Skip to content

README.md

Latest commit

 

History

History
87 lines (67 loc) · 3.96 KB

README.md

File metadata and controls

87 lines (67 loc) · 3.96 KB
Raw

Data preparation

All URLs verified live 2026-07-21. Prereqs: repo env (python 3.9 + requirements.txt) plus sox and ffmpeg binaries on PATH (Scaper needs both).

Target layout (paths hardcoded in configs/*.json):

data/
├── MixLibriSpeech/
│   ├── LibriSpeech/                        # step 1 (SPEAKERS.TXT is required)
│   ├── librispeech_scaper_fmt/             # step 2
│   ├── wham_noise/{tr,cv,tt}/              # step 3
│   ├── jams/                               # step 4
│   ├── librispeech_dvector_embeddings/     # step 5
│   └── CIPIC/                              # step 6: 45 *.sofa (split lists: in repo)
├── RRBRIR/                                 # step 6: 5 *.sofa (split lists: in repo)
├── ASH-Listening-Set-8.0/BRIRs/            # step 6
└── CATT_RIRs/Binaural/16k/                 # step 6

1. LibriSpeech

mkdir -p data/MixLibriSpeech && cd data/MixLibriSpeech      # NOT .../LibriSpeech —
wget https://www.openslr.org/resources/12/train-clean-360.tar.gz   # tarballs already
wget https://www.openslr.org/resources/12/dev-clean.tar.gz         # contain LibriSpeech/
wget https://www.openslr.org/resources/12/test-clean.tar.gz
for f in *.tar.gz; do tar xzf "$f"; done
cd -

2. Scaper format (symlinks)

# script's dset list also includes train-clean-100 (unused by configs);
# create a placeholder so the loop doesn't crash, or edit the list in the script
mkdir -p data/MixLibriSpeech/LibriSpeech/train-clean-100
python -m src.datasets.librispeech2scaper --root_dir data/MixLibriSpeech

Asserts per-speaker output dirs don't exist — delete librispeech_scaper_fmt/<subset> before re-running a subset.

3. WHAM! noise (16 kHz)

cd data/MixLibriSpeech
# old storage.googleapis.com URL is dead; this is the current mirror from wham.whisper.ai
curl -L -O https://my-bucket-a8b4b49c25c811ee9a7e8bba05fa24c7.s3.amazonaws.com/wham_noise.zip  # 18.2 GB
unzip wham_noise.zip    # -> wham_noise/{tr:20000, cv:5000, tt:3000} wavs, 37.7 GB
cd -

4. Generate jams

python -m src.datasets.generate_jams    # from repo root (relative paths hardcoded)

test/dev/train = 10k/5k/100k specs, seeds 1/2/42, WHAM labels tt/cv/tr. Writes specs only (write_audio=False); audio is rendered on the fly during training.

5. Speaker embeddings

python -m src.datasets.dvector_embeddings   # defaults match the layout above

Processes every subset under librispeech_scaper_fmt/; asserts output dirs don't exist.

6. HRTFs / BRIRs

Set Source Notes
CIPIC sofacoustics.org/data/database/cipic all 45 subject_*.sofa (82 MB); SOFA format, not UC Davis .mat
RRBRIR IoSR-Surrey/RealRoomBRIRs the 5 UniS_*_BRIR_16k.sofa files (47 MB)
ASH ShanonPearce/ASH-Listening-Set v8.0 ASH-Listening-Set-8.0.zip (230 MB); unzip at data/; covers every room hardcoded in multi_ch_simulator.py
CATT iosr.uk/software CATT_RIRs.zip (71 MB); unzip at data/ — its CATT_RIRs/Binaural/16k/<room>/ layout matches the configs as-is (407 wavs, rooms = RT60 0_0s…1_0s, azimuths ±90°/5°) 
cd data/MixLibriSpeech/CIPIC && curl -s https://sofacoustics.org/data/database/cipic/ \
  | grep -oE 'subject_[0-9]+\.sofa' | sort -u \
  | xargs -I{} wget -q https://sofacoustics.org/data/database/cipic/{} && cd -

Split lists: the CIPIC/RRBRIR {train,val,test}_hrtf.txt files (one SOFA filename per line, resolved relative to the txt's dir — multi_ch_simulator.py:31-35) are committed in this repo: CIPIC 31/5/9 subjects; RRBRIR train={Room_D,Room_B}, val={Room_C}, test={Room_A}. ASH/CATT need no lists — their room splits are hardcoded in multi_ch_simulator.py.