New language supports for Telugu and Kannada. These are experimental lite recognition models. Their file sizes are only around 7% of other models and they are ~6x faster at inference with CPU.

This release is also a preparation for user-created models/architectures in the future.