Received: from GLINDA.OZ.CS.CMU.EDU by A.GP.CS.CMU.EDU id aa04746; 28 Feb 95 14:08:44 EST Date: Tue, 28 Feb 95 14:07:11 EST From: AI.Repository@GLINDA.OZ.CS.CMU.EDU To: ai+ai-postdoc@cs.cmu.edu Subject: Postdoc: Speech Dialog at Erlangen-Nuerenberg University Sender: ai@A.GP.CS.CMU.EDU From: lieske@forwiss.uni-erlangen.de (Christian Lieske) Subject: post-doctoral fellowships HCM/MSDoS at Erlangen-Nuernberg University Date: Fri, 24 Feb 1995 09:46:46 GMT Organization: Regionales Rechenzentrum Erlangen, Germany == post-doctoral fellowships HCM/MSDoS at Erlangen-Nuernberg University == MSDoS -- Modelling Spontaneous Dialogs of Speech Spontaneous speech is a very important research topic if speech technology is to be used in real applications. Even though there are well known large differences between read speech and spontaneous speech, our presently available speech understanding and dialog system EVAR -- like most other systems -- assumes grammatically correct input. Typical phenomena of spontaneous speech that make it ungrammatical are corrections, filled pauses, restarts, non-speech phenomena, ungrammatical order of constituents. In addition, words can occur that are unknown to the recognition system. Recognition results can greatly be improved using stochastic language models (Again most results for language models refer to read speech). These models give for each word of the recognition lexicon a probability of appearence, given the words uttered so far. Normally this is approximated by estimating these probabilities for a large training corpus, given the last or the last 2 words. This approach is to be adopted for spontaneous speech. Another important research topic is a greater robustness and flexibility of currently implemented dialog systems especially with respect to the handling of user-interrupts for topic shift. Topic shifts imply a change in the language model used for recognition. The fact that there is a topic shift has to be decided by the dialog module based on the results from the word recognition module. Thus these two tasks are losely coupled. The necessary steps include: * Collection and transliteration of a corpus of spontaneous human-machine-dialogs. This is to be done, using our EVAR system. * Creation of a language model for spontaneous speech that takes the spontaneous speech phenomena described above into account. The main problem will be how to process partially uttered words, unknown word, and non-speech events, i.e. how to process parts of speech that cannot be analyzed and create a language model of high predictive power for the following speech events. * Extension of the current dialog system to achieve greater robustness and flexibility. For this a careful analysis of the user input that leads to system errors and an incremental improvement of the system has to be done. We expect the system to have a initial failure rate of app. 40\%. This is not surprising because the system was not trained so far with spontaneous dialogs. Certain discourse behaviors like topic shift are not modelled yet in the system. The necessary steps include changes in the dialog memory and adaptation of the recognition module in order to recognize special vocabulary indicating topic shift. For these tasks we have two HCM post-doctoral fellowships for a duration of 9 month each. The task of collecting a speech corpus will be split between the two researchers. It is planned that the researcher working on language models will start working app. one year before the person working on the dialog robustness who can then already start working on the first part of the corpus. This second researcher can then collect his data with the already improved version of EVAR. Together with 5 other research groups we applied for the HCM network ``SPIN -- Spontaneous Speech Recognition in Real Environments''. If this network is accepted, the two researchers from the MSDoS proposal will greatly benefit from a close cooperation with the researchers involved in the SPIN network. However, even though the topics covered in the two proposals are strongly related to each other, there is no overlap of the research tasks between the MSDoS proposal and the SPIN network. Both proposals can be worked upon independently. People interested in the Project should write to: Prof. Dr.-Ing. H. Niemann Lehrstuhl fuer Mustererkennung (Informatik 5) Universitaet Erlangen-Nuernberg Martensstrasse 3 D-91058 Erlangen F.R. of Germany e-mail: niemann@informatik.uni-erlangen.de or to Elmar Noeth Lehrstuhl fuer Mustererkennung (Informatik 5) Universitaet Erlangen-Nuernberg Martensstrasse 3 D-91058 Erlangen F.R. of Germany e-mail: noeth@informatik.uni-erlangen.de --- Christian Lieske |Tel.: 049/9131/691 137 |Fax.: 049/691 185 Bayerisches Forschungszentrum | fuer wissensbasierte Systeme | - FORWISS - | Am Weichselgarten 7 | | D-91058 Erlangen-Tennenlohe | ------------------------------------------------------------------------------- This message | Submissions ai+ai-postdoc@cs.cmu.edu was sent via | Subscribe/Unsubscribe ai+query@cs.cmu.edu the AI-POSTDOC | Available mailing lists include mailing list. | AI-JOBS, LISP-JOBS, PROLOG-JOBS, AI-POSTDOC, AI-PREDOC