User modeling is an iml>ortant COlnponents of dialog systems. Most previous approaches are rule-based methods, hi this paper, we proimse to represent user models through Bayesian networks. Some advantages of the Bayesian approach over the rule-based approach are as follows. First, rules for updating user models are not necessary because upility theory; this provides us a more formal way of dealing with uncertainties. Second, the Bayesian network pro: rides more detailed information of users' knowledge, because the degree of belief on each concept is provided in terms of prol~ability. We prove these advantages through a prelinfinary experiment.

No abstract

SUMMARYWe propose a method for creating an N-gram language model for use in a speech-operated question-answering system. We note that input questions to such a system frequently consist of an initial section, relating to the query topic, and a formulaic sentence final expression that is used in questions (a fixed phrase). While we are able to model the initial sections adequately using the target query newspaper corpus, we are not able to model the fixed phrases adequately with this data source. In this paper we frame the problem as one of adapting a language model created using a generic corpus to fixed phrases and propose a language model adaptation method that makes use only of a list of fixed phrases created by hand, rather than attempting the more difficult task of collecting an adaptation corpus. In the proposed method we determine which sections in the generic corpus correspond to N-gram sequences on the list of fixed phrases, and perform language model adaptation by amplifying the probabilities of those N-grams; this is equivalent to performing maximum a posteriori (MAP) estimation treating these partial N-gram sequences from the generic corpus itself as posterior information. We perform recognition experiments with spoken questions consisting of input to a question-answering system and confirm the effectiveness of the proposed method.

No abstract

Deep neural networks (DNNs) have achieved significant success in the field of automatic speech recognition. One main advantage of DNNs is automatic feature extraction without human intervention. However, adaptation under limited available data remains a major challenge for DNN-based systems because of their enormous free parameters. In this paper, we propose a filterbank-incorporated DNN that incorporates a filterbank layer that presents the filter shape/center frequency and a DNN-based acoustic model. The filterbank layer and the following networks of the proposed model are trained jointly by exploiting the advantages of the hierarchical feature extraction, while most systems use predefined mel-scale filterbank features as input acoustic features to DNNs. Filters in the filterbank layer are parameterized to represent speaker characteristics while minimizing a number of parameters. The optimization of one type of parameters corresponds to the Vocal Tract Length Normalization (VTLN), and another type corresponds to feature-space Maximum Linear Likelihood Regression (fMLLR) and feature-space Discriminative Linear Regression (fDLR). Since the filterbank layer consists of just a few parameters, it is advantageous in adaptation under limited available data. In the experiment, filterbank-incorporated DNNs showed effectiveness in speaker/gender adaptations under limited adaptation data. Experimental results on CSJ task demonstrate that the adaptation of proposed model showed 5.8% word error reduction ratio with 10 utterances against the un-adapted model.

No abstract

The lecture is one of the most valuable genres of audiovisual data. Though spoken document processing is a promising technology for utilizing the lecture in various ways, it is difficult to evaluate because the evaluation require a subjective judgment and/or the verification of large quantities of evaluation data. In this paper, a test collection for the evaluation of spoken lecture retrieval is reported. The test collection consists of the target spoken documents of about 2,700 lectures (604 hours) taken from the Corpus of Spontaneous Japanese (CSJ), 39 retrieval queries, the relevant passages in the target documents for each query, and the automatic transcription of the target speech data. This paper also reports the retrieval performance targeting the constructed test collection by applying a standard spoken document retrieval (SDR) method, which serves as a baseline for the forthcoming SDR studies using the test collection.

A speech-driven information retrieval system is expected to be useful for gathering information with greater ease. In a conventional system, users have to decide on the contents of their utterance before speaking, which takes quite a long time when their request is complicated. To overcome that problem, it is required for the retrieval system to handle a spontaneously spoken query directly. In this work, we propose an extension technique of spoken content retrieval (SCR) for effectively using spontaneously spoken queries. Acoustic features of meaningful terms in the retrieval may have prominence compared to other terms. Also, those terms will have linguistic specificity. From this assumption, we predict the contribution of terms included in spontaneously spoken queries using acoustic and linguistic features, and incorporate it in the query likelihood model (QLM) which is a probabilistic retrieval model. We verified the effectiveness of the proposed method through experiments. Our proposed method was successful in improving retrieval performance under various conditions.