From speech recognition to understanding: Shifting the paradigm to achieve natural human–machine communication

Automatic speech recognition has made substantial progress in the last decade due to applications of statistical methods. The task of speech recognition is often described as converting a sequence of sounds into a sequence of linguistic events (words). This is in essence transcription: given an acoustic signal, the system determines which word it represents. The implicit assumption in this traditional formulation of speech recognition is that the speech is well‐formed, free from extraneous components such as disfluency, partial words, repairs, hesitation, etc. The statistical framework that supports the formulation is the Bayes decision theory. For natural human–machine communications, however, it is expected that the spoken input to the system will inevitably contain extraneous components, treatments of which have to be developed in a broadened framework utilizing the Neymann–Pearson Lemma. The departure also inspired a new formulation based on the signal detection theory for robust speech understanding....