The use of menu based speech recognition in which callers get the desired information or are routed to the right department is limited by the a-priori knowledge of these menus by the callers. Automatic speech recognition has always been promoted as superior to classic DTMF-IVR (push 1 for John, 2 for ..) because it is faster and more user friendly. Instead of listening to endless menu options and pushing the corresponding key(s), a user may just say the desired item. This works fine but the problem is that often people do not know what to say. Classification based on the results of mostly imperfect speech recognition of open questions may help to overcome this deadlock. In situations were the caller can explain his/her problem but does not know were to address it too, this approach may be very useful. Moreover, classification of incoming calls can be used to generate automatically management reports about what time people spoke about which subject. In this paper we will discuss the design and the results of the first pilot implementation of a speech recogniser in an Information Retrieval (IR) system at an assurance company. Incoming calls are recognised and matched with previous, already classified calls. The labels of these classified calls are used to calculate a label (i.e. class) for this new call.
[1]
Giuseppe Riccardi,et al.
How may I help you?
,
1997,
Speech Commun..
[2]
Wessel Kraaij,et al.
Variations on language modeling for information retrieval
,
2005,
SIGF.
[3]
Wessel Kraaij,et al.
Evaluation of a Dutch stemming algorithm
,
1994
.
[4]
Roeland Ordelman,et al.
Dutch speech recognition in multimedia information retrieval
,
2003
.
[5]
K. Sparck Jones,et al.
Simple, proven approaches to text retrieval
,
1994
.
[6]
Dietrich Klakow,et al.
Information retrieval based call classification
,
2003,
INTERSPEECH.
[7]
Min Tang,et al.
Call-type classification and unsupervised training for the call center domain
,
2003,
2003 IEEE Workshop on Automatic Speech Recognition and Understanding (IEEE Cat. No.03EX721).