Speech: Evidence for the Design of Human-
Computer Speech Interfaces
University of Hamburg
At human-computer speech interfaces, irritations caused by system malfunctions cannot be completely avoided. These irritations are not just local problems which can be easily overcome; they constitute severe problems for human-to- computer communication in at least two ways:
Firstly, the acoustic characteristics of the users' utterances have been found to be very different if they constitute repetitions or reformulations of previous utterances. That is, if the system claims not to have understood a contribution by the speaker, the speaker will repeat his utterance, however, possibly with a different stress pattern, different phrasal intonation, with a strong emphasis on exact pronunciation or even hyper-articulation, and short pauses between the words. Some of these properties may cause severe problems for current automatic speech processing systems; for instance, Levow ( 1998) finds that the error rate in speech recognition rises from 16% to 44% for repetitions. That means that the characteristics of an utterance are very different if it constitutes a repetition of a previous contribution, and that these differences cannot be neglected in human-computer interaction (HCI).
The second problem concerns the fact that speakers may become emotionally involved when they are repeatedly confronted with errors by the automatic speech processing system such that the system's malfunctions may provoke emotional responses in the user; thus, the speakers' attitude towards the system may change over time. This change in attitude may have global consequences for the prosodic, lexical, and conversational properties of the speakers' utterances. For instance, the average pitch may rise, the local properties as the above may occur also when no irritation directly precedes the current utterance,