Variational Recurrent Auto-Encoder (VRAE) that learns to map incorrectly spelled tokens to correctly spelled ones by creating an intermediate low-dimensional hash.
The correction suggestion would come out of a k-Nearest-Neighbor search with the intermediate hash representation.
Such a representation could, for example, be 8xfloat16 per token.
This will also require corruption rules in the training grammar.
Research options for neural spelling correction OR error-labeling.