search

k2-fsa / sherpa

Speech-to-text server framework with next-gen Kaldi
https://k2-fsa.github.io/sherpa
Apache License 2.0
515 stars 103 forks source link

Any support for ctc guide decode with blank skip? #386

Closed drawfish closed 1 year ago

drawfish commented 1 year ago

It seem that currently shepra only add this feature for triton inference. Is there any plan for support for ctc guide decode with blank skip for transducer?

csukuangfj commented 1 year ago

It seem that currently shepra only add this feature for triton inference. Is there any plan for support for ctc guide decode with blank skip for transducer?

We don't have time to add it to sherpa now, but we welcome contributions from the community.

drawfish commented 1 year ago

It seem that currently shepra only add this feature for triton inference. Is there any plan for support for ctc guide decode with blank skip for transducer?

We don't have time to add it to sherpa now, but we welcome contributions from the community.

I see. If I have time, I will add this. Well, this issue is related to another issue in icefall. How can I check if the export jit model is correct after I modified the code of the frame_reduce.py?

yfyeung commented 1 year ago

You can compare the output of frame_reducer.py between the old and your versions.

Like: https://github.com/k2-fsa/icefall/blob/d5ad908562f432df54ec4592f89e3218bab8940f/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py#L127-L170

yfyeung commented 1 year ago

It seem that currently shepra only add this feature for triton inference. Is there any plan for support for ctc guide decode with blank skip for transducer?

https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/export_onnx.py This recipe supports exporting blank-skip model for both onnx and triton. Note torch.onnx.export() will use torch.jit.trace().

drawfish commented 1 year ago

Good ~