TSE-with-ref-selection

target speaker extraction with speaker reference selection

Dependencies

pip install -r requirements.txt

Prepare Dataset

git clone https://github.com/JorisCos/LibriMix.git

Reference Selection

Method	ACC(ov=30%)
Speaker Diarization and use longest segments	86.0
Split into 5s chunks and clustering	79.5
Overlap detection and clustering with 5s sliding window	99.5
Split into regions, overlap detection and clustering with regions	89.2

Speaker Extraction

refer to SPEX+

# train
./train.sh

# evaluation
python cse_test.py

Method	SI-SDR(model=x8515, ov=30%)	SI-SDR（model=xmax, ov=30%,use ground truth segmentation)
Speaker Diarization and use longest segments	8.77	13.62
Split into 5s chunks and clustering	8.75	14.00
Overlap detection and clustering with 5s sliding window	8.89	14.32
Split into regions, overlap detection and clustering with regions	9.11	14.44