AlexWu/tseomni

TSE-Omni — Test-set Data

Supporting data for TSE-Omni. This dataset repo holds the test-set scp lists and the FunCodec target features used at inference/evaluation time. It does not contain model weights, nor the raw VoxCeleb2 / LRS3 / YGD audio (which you must obtain from their original sources).

Contents

dump/wavs/<testset>/{mix_clean,aux_s1,s1}.scp   # plain-text lists (uncompressed)
funcodec.tar.gz                                 # FunCodec .npy features (~154 MB)

Test sets: vox2_en_2spk_test_list-scale, lrs3_en_2spk_test_list, lrs3_en_2spk_test_verylong_list, ygd_en_2spk_test_list.

dump/ — scp lists

Two-column Kaldi-style scp: <utt_id> <path>. The paths use placeholders for the raw corpora you provide locally:

placeholder	replace with
<VOX2_ROOT>	your VoxCeleb2 mixture / audio_clean root
<LRS3_ROOT>	your LRS3 mixture / wav root
<YGD_ROOT>	your YGD data root

Example (replace placeholders to match your machine):

sed -i 's#<VOX2_ROOT>#/your/path/to/voxceleb2#g' dump/wavs/*/*.scp

funcodec.tar.gz — target features

FunCodec-encoded target-speech codec (funcodec/<set>/test*/0/*.npy) plus the visual/gesture feature lists. Extract into the code repo's data/ directory:

tar xzf funcodec.tar.gz -C /path/to/tse-omni-code/data/

This yields data/funcodec/..., which matches the default paths in recipes/inference.sh. The .npy filenames correspond 1:1 to the mix_clean.scp entries (verified: 2000/2000 for vox2, etc.). The scp files inside the archive use relative funcodec/... paths; visual-feature lists for YGD use the <YGD_ROOT> placeholder.

Usage with the code

# 1) get scp lists
git clone https://huggingface.co/datasets/AlexWu/tseomni
cp -r tseomni/dump  /path/to/tse-omni-code/data/
# 2) get FunCodec features
tar xzf tseomni/funcodec.tar.gz -C /path/to/tse-omni-code/data/
# 3) point the scp placeholders at your local corpora (see above), then run inference