This is a course project for SNA by Igor Fedorov and Evgeny Shabalin.
The speaker embedder is adopted from CLOVA and we use VoxCeleb-1 for household simulation.
Our graph-based algorithm is an adaptation of Links for given scenario. Precisely, we discard the node-splitting part and operate only with enrichment & merging.
Our algorithm outperforms simple enroll enrichment in terms of average EER:
(0.272 - 0.196 - 0.175 for n=2, 0.415 - 0.387 - 0.299 for n=4, 0.565 - 0.455 - 0.401 for n=6 after 1000 simulations)
Implementation is mainly written in embs.ipynb