NAIST Graduate School of Information Science
| Date: | Thursday, Oct. 27, 2005 |
|---|---|
| Time: | 13:30 - 16:20 |
| Place: | L1 Lecture Room |
| Language: | English (Oral Presentation), English/Japanese (Question) |
| Chairperson: | Thomas Clouqueur (Computer Design and Test Lab. : PD) Muneyuki Sakata (Image Processing Lab.: PD) |
There are various applications for automatic speech recognition (ASR), e.g. dictation systems, speech-controlled dialogue systems, speech-to-speech translation systems, human machine interfaces, a.s.o. However, there are only very few commercial products which make use of ASR technology. One reason is the fact, that it is not possible to use one and the same ASR system for any application and that the development costs of the main components of an ASR system, especially the acoustic model are very high. This is due to the fact, that robust training of the acoustic model (statistical model with hundreds of thousands of parameters) requires a huge amount of transcribed speech data. However, recording and transcription of speech data is a very costly and time-consuming process. Therefore, we propose a cost-effective method for the construction of task-adapted acoustic models. Instead of collecting large amounts of speech data for a new recognition task, the central idea is to employ a subset of existing speech data for training. The selection of appropriate speech data can be realized by an utterance-based selective training algorithm. An introduction of the algorithm as well as results of evaluation experiments for obtaining an infant- and elderly-dependent will be given in the talk.
To test Integrated Circuits, compactors are used to reduce the volume of
test response to be transfered out of the chip and checked by the
tester. Such compaction complicates the debug of the design because the
location of the errors in the test response is hidden and partially lost
in the compacted signature.
In this presentation, we will study how the errors can be retrieved from
the compacted signature. An algorithm for error diagnosis will be
proposed and evaluated by measuring the resolution obtained when
identifying error patterns.
In the next-generation home network systems(HNS), interoperability among multi-vendor appliances is a challenging issue to implement value-added integrated services. This paper presents a service-oriented framework to enable both evolution of HNS and the appliance interoperability. The key idea is to construct the integrated services by combining the existing services deployed by the appliances. This allows to eliminate the reference model, which had been an obstacle of the evolution of the HNS.
The technology of MANET(Mobile Ad-hoc NETwork) is used in the area of
the mesh network. The mesh network is the network that wireless access
points are connected each other by not wired network infrastructure but
wireless network. Therefore, the area covered wireless network looks
like "mesh". This mesh network mechanism makes the establishment of
wireless network easy.
Currently, we are investigating the security system that is formed the
mesh network by networked camera systems. In this presentation, I
introduce the consideration about the mechanism needed for the security
system using the mesh network. I propose the network architecture of the
networked camera systems configuring the mesh network.
I will talk about a data obfuscation method
for protecting software against malicious reverse engineering attacks.
Our method first calculates a table relating
input of a function (to be obfuscated) to output of the function.
Then we make the each input hard to invert
so that the relation between the input and the output will be obscure.
We believe that our method drastically increases the cost of
reverse engineering attacks since the function
has only a table relating hard-to-invert input to the output
instead of expressions that give attackers clues to obtain secret data.
In this presentation, I'll first talk about
the problem of conventional data obfuscation methods and introduce our idea.
Then I will be discussing difficulty of reverse
engineering attacks of our method
and the overhead that is imposed by applying our method.
The representation method of the location information exchanged on the
Internet becomes more multifaceted because many kinds of devices can
be used to sense objects' location. A location information user needs
to transform the representation style of the acquired information into
the one corresponding to his demand individually. This situation has
spoiled the reusability of location information.
In order to promote general-purpose circulation of location
information, we need to understand what is "place"
essentially; the meaning of expressing place and transforming a
certain representation style into another. In this presentation, I
redefine "place", and raise some
concept. And using these definitions and concepts, I discuss the
essential meaning of expressing "place" and transforming
representation style. By developing the argument based on these
concepts and considerations, it is expected that the effort towards
the improvement in reusability of location information can be
promoted.
I have been studying about the method of estimating the group attention defined as the intersection of measured multiple face-normal vectors. In this presentation, I will firstly present the developed system. The developed prototype system can measure the facial information such as head motion and gaze direction, and estimate the group attention in real-time. Then I will present a method for estimating the group attention. The method for estimating the group attention was improved in order to measure multiple group attentions. The feasibility of new algorithm is confirmed through simulation.
3D model of outdoor environment is applied in number of field such as
site simulation, car navigation system and virtual walk-through.
We proposed a 3D modeling method of outdoor environment by using a laser
rangefinder.
Our proposed method is that range data are acquired at multiple points
in the outdoor environments, are registered simultaneously.
We mainly aim at modeling of urban site, register the range data by
using planar regions such as roads and walls for noises removal and data
reduction.
However, outdoor environment is complex, registration fails when the
planar portions which overlaps between range data do not exist.
Moreover, data acquisition without unobserved portions require huge
manpower.
View planning method which instructs data acquisition position for
efficient data acquisition and success of registration to a sensor
operator is applied to the proposed method.