How to Drink from a Fire Hose: One Person Can Annoscribe One Million Utterances in One Month

David Suendermann,  Jackson Liscombe,  Roberto Pieraccini
SpeechCycle, New York, USA


Transcription and semantic annotation (annoscription) of utterances is crucial part of speech performance analysis and tuning of spoken dialog systems and other natural language processing disciplines. However, the fact that these are manual tasks makes them expensive and slow. In this paper, we will discuss how annoscription can be partially automated. We will show that annoscription can reach a throughput of one million utterances per person month under certain assumptions.