Word-Level Speech Recognition With a Letter to Word Encoder