Fix the bug of Embedding norm and update the ERR

HarryVolek / PyTorch_Speaker_Verification

PyTorch implementation of "Generalized End-to-End Loss for Speaker Verification" by Wan, Li et al.

BSD 3-Clause "New" or "Revised" License

576 stars 164 forks source link

Fix the bug of Embedding norm and update the ERR #30

Closed mazzzystar closed 5 years ago

mazzzystar commented 5 years ago

As #27 mentioned, your implementation is a little different with the original paper, and I experimented on TIMIT with L2 norm being calculated per embedding as said in paper. With no other changes to your code, the ERR across 10 epochs decreases from 0.0518 to 0.0377.

mazzzystar commented 5 years ago

Only the first two commit matters.

c6fa12e   Update speech_embedder_net.py
26bd21a  Fix bug and update ERR

Ignore the third commit, it was intended for creating Embedding without webRTC VAD. I'm glad if you could check my implementation to see if correct.

HarryVolek commented 5 years ago

Thanks for the contribution mazzzystar. I merged the first 2 commits. I can take a look at the third when I get a chance.