D-vector trained on VoxCeleb1

This release is to address the module loading issue after upgrading torchaudio to 0.8.0

Pretrained models

The model was trained on VoxCeleb1 dataset.

Model details:

• 40-dim log mel spectrogram as input
• 3-layer LSTM with hidden dimensions being 256
• 256-dim attentive pooled speaker embedding

Training details:

• 64 speakers, 10 utterances per speaker in a batch
• 250K steps