Fangjun Kuang

@danpovey Do you have time to design one?

https://github.com/k2-fsa/sherpa#rtf-test Can you use the above [colab notebook](https://colab.research.google.com/drive/1JX5Ph2onYm1ZjNP_94eGqZ-DIRMLlIca?usp=sharing) to measure RTF for CPU. (The default is using GPU). A PR is welcome if you can do that. Thanks

> @csukuangfj Actually I was about to create an issue to tackle it (to also have an RTF for streaming model, etc) :) That would be very nice of you....

Thanks!

> @csukuangfj Also, I was thinking to maybe add the RTF in Results.md in `icefall`, to give a sense of how much CPU/GPU usage it needs (so not only comparing...

> @csukuangfj I tried to add the RTF for the streaming pipeline. However, I keep having `sent 1011 (unexpected error) keepalive ping timeout; no close frame received` any idea why...

It is doable. We can output token force alignment information with icefall. See https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/transducer_stateless For the confidence information, you have to change the code to return the scores.

Do you want to return the alignment information in frame numbers or in seconds? Also, do you want the origin to be relative to the start of a segment or...

@EmreOzkose Great work! Thanks! For > 1. change ActivationBalancer() forward > 2. change DoubleSwish() forward: I recommend using ```python3 if torch.jit.is_scripting() or torch.onnx.is_in_onnx_export(): return x ``` (See https://pytorch.org/docs/stable/onnx.html#torch.onnx.is_in_onnx_export) For >...

@EmreOzkose Please have a look at https://github.com/k2-fsa/icefall/pull/501