Materials for challenge participants

The following are materials that you should obtain in order to participate in the challenge. Please download corresponding items.

Data & Task files (common for SE and ASR tasks)

• Data (i.e. dev, eval and training data)
  • Step 1: Obtain WSJCAM0 and Multi-Channel WSJ from LDC.
  • Step 2: Use a shell script (prepare_reverb_data.sh) in this zip file to reorganize downloaded WSJCAM0 and MCWSJ into an appropriate directory structure (which is also compatible with the Kaldi recipe) assumed in the challenge. Note that the shell script requires Matlab, SoX（Sound eXchange) to decode .flac files, and an internet connection to collect necessary items.
• Task files
  • If you are using Kaldi, the task files will be downloaded in the Kaldi recipe. So you do not have to download the task files by yourself from here.
  • If you are NOT using Kaldi (doing only speech enhancement evaluation or doing ASR evaluation but with your own ASR system), please prepare the task files as follows.
  • Task files for the dev set: They are included in the evaluation tools of both SE and ASR tasks
  • Task files for the eval set: Please download them from here
  Please copy the content of taskFiles_et.tgz to the directory containing the task files (i.e. "/taskfiles" for SE tool, and "/taskFiles" for the ASR tool).
  Please read the following readme files for instructions about how to use the evaluation task files,
  • readme_se.txt
  • readme_asr.txt

Evaluation tools for SE/ASR task

• For ASR task
  • Baseline ASR system based on Kaldi (new)
  • (Original HTK-based baseline system, maybe obsolete) Evaluation tool and Baseline ASR system developed by the organizers in 2014. (ver.2.0)
  • Result table format for ASR task and instructions to fill the table
• For Enhancement task
  • Evaluation tool
  • Result table format for SE task and instructions to fill the table
  • There is a memory leak in the SE evaluation script that causes "out of memory" errors on some machines. If you are facing such issues, you can use the SRMR patch.
    

If you have any questions regarding the challenge data and baseline system, please contact us.