At the moment this repository is still under construction

This repository includes the scripts required to create the sms_wsj database and a baseline ASR system using KALDI (http://github.com/kaldi-asr/kaldi).

If you are using this code please cite the following paper:

@Article{SmsWsj19,
  author    = {Lukas Drude, Jens Heitkaemper, Christoph Boeddeker, Reinhold Haeb-Umbach},
  title     = {{SMS-WSJ}: Database, performance measures, and baseline recipe for multi-channel source separation and recognition},
  year      = {2019},
}

Does not work with Windows.

Clone this Repo and install the package

$ git clone https://github.com/fgnt/sms_wsj.git
$ cd sms_wsj
$ pip install --user -e ./

Set your KALDI_ROOT

$ export KALDI_ROOT=/path/to/kaldi

We assume that the KALDI wsj baseline has been created with the run.sh script. To build the database the structures created during the first stage of the run.sh script are required. The ASR baseline uses the language models created during the same stage. Afterwards you can create the database:

$ make SMS_WSJ_DIR=/path/to/write/db/to WSJ_DIR=/path/to/wsj

If desired the number of parallel jobs may be specified using the additonal input num_jobs. Per default 16 parallel jobs are used.

The RIRs are downloaded by default, to generate them yourself see here.

Use the following command to train the baseline ASR model:

$ python -m sms_wsj.train_baseline_asr with egs_path=$KALDI_ROOT/egs/ json_path=/path/to/sms_wsj.json

The script has been tested with the KALDI Git hash "7637de77e0a77bf280bef9bf484e4f37c4eb9475"

A: The example ID is a unique identifier for an example (sometime also known as utterance ID). The example ID is a composition of the sperakers, the utterances and an scenario counter:

The python code in this repository requires python 3.6. However, Kaldi runs on python 2.7. To solve this mismatch Kaldi has to be forced to switch the python version using the path.sh. Therefore, add the follwing line to the ${KALDI_ROOT}/tools/envh.sh file:

export PATH=path/to/your/python2/bin/:${PATH}

To generate the RIRs you can run the following command:

$ mpiexec -np $(nproc --all) python -m sms_wsj.database.create_rirs with database_path=cache/rirs

The expected runtime will be around 1900/(ncpus - 1) hours. When you have access to an HPC system, you can replace mpiexec -np $(nproc --all) with an HPC command. It is enough, when each job has access to 2GB RAM.