At&t Intellectual Property I, LP

Lucent Technologies Inc.

At+t Corp.

Nuance Communications, Inc.

Richard William Sproat

Murat Saraclar

Robert E Coyne

Jianying Hu

Hao Chen

Michael Dennis Riley

Michiel A U Bacchiani

Sameer Maskey

Fernando Carlos Pereira

Jan Pieter VanSanten

Brian E Roark

Sameer Raj Maskey

A system and method are disclosed for retrieving audio segments from a spoken document. The spoken document preferably is one having moderate word error rates such as telephone calls or teleconferences. The method comprises converting speech associated with a spoken document into a lattice representation and indexing the lattice representation of speech. These steps are performed typically off-line. Upon receiving a query from a user, the method further comprises searching the indexed lattice representation of speech and returning retrieved audio segments from the spoken document that match the user query.

System and method of lattice-based search for spoken utterance retrieval

The invention relates to a method of converting a set of words into a three-dimensional scene description, which may then be rendered into three-dimensional images. The invention may generate arbitrary scenes in response to a substantially unlimited range of input words. Scenes may be generated by combining objects, poses, facial expressions, environments, etc., so that they represent the input set of words. Poses may have generic elements so that referenced objects may be replaced by those mentioned in the input set of words. Likewise, a character may be dressed according to its role in the set of words. Various constraints for object positioning may be declared. The environment, including but not limited to place, time of day, and time of year, may be inferred from the input set of words.

Text-to-scene conversion

Systems and methods relate to generating a language model for use in, for example, a spoken dialog system or some other application. The method comprises building a class-based language model, generating at least one sequence network and replacing class labels in the class-based language model with the at least one sequence network. In this manner, placeholders or tokens associated with classes can be inserted into the models at training time and word/phone networks can be built based on meta-data information at test time. Finally, the placeholder token can be replaced with the word/phone networks at run time to improve recognition of difficult words such as proper names.

System and method of using meta-data in speech processing

A system for generating a scene description from a set of words includes modules that perform the following functions: performing a linguistic analysis on a set of words to generate a structure representation of semantic relationships between words within the set of words, converting the structure to a set of description elements representative of objects to be depicted in a scene and relationships between objects, depicting a spatial relationship by using at least one spatial tag of an associated one or more of the object, and assigning a set of depictors to each of the description elements. Each depictor in the set of depictors may include a reference to an object to be modified, parameters used in modifying the object and a procedure for modifying the object. The system may also include a module that generates the same description by execution of the procedure for the modification of the objects.

Text-to scene conversion

A system is disclosed for generating a scene description from a set of words. The system includes modules that perform the following functions: performing a linguistic analysis on a set of words to generate a structure representation of semantic relationships between words within the set of words, converting the structure to a set of description elements representative of objects to be depicted in a scene and relationships between objects, depicting a spatial relationship by using at least one spatial tag of an associated one or more of the object, and assigning a set of depictors to each of the description elements. Each depictor in the set of depictors may include a reference to an object to be modified, parameters used in modifying the object and a procedure for modifying the object. The system may also include a module that generates the same description by execution of the procedure for the modification of the objects.

The invention relates to a method of converting a set of words into a three-dimensional scene description, which may then be rendered into three-dimensional images. The invention may generate arbitrary scenes in response to a substantially unlimited range of input words. Scenes may be generated by combining objects, poses, facial expressions, environments, etc., so that they represent the input set of words. Poses may have generic elements so that referenced objects may be replaced by those mentioned in the input set of words. Likewise, a character may be dressed according to its role in the set of words. Various constraints for object positioning may be declared.

A technique analyzing loosely constrained text blocks, such as e-mail signature blocks by performing a two-dimensional geometrical analysis and a one-dimensional language analysis in order to classify sub-blocks at the loosely constrained text block into particular functional classes. The present technique may also be utilized to identify a personal name from a user name in a loosely constrained text block, such as an e-mail signature block.

E-mail signature block analysis

A technique for segmenting a loosely constrained text block, such as an e-mail signature block into sub-blocks by performing line segment extraction and connected component analysis on the foreground characters and background characters and recursively repeating connected component analysis on both the foreground and background characters and line segment extraction on the background characters until a text output includes no mixed reading blocks. A technique for correcting over segmentation errors in a line of text from a loosely constrained text block which has undergone geometrical analysis.

E-mail signature block segmentation

A system and apparatus are disclosed for identifying polysemous terms and for measuring their degree of polysemy. A polysemy index provides a quantitative measure of how polysemous a word is. A list of words can be ranked by their polysemy indices, with the most polysemous words appearing at the top of the list. A polysemy evaluation process collects a set of terms near a target term. Inter-term distances of the set of terms occurring near the target term are computed and the multi-dimensional distance space is reduced to two dimensions. The two dimensional representation is converted into radial coordinates. Isotonic/antitonic regression techniques are used to compute the degree to which the distribution deviates from unimodality. The amount of deviation is the polysemy index. A corpus can be preprocessed using the polysemy indices to identify words having clearly separated senses, allowing an information retrieval system to return a separate list of documents for each sense of a word. Self-organizing sense disambiguation techniques can use the polysemy indixces to select canonical contexts for the various senses identified for a given word. Contexts are selected containing terms falling in radial bins near each peak. Such contexts can then be used for subsequent training of a classifier.

Method and apparatus for measuring the degree of polysemy in polysemous words

A method for automatically converting a decision tree into one or more weighted finite-state transducers. Specifically, the method in accordance with an illustrative embodiment of the present invention processes one or more terminal (i.e., leaf) nodes of a given decision tree to generate one or more corresponding weighted rewrite rules. Then, these weighted rewrite rules are processed to generate weighted finite-state transducers corresponding to the one or more terminal nodes of the decision tree. In this manner, decision trees may be advantageously compiled into weighted finite-state transducers, and these transducers may then be used directly in various speech and natural language processing systems. The weighted rewrite rules employed herein comprise an extension of conventional rewrite rules, familiar to those skilled in the art.

Compilation of weighted finite-state transducers from decision trees

The present invention provides a method of expanding a string of one or more digits to form a verbal equivalent using weighted finite state transducers. The method provides a grammatical description that expands the string into a numeric concept represented by a sum of powers of a base number system, compiles the grammatical description into a first weighted finite state transducer, provides a language specific grammatical description for verbally expressing the numeric concept, compiles the language specific grammatical description into a second weighted finite state transducer, composes the first and second finite state transducers to form a third weighted finite state transducer from which the verbal equivalent of the string can be synthesized, and synthesizes the verbal equivalent from the third weighted finite state transducer.

Growing community of inventors

Berkeley Heights, NJ, United States of America

Richard William Sproat