Implications of an Early Experiment in Speech Understanding

A survey of the literature on speech recognition reveals two distinct goals for long range research, transcription and understanding. Transcription is defined as the process of converting speech to ordinary text without regard for meaning. Understanding, on the other hand is taken to mean translating speech into appropriate actions such as correctly answering the speaker's question or properly executing his command. There are certain legitimate technical uses for transcription such as very low bit rate transmission and some very specialized types of stenography. However, these applications are rather artificial and, even so, arguably impossible to achieve without some ability to understand the message. When pressed on the issue, most speech researchers would, I suspect, acknowledge the centrality of understanding the message. In their laboratories, though, these investigators concentrate their efforts on transcription with the intention of combining their results with those of others working on understanding at some time in the future. Implicit in this behavior is the assumption that transcription and understanding are distinct processes connected at a simple interface.