Verizon Corporate Services Group Inc.

Bbn Technologies Corporation

Bolt Beranek and Newman Inc.

Bbn Corporation

Herbert Gish

Salim Estephan Roukos

Marcos Antonio Bergamo

Robert David Preuss

Patrick M Peterson

William R Belfield

Alan G Derr

A W Huggins

Phillippe Jeanrenaud

Jan R Rohlicek

John W McDonough

Kenney Ng

Patrick Mangan Peterson

Alan Gregory Derr

A system is provided that trains a speech processing device to recognize any spoken language. The speech processing device is trained by receiving a group of speech frames (). The group of speech frames is grouped into a group of segments (). A model of each of the segments is created () and the segments are compared (). The segments are grouped into clusters based on comparisons of the segments (). A model for each of the clusters is built (). The built cluster models enable the labeling of segments from speech in the language. Labeled segments (i.e., recognizer output) may be used to train topic classifiers. Speech topics may be classified on the basis of the generated labeled segments.

Self-organizing speech recognition for information extraction

An improved topic discriminator includes an integrated speech recognizer or word and phrase spotter as part of a speech event detector, and a topic classifier trained on topic-dependent event frequencies. The event frequencies are determined from either or both transcribed data with a text event detector and untranscribed speech data with a speech event detector. In accordance with another aspect of the present invention, the phrase spotter is used to detect the presence of phrases without the need of parsing the output of a speech recognizer's hypothesized transcription. In accordance with another aspect of the invention, an improved technique is used to select a subset of the potential speech events on whose event frequencies the topic discrimination decision is made. Finally, in accordance with another aspect of the invention, an improved method of topic modeling is used to improve the performance of the topic discriminator.

Topic discriminator using posterior probability or confidence scores

A code signal representing a sequence of code symbols carrying digital information is generated with the frequency components of the code signal being essentially confined to a preselected signalling band lying within the bandwidth of an audio signal within which the code signal is to be embedded. The audio signal is continuously frequency analyzed over a frequency band encompassing the signalling band and the code signal is dynamically filtered as a function of the analysis thereby to provide a modified code signal with frequency components which are, at each time instant, essentially a preselected small proportion of the levels of the corresponding audio signal frequency components. Accordingly, the modified code signal can be combined with the audio signal to obtain a composite audio signal which is not readily distinguishable from the original audio signal by listening. Furthermore, the digital information can be recovered from the composite audio signal by a procedure which is essentially the compliment of that used to obtain the composite audio signal.

Embedded signalling

Growing community of inventors

Newton, MA, United States of America

Herbert Gish