Current approach of speech recognition is about as old (maximum likelihood using WFST). In the old papers about problem, there was vocabulary size of 64K words

Current approach of speech recognition is about as old (maximum likelihood using WFST). In the old papers about problem, there was vocabulary size of 64K words, because nothing was working for bigger vocabularies.