Hi Erick,
I think it might be possible to handle the part about just capturing the speech without trying to get a correct hypothesis out of it – perhaps a Rejecto-only model would handle this requirement. The problem is that the audio session interference is likely to have a negative effect on the behavior of audio buffering and transference into the engine. That is, if the NDK is preventing OpenEars from doing recognition due to the session changes, it will probably also affect whether it can save WAVs which have correct data in them. Have you given it a try? It might be worth some experimentation.