The point of word embeddings is that they capture semantic similarity. Basing the similarity calculation on spelling alone will fail to do that. Will ‘beautiful’ be near ‘gorgeous’?

You can try training on a language model task in parallel where you sample misspelled words during training time.