Speech:Spring 2016 Modeling Group

From Openitware
Jump to: navigation, search


Groups


Group Member Logs

Modeling Group Schedule:

Members Monday Tuesday Wednesday Thursday Friday Saturday Sunday
Ben Leith 9:00am - 5:30pm 9:00am - 5:30pm 2:00pm - 5:00pm and 5:30pm - 8:30pm 9:00am - 5:30pm 9:00am - 5:30pm off off
Ryan O'Neal 11:00am- 5:00pm 9:00am - 4:00pm 9:00am - 12:00pm and 2:00pm - 8:30pm 9:00am - 4:30pm 9:00am - 4:30pm off off
Jon Shallow off 1:00pm - 3:00pm and 5:30pm - 8:30pm 9:00am - 12:00pm and 2:00pm - 5:00pm 5:30pm - 8:30pm off Drill weekend once a month. On call 24/7 basis with NHARNG. Drill weekend once a month. On call 24/7 basis with NHARNG.
James Schumacher 4:15am - 9:00am 12:00pm - 3:00pm and 5:30pm - 8:30pm 9:00am - 12:00pm and 2:00pm - 5:00pm off 4:00am - 9:00am off off

04 FEB 16

Investigate an error in the decode process of 0281 sub-experiment 005: link here

  1. Error details:
    1. Segmentation fault (core dumped)
  2. Investigation
    1. Read logs in /mnt/main/Exp/0281/005/DECODE/decode.log
      1. Noticed several warnings:
        1. Line 256: WARNING: "cont_mgau.c", line 667: Weight normalization failed for 9 senones
        2. Line 260: WARNING: "cont_mgau.c", line 781: 73 densities removed (9 mixtures removed entirely)
        3. Line 265: WARNING: "tmat.c", line 192: Normalization failed for tmat 0 from state 0
        4. Line 266: WARNING: "tmat.c", line 192: Normalization failed for tmat 0 from state 1
        5. Line 267: WARNING: "tmat.c", line 192: Normalization failed for tmat 0 from state 2
        6. Line 268: WARNING: "tmat.c", line 192: Normalization failed for tmat 1 from state 0
        7. Line 269: WARNING: "tmat.c", line 192: Normalization failed for tmat 1 from state 1
        8. Line 270: WARNING: "tmat.c", line 192: Normalization failed for tmat 1 from state 2
        9. Line 271: WARNING: "tmat.c", line 192: Normalization failed for tmat 2 from state 0
        10. Line 272: WARNING: "tmat.c", line 192: Normalization failed for tmat 2 from state 1
        11. Line 273: WARNING: "tmat.c", line 192: Normalization failed for tmat 2 from state 2
      2. As well as several errors:
        1. Line 304: ERROR: "wid.c", line 282: <UNK> is not a word in dictionary and it is not a class tag.
        2. Line 305: ERROR: "wid.c", line 282: <B_ASIDE> is not a word in dictionary and it is not a class tag.
        3. Line 306: ERROR: "wid.c", line 282: <E_ASIDE> is not a word in dictionary and it is not a class tag.
      3. And this final system_error at the end of the log:
        1. Line 403: SYSTEM_ERROR: "corpus.c", line 562: fopen(/mnt/main/Exp/0281/005/etc/005_decode.fileids,r) failed; No such file or directory
        2. error is the corpus.c file at line 562. It attempts to open "/mnt/main/Exp/0281/005/etc/005_decode.fileids" with read permissions.
        3. This is where we believe the primary error is at.
    2. . Checked /mnt/main/Exp/0281/005/etc for the 005_decode.fileids file
      1. File does not exist in directory, as expected. However, there is a 001_decode.fileids.
        1. We believe this was user error and is due to not correctly matching the sub_experiment number in the head -1000 001_train.fileids > 001_decode.fileids command from the tutorial.
        2. We believe if the file is renamed to 005_decode.fileids we can re-run the decode and expect to be able to score it successfully.
  3. Fix
    1. Unable to fix at this at time
      1. The tutorial says not to run any commands from root. Ben is the only person with a user id as of now, and he is at work. Ben will attempt our proposed fix when he is free.

10 Feb 16

Ran the decode on the 0283/002 experiment. Once that finished, we ran the scoring tool. Check out this link for information on the results for this sub experiment.

We came up with the following tasks to perform:

  1. Duplicate best result from the Spring 2015 semester.
    1. We'll use a 125 Hr train instead of a 256 Hr train to conserve time.
    2. We'll also use the same configuration settings.
    3. Ultmately, we'll generate a baseline that we can refer to.
  2. Run another train with the density set to 128, other settings the same as the baseline
  3. Run another train with the convergence ratio set to 0.001, other settings the same as the baseline
  4. Run another train with the variance normalization set to yes
  5. Run the other combinations and determine the best result. That result will become the next baseline.


Assigned machine is: TBD

17 Feb 16

Task: Score the decode in 0238/003. Prepare the hypothesis transcript and use SCLite to score the decode. Start a new train 0283/004. Configure the sphinx_train.cfg for variance normalization to yes, senomes to 8000, convergence ratio to 0.001 and change density 64.

Concerns: Hope that this will reduce real time factor and error rate.

Results: Train finished on 2/20. Started decode on 2/20. Decode took 5 hours and 24 minutes. Analyzing the results indicates that the Error rate is 35.2% which did not change very much from the last Error rate which is 36.3%.

2 Mar 16

We decided to create a new corpus (fixed_30k) to train on after the Data group discovered that there were discrepancies between the audio utterances and the transcripts after 32,000 utterances.

Steps:

  • Create new corpus directory structure
   cd /mnt/main/corpus/switchboard
   mkdir fixed_30k
   cd fixed_30k
   mkdir test
   mkdir train
   cd train
   mkdir audio
   mkdir info
   mkdir trans
   cd audio
   mkdir conv
   mkdir utt
  • Copy the first 247 files from
   /mnt/main/corpus/switchboard/256hr/train/audio/conv to /mnt/main/corpus/switchboard/fixed_30k/train/audio/conv
   cd /mnt/main/corpus/switchboard/256hr/train/audio/conv
   ls | head -247 | xargs cp -t /mnt/main/corpus/switchboard/fixed_30k/train/audio/conv
  • Copy the first 32000 audio utterances from
   /mnt/main/corpus/switchboard/256hr/train/audio/utt to /mnt/main/corpus/switchboard/fixed_30k/train/audio/utt
   cd /mnt/main/corpus/switchboard/256hr/train/audio/utt
   ls | head -32000 | xargs cp -t /mnt/main/corpus/switchboard/fixed_30k/train/audio/utt
  • Copy the first 32000 lines in the transcript from
   /mnt/main/corpus/switchboard/256hr/train/train/train.trans to /mnt/main/corpus/switchboard/fixed_30k/train/train.trans
   cd /mnt/main/corpus/switchboard/256hr/train/trans
   head -32000 train.trans > /mnt/main/corpus/switchboard/fixed_30k/train/train.trans

Currently, this corpus isn't playing nice with the prepareTrainExperiment perl script.