Advances in speech transcription at IBM under the DARPA EARS program

This paper describes the technical and system building advances made in IBM's speech recognition technology over the course of the Defense Advanced Research Projects Agency (DARPA) Effective Affordable Reusable Speech-to-Text (EARS) program. At a technical level, these advances include the development of a new form of feature-based minimum phone error training (fMPE), the use of large-scale discriminatively trained full-covariance Gaussian models, the use of septaphone acoustic context in static decoding graphs, and improvements in basic decoding algorithms. At a system building level, the advances include a system architecture based on cross-adaptation and the incorporation of 2100 h of training data in every system component. We present results on English conversational telephony test data from the 2003 and 2004 NIST evaluations. The combination of technical advances and an order of magnitude more training data in 2004 reduced the error rate on the 2003 test set by approximately 21% relative-from 20.4% to 16.1%-over the most accurate system in the 2003 evaluation and produced the most accurate results on the 2004 test sets in every speed category.     

Choice between fixed and variable delays with different reward amounts.

Choice between a variable and a fixed food source was studied in pigeons. The variable source yielded different durations of access to food (changed across experimental conditions) uniquely associated with equiprobable delays of either 20 sec or 60 sec, timed from the choice point. A comparison, elapsing alternative (time-left procedure [J. Gibbon and R. M. Church; see PA, Vol 67:2812]), was associated with fixed access to feeder after a 60-sec delay, timed from the beginning of the trial. Preference was best characterized as reflecting an average of the 2 local rates of reinforcement in the variable alternative (the average of the ratio of each standard amount and its associated delay). The averaging rule characterized initial choice link data well, but terminal link responding showed that Ss also remembered the 2 delay–amount combinations separately. (PsycINFO Database Record (c) 2010 APA, all rights reserved)     

Comparing Universal Covers in Polynomial Time

The universal cover T _G of a connected graph G is the unique (possibly infinite) tree covering G, i.e., that allows a locally bijective homomorphism from T _G to G. It is well-known that if a graph G covers a graph H, then their universal covers are isomorphic, and that the latter can be tested in polynomial time by checking if G and H share the same degree refinement matrix. We extend this result to locally injective and locally surjective homomorphisms by following a very different approach. Using linear programming techniques we design two polynomial time algorithms that check if there exists a locally injective or a locally surjective homomorphism, respectively, from a universal cover T _G to a universal cover T _H (both given by their degree matrices). This way we obtain two heuristics for testing the corresponding locally constrained graph homomorphisms. Our algorithm can also be used for testing (subgraph) isomorphism between universal covers, and for checking if there exists a locally injective or locally surjective homomorphism (role assignment) from a given tree to an arbitrary graph H.     

A network approach to the French system of legal codes—part I: analysis of a dense network

We explore one aspect of the structure of a codified legal system at the national level using a new type of representation to understand the strong or weak dependencies between the various fields of law. In Part I of this study, we analyze the graph associated with the network in which each French legal code is a vertex and an edge is produced between two vertices when a code cites another code at least one time. We show that this network distinguishes from many other real networks from a high density, giving it a particular structure that we call concentrated world and that differentiates a national legal system (as considered with a resolution at the code level) from small-world graphs identified in many social networks. Our analysis then shows that a few communities (groups of highly wired vertices) of codes covering large domains of regulation are structuring the whole system. Indeed we mainly find a central group of influent codes, a group of codes related to social issues and a group of codes dealing with territories and natural resources. The study of this codified legal system is also of interest in the field of the analysis of real networks. In particular we examine the impact of the high density on the structural characteristics of the graph and on the ways communities are searched for. Finally we provide an original visualization of this graph on an hemicyle-like plot, this representation being based on a statistical reduction of dissimilarity measures between vertices. In Part II (a following paper) we show how the consideration of the weights attributed to each edge in the network in proportion to the number of citations between two vertices (codes) allows deepening the analysis of the French legal system.     

Comparing and evaluating the effectiveness of mobile Web adequacy evaluation tools

Universal Access in the Information Society - The number of Web sites is growing exponentially and so are the people who are accessing them on mobile devices including people with special...     

Automatic Lighting Design using a Perceptual Quality Metric

Lighting has a crucial impact on the appearance of 3D objects and on the ability of an image to communicate information about a 3D scene to a human observer. This paper presents a new automatic lighting design approach for comprehensible rendering of 3D objects. Given a geometric model of a 3D object or scene, the material properties of the surfaces in the model, and the desired viewing parameters, our approach automatically determines the values of various lighting parameters by optimizing a perception-based image quality objective function. This objective function is designed to quantify the extent to which an image of a 3D scene succeeds in communicating scene information, such as the 3D shapes of the objects, fine geometric details, and the spatial relationships between the objects.
Our results demonstrate that the proposed approach is an effective lighting design tool, suitable for users without expertise or knowledge in visual perception or in lighting design.     

Dufour Gland Contents of Ants of the Cataglyphis bicolor Group

The species of desert-dwelling ants of the Cataglyphis bicolor (Hymenoptera: Formicidae) group are difficult to distinguish by morphological features. Analysis of the secretion from the Dufour glands of workers of a number of colonies was undertaken to see if it provided a clear test of species. Linked 6c-ms showed in all samples straight and branched-chain alkanes, linear alkenes, ketones, aldehydes, acetates, and a group of C₂₂ to C₂₈ esters not previously identified in this genus. Contents of the Dufour glands of C. savignyi from Tunisia and Egypt were similar, and comprised straight and branched-chain alkanes, alkenes and small amounts of esters. C. bicolor from Tunisia contained compounds similar to C. savignyi but was distinguished from the latter by larger amounts of the esters. The major compound in the glands of C. viaticus was tridecane, in contrast to the pentadecane of other species. It also contained a branched alkane, 3-methyltridecane as a major component. Branched-chain esters and a wide variety of acetates were also found in this species. C. diehlii had a limited range of compounds, with branched alkanes almost completely absent and high proportions of pentadecene and dodecyl acetate. C. bombycinus, a sympatric species, but recognized as not belonging to the bicolor group by its different mandibular gland substances, was notable in having butanoate esters in its Dufour glands. Despite these differences among species, both the great variability of individuals from a single colony and the among between conspecific colonies make species diagnosis from a few individuals difficult, in contrast with postpharyngeal glands, which, as recently reported, give a clearer indication of species.     

The cascaded nature of lexical selection and integration in auditory sentence processing.

An event-related brain potential experiment was carried out to investigate the temporal relationship between lexical selection and the semantic integration in auditory sentence processing. Participants were presented with spoken sentences that ended with a word that was either semantically congruent or anomalous. Information about the moment in which a sentence-final word could uniquely be identified, its isolation point (IP), was compared with the onset of the elicited N400 congruity effect, reflecting semantic integration processing. The results revealed that the onset of the N400 effect occurred prior to the IP of the sentence-final words. Moreover, the factor early or late IP did not affect the onset of the N400. These findings indicate that lexical selection and semantic integration are cascading processes, in that semantic integration processing can start before the acoustic information allows the selection of a unique candidate and seems to be attempted in parallel for multiple candidates that are still compatible with the bottom-up acoustic input. (PsycINFO Database Record (c) 2010 APA, all rights reserved)