Fangjun Kuang
Fangjun Kuang
Recently, two new models are added to sherpa-ncnn: - https://huggingface.co/csukuangfj/sherpa-ncnn-conv-emformer-transducer-2022-12-04 (Support English) - https://huggingface.co/csukuangfj/sherpa-ncnn-conv-emformer-transducer-2022-12-06 (Support both English and Chinese) - [ ] The README.md needs to be updated to include...
This PR only defines the interface for the decoder. Please have a look at - https://github.com/k2-fsa/sherpa/blob/master/sherpa/csrc/online-transducer-decoder.h - https://github.com/k2-fsa/sherpa/blob/master/sherpa/csrc/online-transducer-greedy-search-decoder.cc - https://github.com/k2-fsa/sherpa/blob/master/sherpa/csrc/online-transducer-modified-beam-search-decoder.cc ## TODOs - [ ] Rename `decode.h` and `decode.cc` to...
Fixes https://github.com/k2-fsa/k2/issues/888 and https://github.com/k2-fsa/k2/issues/854 --- It seems to work.
This PR shows that we can also use k2 for decoding with models from other frameworks. For n-gram LM rescoring and attention decoder rescoring, it shares a lot of code...
As we are doing autograd in C++ and because of the change introduced in https://github.com/pytorch/pytorch/pull/28287 the current code won't compile for torch 1.3.1 and below. You can see that the...
See https://github.com/k2-fsa/snowfall/pull/166#discussion_r615521796 > ... Incidentally, the opposite to this function would, I guess, be convert_attr_to_linear(), which would have to call ExpandArcs(). However this would have to be a function in...
#690 changes the format of the string representation of FSAs in k2. We should update the documentation accordingly.
See https://github.com/k2-fsa/k2/pull/599#discussion_r560670103 > An issue for later is to try to use a weak dictionary to introduce a caching mechanism for the results of get_forward_scores and get_backward_scores and get_arc_post (and...