81.1   13.4    5.0   13.5   32.6   95.8 | `-'

Week Ending March 1, 2016
2/24/16:


 * Task: To increase dictionary size limit and decrease the WER in new sub-experiment 005.


 * Results: TBD


 * Plan: Modify scripts to improve vocabulary to 30000 words.
 * Concerns: Looking for the correct scripts in caesar.  Being able to find the right commands and modify the scripts.

2/28/16:


 * Task: Conduct group meeting and add one team member from other group to perform train experiment.  Finish proposal.

need to be fixed. Verified the train was conducted with Neils group (Systems group) so we built the Language Model did the decode and scoring to give him a good idea of that process. His results should be in the Systems Experiment Exp Log and/or his Log
 * Results: Neil met with our group in Concord. We went over the tutorial and identified a couple inconsistencies that

Went over the proposal and created tasks for each group member throughout the semester. Came up with my tasks: Ryan *Coordinate with other teams to disseminate information and act as a train/decode/score advisor. * Research the feasibility of switching to phonetically-tied mixture acoustic models (PTM) in order to reduce training time while preserving accuracy *Continue to research Sphinx language models and variables that affect the WER. *Look for inconsistencies in wiki tutorials and update and modify the wiki tutorials with proper wiki markup format.

Finished the proposal with the other implementations tasks and sending/answering email with Ben over what the word count is in our vocabulary and the train time. Found out from James the word count is 20,004 (004) vs. 23,573 (005). Train time is the same. Assisted the master editor with his communications among other groups about their final proposal changes.


 * 0283/005 Results:

| Sum/Avg | 1000 16335 | 80.4   13.7    6.0   13.4   33.0   94.7 | |=================================================================|    |  Mean   | 45.5  742.5 | 80.7   13.8    5.5   15.2   34.5   95.1 | | S.D.   | 21.2  354.6 |  5.5    3.8    2.5    8.3   10.7    4.6 | | Median | 39.0  630.0 | 81.0   12.9    5.0   13.0   32.3   95.8 | `-'

Word error rate went down .9% from 0283/004 sub-experiment WER: 33.9

Week Ending March 8, 2016
3/2/16:

Come up with a way to decrease WER on a 256 hour train.
 * Task:

TBD
 * Results:

We found out that corpus data approximately 11,000 utterances and starting at the first 32,000 files are corrupt. We have also been instructed that our 125 Hr train configurations will have little effect on more data from a 256 hr train. We need to start running 256 hr trains and cannot use the same configurations. We will create a 256 Hr train. Copy the first 30000 utterances are copied over from corpus directory to new corpus directory. Create a new directory in caesar as: /mnt/main/corpus/switchboard/fixed/30k. Possibly reducing the word count from 10 to 5 will cause the train to run much longer than previous experiments. This configurations might not be as effective or will take much longer than a 125 hr train.
 * Plan:
 * Concerns:

3/3/16:

Come up with a plan to modify the new train configurations and make sure that the train will work effectively with the new corpus.
 * Task:


 * Plan: Must research and email Prof. Jonas about some of the problems with the new corpus and the new train.  James had pointed out that the audio time did not match up with 320.752 hours : 311.76113 hours.  James indicated there is a 2.88% increase in audio hours from the /utt compared to /trans/train.trans.

The data group is responsible for checking the integrity of the audio files.
 * Results:

Being able to get this train to run.
 * Concerns:

3/8/16: Answering emails and reading emails from the data group, Prof Jonas and other members of the modeling group. Read logs from previous semesters.
 * Task:

Need to learn the different mixture weights and other configurations in the sphinx model building tutorial and read more logs from previous semesters to look for new possible configurations. This in combination with inconsistent wiki entries in the wiki logs will provide a better explanation into what train configrations actually work and don't work.
 * Plan:

After perusing these logs and the sphinx tutorial I have still to find these configurations and explanations that will guide me through becoming more active in the modeling group.
 * Results:

Constantly steering off into other group activities has caused me to become less active in the modeling group which is not necessarily the best practice for team research. If I am able to use spring break to catch up by working only 3 days instead of 5 will help me keep up tremendously. Finishing the internship next week by the 18th will determine my ability to spend more time on Capstone research throughout the semester.
 * Concerns:

3/11/16:

Work with James and Jon to find out about the new corpus and to offer assistance with that process.
 * Task:

Continue reading the logs to find out if the new corpus is functional or find out what needs to be implemented for it to be set up correctly.
 * Plan:

Was unable to work with Jon and James due to a scheduling conflict.
 * Results:


 * Concerns: I will reach out to them when the weekend is over. Jon is working all weekend.  I will set up a meeting and go over the new corpus details with them Monday.

