Development fragmentation

[trunglebka]

Currently we have 3 versions of sherpa: this repo, ncnn, onnx.
Do you think that this seperation may cause fragmentation?

(I don't have much expertise in C++ has not used ncnn, onnx versions yet but some time encountered this problem in general programing)

[csukuangfj]

sherpa, this repo, uses libtorch. It is targeted for x86 servers.

sherpa-ncnn uses ncnn. It is mainly targed for embedded systems and mobile systems, e.g., raspberry pi, android, iOS.

sherpa-onnx uses onnxruntime. It can support both embedded systems and mobile systems as well as x86 servers

They are separated so that it simplifies the building system and also simplifies maintenance.

[trunglebka]

Thanks you. Just worry about maintenance burden but look like it is tied to ML framework too much to put together

[trunglebka]

@csukuangfj Related to the OP - since there are many implementations, can you tell me which one has more priority?
Currently I'm using this repo as starting point. I prefer solution that not being locked by Pytorch (due to installation), but I don't know if sherpa-onnx is keep pace with sherpa and that any performance gain/loss with sherpa-onnx.

[csukuangfj]

since there are many implementations, can you tell me which one has more priority?

We would support all three: sherpa, sherpa-ncnn, and sherpa-onnx.

You can choose anyone you like.

[trunglebka]

I can see that from commit history, but which will get icefall's new recipes implement first?

[csukuangfj]

I would say sherpa.

Because exporting a model to sherpa is super straightforward, but it is not the case for sherpa-onnx and sherpa-ncnn.

[trunglebka]

Thanks, I asked this because the last time I checked I want to use streaming Zipformer but it had not yet implemented at that time and waiting is not happy (^3^)

[csukuangfj]

I think streaming zipformer (both v1 and v2) is already supported in sherpa. We are supporting sherpa-onnx right now.
(We have not exported the v2 model to sherpa-ncnn yet)