The lip-sync issue is still the major hurdle why we cannot declare the prototype as ready for the masses. But it is only a matter of time before we also solve this issue. At the same time, we revamp the omx pipeline to simplify our code and to get rid of the external libraries for audio decoding.
The next challenge is to build a lightweight preference-based model for TV item recommendations. It does not have to be perfect, but it should be able to find most of the items the user is interested in, without the need to search all items manually. It sure will be a trade-off between complexity and resources, since time and space is very limited on a single-board computer but it should at least provide better results than a simple keyword-based search.
The design is not finished yet, but it should at least utilize some kind of freely available meta data and it should have support for on-line learning. We currently do some tests with artificial neural networks and the results are quite promising.