Added zipformer ctc streaming support for librispeech at egs/librispeech/ASR/zipformer_ctc_streaming

[uni-manjunath-ke]

Added zipformer ctc streaming support for librispeech (egs/librispeech/ASR/zipformer_ctc_streaming) Pls refer https://github.com/k2-fsa/icefall/issues/999#issuecomment-1569761925

[yaozengwei]

Thanks for your contribution!

I have some questions. It looks like the streaming zipformer code zipformer.py is copied from pruned7_steaming recipe, which should be a soft link instead. I am also confusing about why it has some codes that should not be used, like transformer.py and subsampling.py.

BTW, we did plan to support ctc model in recipe zipformer, trained with pruned-rnnt loss for higher performance. We will also support streaming CTC decoding, like streaming_decode.py.

[uni-manjunath-ke]

Hi @yaozengwei

1. Ya correct. I have copied zipformer.py from pruned7_steaming recipe. Sure, we can create a soft link.
2. Basically, we took a clone of zipformer-ctc code from https://github.com/desh2608/icefall/tree/zipformer_ctc and modified it to zipformer-ctc-streaming by following the steps given at https://github.com/k2-fsa/icefall/issues/999#issuecomment-1528817694 . This had codecs like transformer.py and subsampling.py.
3. When is the support for ctc model in recipe zipformer is expected to be rolled out if you already have something in pipeline. How about Sherpa support for such zipformer-ctc? Thanks