Export NeMo FastConformer Hybrid Transducer-CTC Large Streaming to ONNX.

[csukuangfj]

Fixes #790

CC @tempops @sangeet2020

Also CC @titu1994

Exported models can be found at
https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models

---

Will add

• C++ runtime and other API bindings for it
• Android demo
• transducer support

[sangeet2020]

Hi @csukuangfj ,

Thanks for the PR.
If I understand correctly, this PR is only meant for exporting the EncDecHybridRNNTCTCBPEModel model types such as one found in nvidia/stt_en_fastconformer_hybrid_large_streaming_multi.

The decoding implementation needs to be done separately? or can I still use the existing implementation https://k2-fsa.github.io/sherpa/onnx/pretrained_models/offline-ctc/nemo/index.html for these models?

[csukuangfj]

If I understand correctly, this PR is only meant for exporting the EncDecHybridRNNTCTCBPEModel model types such as one found in nvidia/stt_en_fastconformer_hybrid_large_streaming_multi.

Yes, you are right.

---

The decoding implementation needs to be done separately?

No, we can reuse the existing CTC decoding code.

However, we need to change the C++ code to support the neural network part.

By the way, the C++ decoding part does not depend on the specific neural network model; therefore, it can be reused for any model trained by the CTC loss.