Eva González-Parada

Topic-Dependent Language Model Switching for Embedded Automatic Speech Recognition

Embedded devices incorporate everyday new applications in different domains due to their increasing computational power.Many of these applications have a voice interface that uses Automatic Speech Recognition (ASR). When the complexity of the language model is high, it is common to use an external server to perform the recognition at the expense of certain limitations (network availability, latency, etc.). This paper focuses on a new proposal to improve the efficiency of the usage of the language model in a recognizer for multiple domains. The idea is based on the selection of a proper language model for each domain within the ASR system.

Rule-Based vs. Behavior-Based Self-Deployment for Mobile Wireless Sensor Networks

In mobile wireless sensor networks (MWSN), nodes are allowed to move autonomously for deployment. This process is meant: (i) to achieve good coverage; and (ii) to distribute the communication load as homogeneously as possible. Rather than optimizing deployment, reactive algorithms are based on a set of rules or behaviors, so nodes can determine when to move. This paper presents an experimental evaluation of both reactive deployment approaches: rule-based and behavior-based ones. Specifically, we compare a backbone dispersion algorithm with a social potential fields algorithm. Most tests are done under simulation for a large number of nodes in environments with and without obstacles. Results are validated using a small robot network in the real world. Our results show that behavior-based deployment tends to provide better coverage and communication balance, especially for a large number of nodes in areas with obstacles.

An empirical study on the performance of bluetooth scatternets

Bluetooth scatternet technology was conceived to provide support for ad hoc multihop networking applications. Many theoretical works in the literature have been dedicated to optimize the formation and operation of Bluetooth scatternets. However, the performance of a real scatternet strongly relies on the implementation of the Bluetooth stack in the interfaces of the network nodes. By analyzing the behavior of a simple scatternet of three nodes, this paper shows evidence of the scalability and delay problems that can appear when scatternets are deployed on existing hardware interfaces.

ECA-based control interface on Android for home automation system

Historically, Embodied Conversational Agents (ECAs) have been used as virtual assistants that make easier the access to information or help in performing complex tasks. Due to their high computational requirements ECAs are usually run on desktop computers, but with the recent development of hand-held devices both in hardware and software, it becames neccessary to move ECAs to that new mobile scenario. Thus, we propose an open-source based platform for developing ECA based interfaces on Android-equipped devices. We also present a prototype for controlling a home automation system.

Efficient Use of Voice Activity Detector and Automatic Speech Recognition in Embedded Platforms for Natural Language Interaction

Nowadays people are faced daily with the management of all types of technological devices that have embedded processors inside. The desire of users and the industry is to achieve a natural interaction with these devices. Researchers have spent many years working in spoken dialog systems which are now used in many applications. In these systems it is crucial to achieve correct speech recognition. Usually the largest research effort focuses on the robustness to noise in all kinds of adverse conditions but often the response time is ignored. This paper focuses on a new approach for the efficient use of voice activity detection and speech recognition in embedded devices for natural language interaction. We propose a new approach to adjust the response time of the recognition system to the requirements of the overall implementation without sacrificing too much accuracy.

AVATAR: An Open Source Architecture for Embodied Conversational Agents in Smart Environments

Due to a growing older population, researchers and industry are paying more attention to the needs of this group of people. Ambient Intelligence (AmI) aims to help people in their daily lives, achieving a more natural interaction of users with an electronic home environment. Embodied Conversational Agents (ECAs) arise as a natural interface between humans and AmI. Our contribution is to present AVATAR: an architecture to develop ECAs based on open source tools and libraries. In the current prototype the virtual agent acts as a natural control interface of the home automation system. In addition, we provide the details to allow its use by Spanish speakers.

Embedded Conversational Engine for Natural Language Interaction in Spanish

Nowadays there is a great interest in the development of Human-Computer Interaction (HCI) systems that allow the user to communicate in natural language, thus reducing the training required to use them. This paper focuses on the design and implementation of a conversational engine aimed at developing conversational agents in spanish for embedded systems. Our proposal adds a new element to a baseline conversational engine: a lemmatizer. This new element improves the naturalness of the dialogue, especially for highly inflectional languages such as Spanish, without increasing the size of the knowledge base. However the use of the lemmatizer implies an increase of both the memory allocated to the whole application and the total response time of the system. In order to alleviate the use of memory, the storage of the knowledge base of the conversational engine in an object-oriented database is also proposed and evaluated.

A Social Potential Fields Approach for Self-Deployment and Self-Healing in Hierarchical Mobile Wireless Sensor Networks

Autonomous mobile nodes in mobile wireless sensor networks (MWSN) allow self-deployment and self-healing. In both cases, the goals are: (i) to achieve adequate coverage; and (ii) to extend network life. In dynamic environments, nodes may use reactive algorithms so that each node locally decides when and where to move. This paper presents a behavior-based deployment and self-healing algorithm based on the social potential fields algorithm. In the proposed algorithm, nodes are attached to low cost robots to autonomously navigate in the coverage area. The proposed algorithm has been tested in environments with and without obstacles. Our study also analyzes the differences between non-hierarchical and hierarchical routing configurations in terms of network life and coverage.

On the Impact of RLC Layer Configuration Parameters in UMTS Internet Access

The main goal of this paper is the analysis of the interaction between TCP and the UMTS link-level layer (RLC) to determine an optimal configuration for the parameters of both protocols. For this purpose, a Web browsing simulation framework has been considered. This framework includes a Web client accessing to remote HTTP servers through the UTRAN and Internet. The impact of the RLC configuration parameters on the behaviour of the multiple TCP connections simulated under this framework is analysed, showing their effect on Web traffic performance.

A new methodology for TCP evaluation in a multiuser web environment

This paper presents a new methodology to evaluate and graphically represent TCP performance in a web environment. The main novelty of this work is that it focuses on web environments involving a large number of connections, where the traffic model is extracted from real traces. In these cases, conventional tools are not efficient to handle the complexity of the analysis. The proposed framework includes: (i) a set of representative web browsing scenarios affected by different types of losses; (ii) a new analysis methodology to cope with the huge data volume related to the simulated connections and; (iii) a new graphical representation to allow an easy visualisation of the simulation results. To test the proposed methodology, an evaluation of the impact of two representative TCP configuration parameters for web traffic has been included.