How streaminig evaluation is performed for Wav2Vec2.0 Model

[steventan0110]

Hi authors,

Thanks for releasing the code for your great work! After reading the paper, I am trying to replicate the simultaneous translation experiment using CIF's output as the pre-decision policy.

However, I have a question about the underlying model used -- Wav2Vec2.0. From my understanding, it requires the complete audio input to encode the feature because it uses normal (non-causal) convnet and tf-encoder and does not support streaming, so how is the inference performed for the wav2vec_cif model? From the paper, it seems that you didn't perform any additional training for Wav2Vec2 to make it support partial input during stream-like inference.

I would greatly appreciate it if you could provide additional details about how evaluation is performed for the streaming case!