A formalization of performance specifications for discrete utterance recognition systems

Many manufacturers of Isolated Word Systems claim "99+ percent recognition accuracy" for their products. However, few are willing to define the database used for this claim. What vocabularies were used? What speakers were employed? What were the ambient noise conditions? It is clear that the potential users of IWR systems need an objective and reliable means of comparing these systems to determine which are appropriate to their application. A guideline is proposed for specifying the database and the system performance.