[feat](ZeroPrompt): illustrative exp of zeroprompt, not production ready

[xingchensong]

This PR is an illustrative example to showcase the efficiency of ZeroPrompt and calculate metrics used in the paper. Due to the addition of some input parameters and significant changes to the forward_chunk API, it's better not to merge this PR and leave it as a reference for others who are interested in ZeroPrompt.

• Paper link: https://arxiv.org/pdf/2305.10649.pdf

• Presentation (In Chinese): 【语音识别延迟优化三部曲之二：ZeroPrompt】 https://www.bilibili.com/video/BV1Cj411z73S/?share_source=copy_web&vd_source=fd63a2f5ad1c9bc6b6658d9a922e20aa

• example output (Aishell-1, 640ms chunksize + 640ms zeroprompt):

  • We can see that using ZeroPrompt allows us to consistently gain extra words at almost every chunk, The First Token Display Time & Last Token Display Time have been advanced by 640ms (one chunk).
  • It’s reasonable for the predictions from the first ZeroPrompt chunk to be inaccurate due to the lack of contextual information. However, this is not a significant issue since most of the errors are homophones of the correct counterparts.
  • Additionally, these errors will be quickly corrected by the Prompt-and-Refine strategy.
    
    

• example output (librispeech, 640ms chunksize + 640ms zeroprompt):

  • We can draw the same conclusion as in Aishell-1 results