Fix relative attention for streaming inference

hirofumi0810 / neural_sp

End-to-end ASR/LM implementation with PyTorch

Apache License 2.0

593 stars 140 forks source link

Fix relative attention for streaming inference #250

Closed hirofumi0810 closed 3 years ago

hirofumi0810 commented 3 years ago

This is for streaming Transformer/Conformer encoder when using relative positional encoding. Related to https://github.com/espnet/espnet/pull/2816. Confirmed the same results when encoding incrementally.