Embedding is mostly zero

[dodobyte]

I plot the embedding vector and it's mostly zeros. Is this expected?

I want to use the embedding in another project. I also plotted their example embeddings and those seem to be distributed significantly better.

• Embedding from this project

• Embeddings from AutoVC

And here's the test code;

from pathlib import Path
from resemblyzer import VoiceEncoder, preprocess_wav
import numpy as np, matplotlib.pyplot as plt

wav = preprocess_wav(Path("367-130732-0005.flac"))

encoder = VoiceEncoder()
embed = encoder.embed_utterance(wav)

plt.plot(embed, 'bo')
plt.show()

Am I doing something wrong? Thanks.

[CorentinJ]

Absolutely, the embeddings are sparse due to the relu at the end of the model. It doesn't make them worse, although I did remove that relu in the development branch I'm working on. Don't worry about it.