Code

Recipe for African Accented French #2813

Code

Open

johnjosephmorgan wants to merge 36 commits intokaldi-asr:masterkaldi-asr/kaldi:masterfrom johnjosephmorgan:african_french_recipejohnjosephmorgan/kaldi:african_french_recipeCopy head branch name to clipboard

+8,281

Commits

Commits on Nov 1, 2018

• Initial commit.

  johnjosephmorgan

  committed
  be8c443

  Copy full SHA for be8c443

Commits on Nov 2, 2018

• Moved results to RESULTS file.

  johnjosephmorgan

  committed
  a4204c9

  Copy full SHA for a4204c9

• Corrected name of transcript file for yaounde answers training data.

  johnjosephmorgan

  committed
  845bd47

  Copy full SHA for 845bd47

• I had made some changes to the corpus transcripts that are not in the version on openslr.org. My scripts were using those new files. I changed my scripts to use the files that are currently on open…

  Show description for 078889d
  johnjosephmorgan

  committed
  078889d

  Copy full SHA for 078889d

• The transcript file for the yaounde answers that is currently on openslr.org is bad. The transcripts are actually the questions instead of the answers. I am including a temporary transcription that…

  Show description for 413f0bc
  johnjosephmorgan

  committed
  413f0bc

  Copy full SHA for 413f0bc

Commits on Nov 20, 2018

• Add decoding with large lm rescoring for monophones?

  johnjosephmorgan

  committed
  055ca6e

  Copy full SHA for 055ca6e

• Fixed bug. eq instead of == in string comparision.

  johnjosephmorgan

  committed
  dd49f32

  Copy full SHA for dd49f32

• Changed comments to echoes.

  johnjosephmorgan

  committed
  8f23a41

  Copy full SHA for 8f23a41

• Do not exit after phonetisaurus align ? output.

  johnjosephmorgan

  committed
  439dfbd

  Copy full SHA for 439dfbd

Commits on Nov 21, 2018

• Experiment with lower bottleneck dimension. Lowered from 128 to 96.

  johnjosephmorgan

  committed
  7fe5495

  Copy full SHA for 7fe5495

• Merge.

  johnjosephmorgan

  committed
  071d7f4

  Copy full SHA for 071d7f4

• Lower l2 regularization from 0.03 to 0.02.

  johnjosephmorgan

  committed
  75e3846

  Copy full SHA for 75e3846

Commits on Nov 26, 2018

• Added clauses to stage conditionals to allow small lm decoding exclusively.

  johnjosephmorgan

  committed
  688f768

  Copy full SHA for 688f768

• Added clauses to stage conditionals to allow small lm decoding exclusively. Also removed mllt lda tri2b steps.

  johnjosephmorgan

  committed
  588d5b7

  Copy full SHA for 588d5b7

Commits on Nov 27, 2018

• More fixing stage conditionals for small lm decoding only.

  johnjosephmorgan

  committed
  3efc399

  Copy full SHA for 3efc399

Commits on Nov 28, 2018

• Inserted a wait before decoding tri3b models with larger lms. The decoding depends on previous decoding?

  johnjosephmorgan

  committed
  35f23b5

  Copy full SHA for 35f23b5

• Inserted a wait before decoding with larger LMs and enhanced lexicon.

  johnjosephmorgan

  committed
  8230717

  Copy full SHA for 8230717

• added another tuning script for trying to improve the chain model WER scores.

  johnjosephmorgan

  committed
  1ee6068

  Copy full SHA for 1ee6068

• Updated experiment script.

  johnjosephmorgan

  committed
  77436d8

  Copy full SHA for 77436d8

• Updated experiment script.

  johnjosephmorgan

  committed
  76e9ed6

  Copy full SHA for 76e9ed6

• Updated experiment script.

  johnjosephmorgan

  committed
  b45cf38

  Copy full SHA for b45cf38

Commits on Nov 29, 2018

• Added ascript to experiment with number of epochs.

  johnjosephmorgan

  committed
  0786dfc

  Copy full SHA for 0786dfc

• UUpdate experiment script.

  johnjosephmorgan

  committed
  38115e6

  Copy full SHA for 38115e6

• Corrected comparison. I ran it too early. Fewer epochs gave better results.

  johnjosephmorgan

  committed
  c615a70

  Copy full SHA for c615a70

• Added current results. They look bad. Why are they so bad?

  johnjosephmorgan

  committed
  a485acf

  Copy full SHA for a485acf

• Minor cleaning.

  johnjosephmorgan

  committed
  569eda1

  Copy full SHA for 569eda1

Commits on Nov 30, 2018

• Put tri2b models back.

  johnjosephmorgan

  committed
  7d5b1e3

  Copy full SHA for 7d5b1e3

Commits on Dec 1, 2018

• Adding Yenda's scripts for kws.

  johnjosephmorgan

  committed
  682f49c

  Copy full SHA for 682f49c

• Adding a script to do l2 regularization tuning experiments.

  johnjosephmorgan

  committed
  e1fe905

  Copy full SHA for e1fe905

Commits on Dec 3, 2018

• Adjusted number of leaves and gaussians after experimenting.

  johnjosephmorgan

  committed
  f49a67b

  Copy full SHA for f49a67b

• Add pruned lm modeling.

  johnjosephmorgan

  committed
  652f04b

  Copy full SHA for 652f04b

Commits on Jan 7, 2019

• I am simplifying the directory and file names in this recipe.

  johnjosephmorgan

  committed
  0f1824c

  Copy full SHA for 0f1824c

Commits on Mar 6, 2019

• Simplifying.

  johnjosephmorgan

  committed
  c2946e5

  Copy full SHA for c2946e5

• Update.

  johnjosephmorgan

  committed
  1081fbd

  Copy full SHA for 1081fbd

• Update.

  johnjosephmorgan

  committed
  ee5726a

  Copy full SHA for ee5726a

Commits on Mar 7, 2019

• No small medium or large LMs.

  johnjosephmorgan

  committed
  334ab41

  Copy full SHA for 334ab41