FYI: RTF for the latest Zipformer transducer

[csukuangfj]

We just created a colab notebook to show the RTF (Real-time factor) of the latest zipformer transducer model.

We use sherpa-onnx for the CPU test. As for the GPU test, we infer RTF from the decoding logs for the librispeech test-clean and test-other dataset.

The results are summarized below:

Please see the following colab notebook for details.

https://github.com/k2-fsa/colab/blob/master/rtf_test_for_zipformer_transducer.ipynb

[armusc]

can I ask you which CPU exactly?

[csukuangfj]

can I ask you which CPU exactly?

It is the cpu provided by colab. Please see the colab notebook for details. We have provided the output of lscpu in the colab notebook.

[armusc]

thanks

[AdolfVonKleist]

@csukuangfj have you also looked at GPU streaming mode with sherpa-onnx? I'm curious if that might also be impacted by the streaming strategy, e.g., if you'd expect a fairly large difference in speed between a.) streaming a bunch of short utterances in sequence (and thus not really fully occupying the GPU) vs b.) streaming long recordings or many short recordings in parallel.