3/14/16:

Work with James and Jon to find out about the new corpus and to offer assistance with that process.
 * Task:

Meet at the True Brew Barista in Concord. Identify the new corpus data directory structure and attempt to resolve any issues or add any new changes. Read through all of the emails over the weekend. Attempt to fix any discrepancies with the audio and transcripts being used.
 * Plan:

I met up with them and we went over the new corpus. Had the server group set up servers for Team Stark and Team Cap. Team Stark will use Obelix, Idefix and Team Cap will use Asterix, Miraculix. Emailed the groups about using the "top" command before running a train to verify if another train is in progress. This is a good practice technique over all.
 * Results:

I will reach out to them when the weekend is over. Jon is working all weekend. I will set up a meeting and go over the new corpus details with them Monday.
 * Concerns:

3/17/16:


 * Task: Email Stark group members to guide them in researching and getting familiar with running a train.


 * Plan: I would like to inform Stark group members about the process of running a train. I would like to help advise on what happens and where files are sent to specific directories and displaying the information in the email and asking people to read over the guides sent by Matt will help explain the importance of the train process.  Language model building requires paying strict attention to the steps and being able to see the details of this process that show people how it works.


 * Results: Sent this email to the entire group:

I strongly advise to look at the CreateUtts.pl script here at:

Https://foss.unh.edu/projects/index.php/Speech:CreateUtts.pl

This is a fundamental script that generates the utterances from /utt directory and also generates train.log, utt.log and conv.log and this is a prime example of using audio transcripts to generate utterances so you can get an understanding of model building from the development of the files on a script by script basis. that's what the key to running a train and getting to know the terminology of the model building process. Just start here and go through all the steps of the script and start learning the "vocab" of language modeling.

and understand that the log files will be:

train.log: utterance start time -- utterance stop time -- diff conv.log: utterance stop time -- length of conversation -- diff utt.log: expected utterance -- actual utterance -- diff

as this: train.log Content & Format  utt.log Content & Format  conv.log Content & Format 


 * Concerns: Not being able to guide Stark team members in understanding what scripts are used and where the files are being sent when running a train.

Week Ending March 22, 2016
3/21/16


 * Task: Going over running a train and the error message that Neil sent me. The error came from a segmentation fault (core dump) when he ran his new train on Asterix  Neil's Train take 2.  This error was caused by an error in the decode process.  I believe that the problems with the new corpus that was set up could be causing this.  Need to investigate and contribute the information I find.


 * Results: TBD


 * Plan: Going to meet Neil on Wednesday and have him collaborate with the entire modeling group.  Might be able to throw in this question during the video conference call with professor Jonas.


 * Concerns: I think we are going to find out that a lot of trains were attempted but unsuccessful during spring break.

3/22/16


 * Task: Going over running a train with Saverna.  She attempted to run this train today.


 * Results: TBD


 * Plan: Going to offer assistance via text if need be.


 * Concerns: I think we are going to find out that a lot of trains were attempted but unsuccessful during spring break.  The problems with other capstone members experiencing the same issues might prove that there is something wrong with either the generateFeats or corpus entirely.

Week Ending March 29, 2016
3/23/16:


 * Task: Setup a video conference call through Google Hangouts with Professor Jonas and the entire capstone class. Go over some of the new corpus scripts set up by Jon, James.  Collaborate with the Stark team to find out who has run a train and who has not run a train.  Try to get priorities set up for server access so that team members who haven't run a train yet will be able to set that up without interference.


 * Results: Discussed many concerns with Professor Jonas over the new corpus that was set up and why the utterances were not being generated.


 * Plan: Going to talk to some people in the Stark team to go over the problems with the new corpus with James, Jon and Ben and see when exactly they can run a train.


