Embeddings do not match original implementation

[dcmoura]

Hey! Thank you for porting OpenL3 to PyTorch! :-)

I run the following commands on wav files of different sample rates (8KHz and 48KHz).
The embeddings of this port do not match the embeddings of the original implementation.
Here is an example with one of your sample files.

Original OpenL3:

$ openl3 --suffix original --content-type env --audio-embedding-size 512 --input-repr mel128 audio chirp_1s.wav

(This) Torch port:

$ torchopenl3 --suffix torch --content-type env --audio-embedding-size 512 --input-repr mel128 chirp_1s.wav

MAE calculation:

$ python3 -c "import numpy as np; mae = np.abs(np.load('chirp_1s_original.npz')['embedding'] - np.load('chirp_1s_torch.npz')['embedding']).mean(); print(mae)"

1.3379873

Tried Python 3.7 and 3.8 on a AWS EC2 g5.2xlarge instance running Ubuntu 22.04.3 LTS.