Injecting custom online features with a nnet3 model using online2

[asumagic]

I have a nnet3 model which uses custom features along side iVectors. It works fine for offline ASR using Kaldi.

Now, I'm trying to get this model working for online decoding using Pykaldi. However, I am a bit stuck as to how this should be done.

In my usecase, it does not seem like OnlineNnetFeaturePipeline would be usable as it is hardcoded in Kaldi to only support MFCCs, PLPs and filterbanks.
Even if it were, I cannot really think of a way the Pykaldi kaldi.feat classes could be leveraged to inject custom online features from Python code (OnlineMatrixFeature seems like the closest match, but it seems to be only useful in specific usecases where some info is known ahead of decoding). Currently, it doesn't seem possible to define custom OnlineFeatureInterfaces from Python code.

So, I do have a few leads, but I don't really know how to glue everything together, and whether I'm missing on something obvious. As it does seem like I may have to modify Pykaldi in the end, I could try to contribute towards a PR if there is upstream interest towards this issue - hence why I'd like to know what a sane approach would be here.