 * Concerns: Main problem is setting up meeting times with other team members and seeing when exactly they can run a train before next capstone meeting and simultaneously trying to set up the corpus to allow trains to be set up and run error free.

3/25/16:


 * Task: Research the corpus structure and collect information about the various types of modifications being done.


 * Results: Reading websites about sphinx modifications:


 * Plan: Go over links through google or other search engine sites.


 * Concerns: Making sure to keep up with email threads.  Contributing a script or two will be challenging and I do not want to modify or change anything without notifying other group members.

3/26/16:


 * Task: Research forced alignment and get in touch with Matt H about force aligning transcripts.


 * Results: Found a good link here align using sphinx 3


 * Plan: Get in touch with Matt H before next capstone meeting.


 * Concerns: The issue will most likely be handled and corrected before I can really contribute much.  I'm going to have to come up with a plan on what areas I can focus on and handle between capstone meetings so that I can successfully contribute a project goal or new modification to language modeling techniques.

3/30/16:


 * Task: Capstone meeting. Meet with Team Stark and assist group members with running a train.


 * Results: Met with Saverna to help reset her password and help her run a train.  Helped with the decode as well.


 * Plan: Going to contact members of Team Stark in identifying what changes need to be made when running a 5hr train vs 145hr.  Going to help Jon with his train.


 * Concerns: None at this time.

Week Ending April 5, 2016
4/2/2016


 * Task: Checking in with Jon and Team Stark.   Servers seems to be down and cannot get a train running.


 * Results: None at this time


 * Plan: Going to read over emails and see what the problem is


 * Concerns: To run a train on Obelix or Idefix might be delayed until weekend is over.

4/3/2016


 * Task: Just checking in and reading emails.


 * Results: Reading emails and keeping track of progress with team Stark.


 * Plan: No other plans other than to see when the servers will be back up.


 * Concerns: Might delay running a train for a couple days.

4/4/2016


 * Task: Talk to Jon and Team Stark members about running a train and seeing how progress is going.  Contact systems group to see what the status of the servers is.  Going ot reset my passwords on Obelix and Idefix because I only reset my password for Caesar in the beggining of the semester and not for the other servers.  These two servers are assigned to Team Stark and I feel I do not need to update my password for the other servers.


 * Results: Found out from Neil about a power strip that was faulty and needed to be replaced.  Jon will start a train tomorrow night and I notified him that I can help with anything he needs.  Successfully updated my passwords on Obelix and Idefix by using passwd command under root login.


 * Plan: To help Jon run his train and any other Stark team members with the language model building process.


 * Concerns: None at his time.

4/5/2016


 * Task: Checking in before Capstone meeting.  Going to see how Jon Trimble's train did.


 * Results: Looked at the results of score.log for Jon Trimble's experiment 0286/007 and found the WER to be "29.6".  Not sure if this is because of the changes to the corpus and the fact that it is a 5 hour train leads to more questions than answers.


 * Plan: I need to touch base with Stark team members and see what exactly needs to be achieved.


 * Concerns: Not sure if running a 5 hour train is going to help with our teams progress and I cannot run a 145 hr train because I am out of time due to servers being down over the weekend and my password issues.

Week Ending April 12, 2016
4/6/2016


 * Task: Meet with Team Stark at Capstone meeting. Gather results from trains conducted over the past few days from other team members and try to formulate a plan for lowering the WER to beat Team Captain America.  Get the template for the URC poster presentation and schedule a meeting with the Modeling group to make that poster over the weekend.


 * Results: Met with Team Stark.  Went over some of the results.  More to follow.


 * Plan: TBD


 * Concerns: None at this time other than getting the poster completed and lowering the WER on future experiments.

4/9/2016


 * Task: Meet up with the modeling group to work on the 2016 Modeling group URC poster.


 * Results: Finished Poster. James emailed it to Professor Jonas.  Learned some new Linux commands from James.  Learned about the difference between Unseen data and Trained data from Jon.  The difference is that Trained data is audio data that the corpus has already heard and processed through train experiments and Unseen data is data that the corpus has not heard or had any train experiments conducted through.


 * Plan: Talk with modeling group members about researching sppech recognition patterns and learning more abou the difference between unseen data and trained data.


