how to export the sherpa model torchscript?

[lucasjinreal]

also, any doc on training Chinese dataset?

[csukuangfj]

Please have a look at https://k2-fsa.github.io/icefall/recipes/index.html

For exporting models, please see https://k2-fsa.github.io/icefall/recipes/librispeech/lstm_pruned_stateless_transducer.html#export-models

All recipes in icefall have a file export.py that can be used to export models for torchscript.

[lucasjinreal]

@csukuangfj Does sherpa support export to PNNX?

[csukuangfj]

@csukuangfj Does sherpa support export to PNNX?

Currently, only the LSTM transducer model supports exporting to PNNX.

Please see https://k2-fsa.github.io/icefall/recipes/librispeech/lstm_pruned_stateless_transducer.html#export-model-for-ncnn for usage.

[lucasjinreal]

@csukuangfj Hi, how does the performance drop compare with LSTM transducer and transformer arch?

[lucasjinreal]

Does any one of these support export to pnnx? Just tested default model perform very well.

[csukuangfj]

@csukuangfj Hi, how does the performance drop compare with LSTM transducer and transformer arch?

The LSTM transducer model is for streaming ASR. It performs the best so far for streaming ASR on the test-clean dataset of LibriSpeech among models trained with icefall.

---

Does any one of these support export to pnnx? Just tested default model perform very well.

No, none of them supports PNNX. They are conformer-based models and are for non-streaming ASR.

[lucasjinreal]

@csukuangfj Hi, may I ask why does LSTM only do well on streaming task, not other tasks? Why can't conformer beat LSTM on this task

[csukuangfj]

The original Conformer needs to see the whole utterance to compute attention, which is not suitable for streaming ASR.

We have tried chunk-based attention for Conformer in icefall and the experiment result is not as good as LSTM-based models.