Lucent Technologies Inc.

At+t Corp.

Christopher John Burges

Eric Cosatto

Hans Peter Graf

John Stewart Denker

Ofer Matan

Yann A Le Cun

Speaker identification is performed using a single Gaussian mixture model (GMM) for multiple speakers&mdash;referred to herein as a Discriminative Gaussian mixture model (DGMM). A likelihood sum of the single GMM is factored into two parts, one of which depends only on the Gaussian mixture model, and the other of which is a discriminative term. The discriminative term allows for the use of a binary classifier, such as a support vector machine (SVM). In one embodiment of the invention, a voice messaging system incorporates a DGMM to identify the speaker who generated a message, if that speaker is a member of a chosen list of target speakers, or to identify the speaker as a &ldquo;non-target&rdquo; otherwise.

Discriminative gaussian mixture models for speaker verification

A method and apparatus is described for improving the efficiency of any machine that uses an algorithm that maps to a higher dimensional space in which a given set of vectors is used in a test phase. In particular, reduced set vectors are used. These reduced set vectors are different from the vectors in the set and are determined pursuant to an optimization approach other than the eigenvalue computation used for homogeneous quadratic kernels. An illustrative embodiment is described in the context of a support vector machine (SVM).

Method and apparatus for improving the efficiency of support vector

A kernel-based method and apparatus includes a preprocessor, which operates on an input data in such a way as to provide invariance under some symmetry transformation.

Eliminating invariances by preprocessing for kernel-based methods

A method and system for forming an interpretation of an input expression, where the input expression is expressed in a medium, the interpretation is a sequence of symbols, and each symbol is a symbol in a known symbol set. In general, the system processes an acquired input data set representative of the input expression, to form a set of segments, which are then used to specify a set of consegmentations. Each consegmentation and each possible interpretation for the input expression is represented in a data structure. The data structure is graphically representable by a graph comprising a two-dimensional array of nodes arranged in rows and columns and selectively connected by directed arcs. Each path, extending through the nodes and along the directed arcs, represents one consegmentation and one possible interpretation for the input expression. All of the consegmentations and all of the possible interpretations for the input expression are represented by the set of paths extending through the graph. For each row of nodes in the graph, a set of scores is produced for the known symbol set, using a complex of optimally trained neural information processing networks. Thereafter the system computes an a posteriori probability for one or more symbol sequence interpretations. By deriving each a posteriori probability solely through analysis of the acquired input data set, highly reliable probabilities are produced for competing interpretations for the input expression.

System and method for automated interpretation of input expressions

Noisy images of graphical or pictorial records, such as hand-written postal addresses, are cleaned by a method of analyzing the shapes of pixel clusters. According to such method, a set of structuring kernels is used to extract primitive shapes. The outcome of this extraction process is used to determine whether each pixel, in turn, is to be treated as part of a image element such as an alphanumeric character.

Method of image enhancement using convolution kernels

A process for creating segments out of an arbitrary string of handwritten alphanumeric script is described, in which the contours of the image are defined by the path a ball or pointer follows when allowed to roll from the top and bottom of an image, down or up either side. From the contours, the initial image cut points are determined. The pointer is provided with a capability to measure ink density in the nearby pixels. A grey scale threshold control is provided which operates in conjunction with the pointer as it rolls or moves, to define ink density above the threshold as a white pixel wherein no image content is present; and ink density below the threshold as a black pixel wherein image content is present.

Alphanumeric image segmentation scheme

Apparatus and processes are described for the automatic recognition of alphanumeric images. A set of cuts are made to the image which include incorrect segmentations. The resulting 'cells' comprising in their totality the created segments of the image are then analyzed to determine which cells are legal neighbors and which are not. All cells which are legal neighbors are then presented as connected nodes. A pruning of nodes which are related to certain predetermined image cuts is effected. Each set of remaining connected nodes is then presented to a recognizer which identifies the image and assigns a specified probability to the output. Many cells which are not legal neighbors are thereby not presented to the recognizer, thus saving substantially on computations per recognized image.

Growing community of inventors

Freehold, NJ, United States of America

Christopher John Burges