Reveberant LibriCHiME-5

This repository contains instructions for creating the reverberant LibriCHiME-5 dataset (audio files) for the UDASE task of the CHiME-7 challenge.

The code to generate the metadata files is available in this repository.

Preparation

You must first:

• Follow the instructions in this repository to extract the audio segments from the original CHiME-5 data.

• Download the LibriSpeech dev-clean and test-clean data. Put the data in a folder with the following structure:

  ├── dev-clean
  │   ├── 1272
  │   │   ├── 128104
  │   │   │   ├── 1272-128104-0000.flac
  │   │   │   ├── ...
  ├── test-clean
  │   ├── 1089
  │   │   ├── 134686
  │   │   │   ├── 1089-134686-0000.flac
  │   │   │   ├── ...
  

• Download the VoiceHome dataset.

Installation

# clone repository
git clone https://github.com/UDASE-CHiME2023/reverberant-LibriCHiME-5.git
cd reverberant-LibriCHiME-5

# activate CHiME environment
conda activate CHiME

Dataset creation

• Set the paths in paths.py:

  • udase_chime_5_audio_path is the path to the audio segments that you should have previously extracted from the CHiME-5 data (see preparation section above).
  • librispeech_path is the path to the LibriSpeech dataset that you should have previously downloaded (see preparation section above).
  • voicehome_path is the path to the VoiceHome dataset that you should have previously downloaded (see preparation section above).
  • reverberant_librichime_5_json_path is the path to the metadata of the reverberant LibriCHiME-5 dataset, you do not need to change it.
  • reverberant_librichime_5_audio_path is the path where you want to store the reverberant LibriCHiME-5 dataset.

• Run

  python create_audio_from_json.py --subset dev

  python create_audio_from_json.py --subset eval

For the dev and eval sets of the reverberant LibriCHiME-5 dataset, we have three subsets dependending on the maximum number of simultaneously-active speakers (1, 2 or 3). These subsets are stored in separate subfolders whose name indicates the maximum number of simultaneously-active speakers.

At the path defined by the variable reverberant_librichime_5_audio_path in path.py, you should obtain:

├── dev
│   ├── 1 (3 561 files)
│   │   ├── [...]_mix.wav
│   │   ├── [...]_noise.wav
│   │   ├── [...]_speech.wav
│   ├── 2 (1 695 files)
│   │   ├── [...]_mix.wav
│   │   ├── [...]_noise.wav
│   │   ├── [...]_speech.wav
│   ├── 3 (195 files)
│   │   ├── [...]_mix.wav
│   │   ├── [...]_noise.wav
│   │   ├── [...]_speech.wav
├── eval
│   ├── 1 (4 182 files)
│   │   ├── [...]_mix.wav
│   │   ├── [...]_noise.wav
│   │   ├── [...]_speech.wav
│   ├── 2 (1 482 files)
│   │   ├── [...]_mix.wav
│   │   ├── [...]_noise.wav
│   │   ├── [...]_speech.wav
│   ├── 3 (192 files)
│   │   ├── [...]_mix.wav
│   │   ├── [...]_noise.wav
│   │   ├── [...]_speech.wav

For each mixture in the dataset:

• [...]_mix.wav is the speech+noise mixture;
• [...]_speech.wav is the reference speech signal;
• [...]_noise.wav is the reference noise signal.

Summing the reference speech and noise signals gives the mixture signal.

As the original CHiME-5 recordings, these audio signals are not normalized.