how to export sherpa-ncnn in pruned_transducer_stateless5

[Yymax-max]

how to export sherpa-ncnn in pruned_transducer_stateless5.Have some example？

[csukuangfj]

how to export sherpa-ncnn in pruned_transducer_stateless5.Have some example？

Please have a look at our documentation:
https://k2-fsa.github.io/icefall/model-export/export-ncnn.html

We only support the following models for exporting to ncnn

• LSTM transducer
• ConvEmformer
• Streaming zipformer

[Yymax-max]

how to export sherpa-ncnn in pruned_transducer_stateless5.Have some example？

Please have a look at our documentation: https://k2-fsa.github.io/icefall/model-export/export-ncnn.html

We only support the following models for exporting to ncnn

• LSTM transducer
• ConvEmformer
• Streaming zipformer

Is there any plan to support？

[csukuangfj]

how to export sherpa-ncnn in pruned_transducer_stateless5.Have some example？

Please have a look at our documentation: https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
We only support the following models for exporting to ncnn

• LSTM transducer
• ConvEmformer
• Streaming zipformer

Is there any plan to support？

Sorry, it is not in the plan.

Streaming zipformer is our current best-performing model and we already support it.

[Yymax-max]

how to export sherpa-ncnn in pruned_transducer_stateless5.Have some example？

Please have a look at our documentation: https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
We only support the following models for exporting to ncnn

• LSTM transducer
• ConvEmformer
• Streaming zipformer

Is there any plan to support？

Sorry, it is not in the plan.

Streaming zipformer is our current best-performing model and we already support it.

ok，thank

[Yymax-max]

how to export sherpa-ncnn in pruned_transducer_stateless5.Have some example？

Please have a look at our documentation: https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
We only support the following models for exporting to ncnn

• LSTM transducer
• ConvEmformer
• Streaming zipformer

Is there any plan to support？

Sorry, it is not in the plan.
Streaming zipformer is our current best-performing model and we already support it.

ok，thank

emm，can you tell me current best-performing model in offline，thanks

[csukuangfj]

Please see
https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md#pruned_transducer_stateless8-zipformer--multidataset

It is also zipformer.

---

By the way, the RESULTS.md from each recipe keeps the results of each experiment.
You can find the results from there.

[Yymax-max]

thank

[joazoa]

@csukuangfj can the new zipformer (offline version --causal 0 ) be exported with ncnn ? If so, how and where can i find some code to do offline decoding with the ncnn version ?

[csukuangfj]

the new zipformer (offline version --causal 0 ) be exported with ncnn

No.

sherpa-ncnn supports only streaming models.

[csukuangfj]

@csukuangfj can the new zipformer (offline version --causal 0 ) be exported with ncnn ? If so, how and where can i find some code to do offline decoding with the ncnn version ?

All supported models are listed in the documentation.
https://k2-fsa.github.io/icefall/model-export/export-ncnn.html

If a model is not listed here, then it is not supported.

[joazoa]

Thank you for the lightning fast answer!

I suppose the "zipformer" in the documentation only refers to the older version (transducer_stateless7_streaming ) and not to the new zipformer with causal=1 ) ?

[csukuangfj]

I suppose the "zipformer" in the documentation only refers to the older version

Yes, you are right. The new version ./zipformer has not been supported yet, but it is in the plan.

[joazoa]

thank you for the clarification! Do you have an idea for when support would be planned ?
I ask because I have 15 different models with stateless streaming 7 that I need to redo for other reasons and I need to decide to use either the previous models or the new one.

[csukuangfj]

We will first support sherpa-onnx. There is an on-going issue about it. k2-fsa/icefall#1094
After it is supported, we will start to support sherpa-ncnn.

Could you try sherpa-onnx first once it is supported?

[joazoa]

Yes, absolutely! I will try it as soon as it's available. I'm using the latest zipformer successfully with sherpa already, very happy with it.

[joazoa]

@csukuangfj Now that the Onnx export is done, what is needed for the ncnn export for the zipformer2 ? How can we (try to) help ?

Could we use https://github.com/Tencent/ncnn/blob/master/tools/onnx/onnx2ncnn.cpp ?