This repository provides a set of tools for generating training data for speech enhancement, speech separation, multi-talker speech recognition, and speaker diarization by simulation. Simulated development and evaluation sets can also be created.
Run the following commands to set up an environment.
conda env create -f environment.yml # Create a conda environment with all dependencies, except for pyrirgen, installed.
conda activate jsalt2020_simulate
./install.sh <your-data-dir> # Specify the location where you want the data to be stored.
source ./path.sh # This is created by install.sh.
./download.sh # Download the clean subset of LibriSpeech.
-
The following Python packages need to be installed in advance.
- webrtcvad
- PySoundFile
- resampy
- pyrirgen
-
Create path.sh with the following line.
export EXPROOT=<your-data-dir> -
Download and untar the LibriSpeech corpus from http://www.openslr.org/12/, untar them, which can be performed with the following script.
./download.sh
See EXAMPLES.md.
See TODO.md.
Please report problems with the GitHub issues of this repository. For discussions, please post your feedback to the workshop's Slack channel.