Fangjun Kuang

Thanks! I thought websocket was a byte stream based protocol, so I need to send the length to figure out the boundary, though the length and the payload can be...

Thanks! The link is very helpful.

> Here are the logs at the time of failure. Takes about 40 seconds to fail Does the issue still exist in the latest master?

I am not saying endpointing is not needed. We can keep it open, I suggest. I was asking if it still has memory issues for long waves in the case...

Thanks for the issue. Yes, I think we should provide a C/C++ runtime without depending on Python so that we can deploy sherpa on embedded devices, such as mobile phones....

> Is there any new progress? @minmin0 Please see See https://github.com/k2-fsa/sherpa/pull/95

For the endpointing stuff, can we implement it by counting the number of contiguous of frames that are decoded to blanks?

> Maybe we can make it more robust by adding a state machine with two states (one for silence, one for non-silence). I think we can add an attribute `num_trailing_silence_frames`...

> But that line of stream.hyp slicing seems to get all previous words starting the context_size position? We are getting the final results in that line. `[context_size:]` means to skip...

https://github.com/k2-fsa/sherpa/blob/d958043f1680bdef666a97699d2aaa9fcd4c91fa/sherpa/csrc/rnnt_beam_search.cc#L73-L77 We are indeed only using the last `context_size` tokens to predict the next token. But think of the very beginning. We need to add `context_size` blank tokens at the...