We plan to add CTC decoding support for streaming models with graph(s) in C++.
As for the models, they need not necessarily come from icefall. As long as there is a torchscript model available (or an ONNX model for sherpa-onnx), we should support it.
As for the graph, it can be an H or an HLG (i.e., TLG). We also need to support using a context graph during the search.
At the very beginning, I suggest we use a streaming zipformer trained by @yaozengwei using transducer + CTC loss and implement a Python version for CTC decoding with graphs as it is easier to debug in Python. After that, we can port the implementation to sherpa and sherpa-onnx.
We plan to add CTC decoding support for streaming models with graph(s) in C++.
As for the models, they need not necessarily come from icefall. As long as there is a torchscript model available (or an ONNX model for sherpa-onnx), we should support it.
As for the graph, it can be an H or an HLG (i.e., TLG). We also need to support using a context graph during the search.
At the very beginning, I suggest we use a streaming zipformer trained by @yaozengwei using transducer + CTC loss and implement a Python version for CTC decoding with graphs as it is easier to debug in Python. After that, we can port the implementation to sherpa and sherpa-onnx.