Or better yet, rank the resulting sentences for each of the possible fits using existing speech API, setting a cutoff to filter out nonsensical results. This might even yield surprises.