Skip to content

Latest commit

 

History

History
74 lines (57 loc) · 2.78 KB

README.md

File metadata and controls

74 lines (57 loc) · 2.78 KB
Raw

Wall Street journal experiment

To reproduce our Wall Street Journal (WSJ) experiments, please follow the instructions below. All the steps should be done at the kaldi WSJ recipe directory (or you can add symlinks to all the files (local, steps, utils, etc) as some people do). In order to perform steps 1, 2, 5 you should source path.sh file from the recipe before you sourced $LVSR/env.sh.

Note

Check that $LVSR environment variable points to this repository and $LD_LIBRARY_PATH includes a path to openfst library.

Steps

  1. Compile a Fuel-compatible dataset file in HDF5 format. This step requires kaldi and kaldi-python.

    First, run prepare data part from the WSJ recipe. You'll get all *.scp files which link waves and text.

    Then, run

    $LVSR/exp/wsj/write_hdf_dataset.sh

    The resulting file wsj.h5 should be put to $FUEL_DATA_PATH folder.

  2. Compile language model FST's from ARPA-format language models provided with WSJ. This step requires kaldi.

    $LVSR/exp/wsj/make_all_wsj_graphs.sh <lmfile> <lmsdir>

    where <lmfile> is the arpa languge model which goes with WSJ dataset. (we placed it to $FUEL_DATA_PATH/WSJ/lm_bg.arpa.gz) and <lmsdir> is a directry to place FST language models (we use data/lms).

  3. Train the model. You don't need kaldi for training and it doesn't use any scripts from the recipe.

    For the End-to-End Attention-based Large Vocabulary Speech Recognition use the model=wsj_paper7 and for the Task Loss Estimation for Sequence Prediction use model=wsj_reward6 in the following script:

    $LVSR/bin/run.py train <model> $LVSR/exp/wsj/configs/<model>.yaml

    This will start training and save the model to the folder <model>.

  4. Decode the model on the validation and training datasets.

    $LVSR/exp/wsj/decode.sh wsj_paper7 <part> <beam-size>

    or

    $LVSR/exp/wsj/decode_tle.sh wsj_reward6 <part> <beam-size>

    for the task loss estimation model.

    We typically use beam size 200 to get the best performance. However, with beam size the scores are typically only 10% worse and decoding is much faster. You can see reports at wsj_paper6/reports directory. For every language model option, beam size and subset there are alignment images and report.txt which contains transcripts, approximate CER and WER and other auxiliary information.

  5. Score the recognized transcripts:

    $LVSR/exp/wsj/score.sh <part> <model>/reports/valid_trigram_200/

    This script produces <part>-text.wer and <part>-text.errs files in the corresponding directory. The first one contains WER and the second one the report of compute-wer kadli script.