Analysis and automatic recognition of false starts in spontaneous speech

The author describes the extent of prosodic phenomena in speech restarts in a multispeaker database of spontaneous, continuous speech, and gives intuitive explanations for them, based on a theory of using prosodics to cue semantic information to a listener. He also gives details (based on the acoustic data) on how to attempt to recognize these phenomena in the context of an automatic speech recognizer. It is shown that simple restarts (i.e., those without inserted or substituted words) could be distinguished acoustically, via an analysis of duration, F0, and spectral detail in the neighborhood of a pause. Including some current refinements, the author expects to be able to identify automatically such restarts with an accuracy exceeding 80%, while keeping false alarms to below 15%. Restarts with changed words may be distinguishable, but the required analysis will need to be much more complex, and beyond the scope of the present work.<>