 * Concerns: Not to divulge any pertinent information about Team Stark operations while simultaneously working with the modeling group on Speech research.

4/11/2016


 * Task: Run a 300 hr Baseline train so I can run another train right after and make a comparison of my changes using audio quantization and my new configurations on the second train and give a further understanding about speech research while trying to get the lowest possible WER.


 * Results: TBD


 * Plan: TBD


 * Concerns: Not to divulge any pertinent information about Team Stark operations while simultaneously working with the modeling group on Speech research. My main concern is to keep in touch with group members and let them know that I am conducting this research as a contribution to both Team Stark and the Capstone Project.  My participation with Modeling group is now limited due to the competition.

4/12/2016


 * Task: Read emails and check in with Stark team members. Check on 300 hr train on Idefix.


 * Results: TBD


 * Plan: TBD


 * Concerns: Not to divulge any pertinent information about Team Stark operations while simultaneously working with the modeling group on Speech research. My main concern is to keep in touch with group members and let them know that I am conducting this research as a contribution to both Team Stark and the Capstone Project.  My participation with Modeling group is now limited due to the competition.

Week Ending April 19, 2016
4/16/2016


 * Task: Finish decode and scoring on train.


 * Results: Secret


 * Plan: Secret


 * Concerns: Not to divulge any pertinent information about Team Stark operations while simultaneously working with the modeling group on Speech research. My main concern is to keep in touch with group members and let them know that I am conducting this research as a contribution to both Team Stark and the Capstone Project.  My participation with Modeling group is now limited due to the competition.

4/17/2016


 * Task: Run trains on seen/unseen data for team Stark.


 * Results: Secret


 * Plan: Secret


 * Concerns: Not to divulge any pertinent information about Team Stark operations while simultaneously working with the modeling group on Speech research. My main concern is to keep in touch with group members and let them know that I am conducting this research as a contribution to both Team Stark and the Capstone Project.  My participation with Modeling group is now limited due to the competition.

Week Ending April 26, 2016
4/22/2016


 * Task: Conduct Google Hangouts meeting with Team Stark


 * Results: Met with other members over the weekend via Google Hangouts.


 * Plan: Come up with new plans to single out independent variables per Ben. Try to coordinate server time with the group.  Try to find out about different Sphinx configurations.


 * Concerns: None at this time other than winning the competition for the lowest WER.

4/23/2016


 * Task: Conduct Google Hangouts meeting with Team Stark


 * Results: Try to configure sphinx with some new ideas pointed out by Ben from his online source material.


 * Plan: Run a train using the secret configurations.


 * Concerns: None at this time.

Week Ending May 3, 2016
5/1/2016
 * Task: Work on report for team stark and modeling group section.


 * Results: Go over reports from experiments in the team stark section.


 * Plan: Continue meeting at capstone class and working with group members and team members on what needs to get done and on what to add in the reports.


 * Concerns: making deadlines before semester ends.

5/3/2016


 * Task: Work on report for team stark and modeling group section.

Areas of Focus [edit]Documentation The modeling group documentation consisted of the Modeling Group Wiki Log, Experiment log and seperate individual group members logs ( Ryan, James, JonBen). Further documentation consisted of tutorials for running a train experiment, decoding and scoring. They required documenting Linux commands and Sphinx configurations in the Wiki tutorials in order to give users a better focus on how to modify their own experiments and compare results from documented experiments.
 * Results: I contributed this part of the report this week


 * Plan: Continue meeting at capstone class and working with group members and team members on what needs to get done and on what to add in the reports.


 * Concerns: making deadlines before semester ends.

Week Ending May 10, 2016
5/8/16


 * Task: Go over emails from other group members.  Look at any changes made to final report.


 * Results: Reading all emails it has been determined that the project report will be due for the last Capstone class. The final version is due before end of semester.


 * Plan: Working on both reports for Team Stark and Modeling Group.


 * Concerns: None at this time.

5/11/16


 * Task: Attend the last meeting of Capstone and continue working on the final report.  Go over anything else related to finishing up the semester.


 * Results: TBD


 * Plan: Working on report. Filing out any other paperwork.


 * Concerns: None at this time.