Binbin Zhang
Binbin Zhang
Add `continuous_decoding=True` when initializing the `Decoder`, please see https://github.com/wenet-e2e/wenet/blob/main/runtime/binding/python/README.md.
Just prepare the nbest candidates and feed it to rescoring module, please refer https://github.com/wenet-e2e/wenet/blob/main/runtime/core/decoder/asr_decoder.cc#L80.
Yes, 2 hours are too few to adapt the wenetspeech model.
aishell-1 should give poorer result than wenetspeech since the model is trained on much less data(200 vs 10,000). You can combine the two hours data with sampled data(50~100h) from wenetspeech,...
目前我们的 runtime 和 binding 的编译暂时还没有考虑 ARM 平台，但理论上 LibTorch 支持的平台 wenet 均可以支持。你们可以调研一下如何解决，解决之后方便的话可以贡献一下方案。
wenetruntime 中没有暴露参数设置的接口，wenetruntime 中默认使用 attention_rescoring 的解码方式，对应 recognize.py 中的 `--mode attention_rescoring`.
Just follow https://github.com/wenet-e2e/wenet/tree/main/runtime/libtorch#run-with-local-build to build. It seems it works on your friend's computer, so please check your env. I have no idea from the log, I think you can align...
Sorry, I have no idea.
Great job, I will go through the details as soon as I can.