Streaming setup in the whisper tokenizer

[zsLin177]

Hello,

Thanks for the nice work! I noticed in your paper that the Whisper encoder supports "Causality for Streaming Inference." However, in the currently released model, I observed that both encoder_causal_attention and quantize_causal_encoder are set to false in the tokenizer's configuration. Whether the current version of the tokenizer supports streaming input for speech encoding?

Additionally, could you explain the impact of setting quantize_causal_block_size to 200?

Thank you for your time and assistance!