At+t Corp.

Samsung Electronics Co., Ltd.

Imagia Cybernetics Inc.

Yoshua Bengio

Leon Bottou

Paul Glor Howard

Yann Andre LeCun

Sang Hyun Yoo

Sanghyun Yoo

Inchul Song

Nicolas Chapados

Nicolas Guizard

Mohammad Havaei

A speech recognition method and apparatus, including implementation and/or training, are disclosed. The speech recognition method includes obtaining a speech signal, and performing a recognition of the speech signal, including generating a dialect parameter, for the speech signal, from input dialect data using a parameter generation model, applying the dialect parameter to a trained speech recognition model to generate a dialect speech recognition model, and generating a speech recognition result from the speech signal by implementing, with respect to the speech signal, the dialect speech recognition model. The speech recognition method and apparatus may perform speech recognition and/or training of the speech recognition model and the parameter generation model.

Speech recognition method and apparatus

A unit is disclosed for generating combined feature maps in accordance with a processing task to be performed, the unit comprising a feature map generating unit for receiving more than one modality and for generating more than one corresponding feature map using more than one corresponding transformation; wherein the generating of each of the more than one corresponding feature map is performed by applying a given corresponding transformation on a given corresponding modality, wherein the more than one corresponding transformation is generated following an initial training performed in accordance with the processing task to be performed and a combining unit for selecting and combining the corresponding more than one feature map generated by the feature map generating unit in accordance with at least one combining operation and for providing at least one corresponding combined feature map; wherein the combining unit is operating in accordance with the processing task to be performed and the combining operation reduces each corresponding numeric value of each of the more than one feature map generated by the feature map generation unit down to one numeric value in the at least one corresponding combined feature map.

Method and system for processing a task with robustness to missing input information

A binary arithmetic coder and decoder provides improved coding accuracy due to improved probability estimation and adaptation. They also provide improved decoding speed through a &ldquo;fast path&rdquo; design wherein decoding of a most probable symbol requires few computational steps. Coded data represents data that is populated by more probable symbols (&ldquo;MPS&rdquo;) and less probable symbols (&ldquo;LPS&rdquo;). In an embodiment, a decoder receives a segment of the coded data as a binary fraction C. It defines a coding interval of possible values of C, the interval extending from a variable lower bound A to a constant upper bound 1. For each position in the decoded symbol string, the decoder computes a test value Z that subdivides the coding interval into sub-intervals according to the relative probabilities that an MPS or an LPS occurs in the position. A first sub-interval extends from the lower bound A to the test value z; the second sub-interval extending from the test value Z to 1. If C is greater than Z, the decoder emits an MPS for the current position in the decoded symbol string and sets the lower bound A to the test variable Z for use during decoding of the next position in the decoded symbol string. If C is less than Z, the decoder emits an LPS and computes a new lower bound A and a new binary fraction C for use during decoding of the next position in the decoded symbol string. The encoder operates according to analogous techniques to compose coded data from original data.

Z-coder: a fast adaptive binary arithmetic coder

A binary arithmetic coder and decoder provides improved coding accuracy due to improved probability estimation and adaptation. They also provide improved decoding speed through a &ldquo;fast path&rdquo; design wherein decoding of a most probable symbol requires few computational steps. Coded data represents data that is populated by more probable symbols (&ldquo;MPS&rdquo;) and less probable symbols (&ldquo;LPS&rdquo;). In an embodiment, a decoder receives a segment of the coded data as a binary fraction C. It defines a coding interval of possible values of C, the interval extending from a variable lower bound A to a constant upper bound,For each position in the decoded symbol string, the decoder computes a test value Z that subdivides the coding interval into sub-intervals according to the relative probabilities that an MPS or an LPS occurs in the position. A first sub-interval extends from the lower bound A to the test value Z; the second sub-interval extending from the test value Z to,If C is greater than Z, the decoder emits an MPS for the current position in the decoded symbol string and sets the lower bound A to the test variable Z for use during decoding of the next position in the decoded symbol string. If C is less than Z, the decoder emits an LPS and computes a new lower bound A and a new binary fraction C for use during decoding of the next position in the decoded symbol string. The encoder operates according to analogous techniques to compose coded data from original data.

A binary arithmetic coder and decoder provides improved coding accuracy due to improved probability estimation and adaptation. They also provide improved decoding speed through a &ldquo;fast path&rdquo; design wherein decoding of a most probable symbol requires few computational steps. Coded data represents data that is populated by more probable symbols (&ldquo;MPS&rdquo;) and less probable symbols (&ldquo;LPS&rdquo;). In an embodiment, a decoder receives a segment of the coded data as a binary fraction C. It defines a coding interval of possible values of C, the interval extending from a variable lower bound A to a constant upper bound,. For each position in the decoded symbol string, the decoder computes a test value Z that subdivides the coding interval into sub-intervals according to the relative probabilities that an MPS or an LPS occurs in the position. A first sub-interval extends from the lower bound A to the test value Z; the second sub-interval extending from the test value Z to 1. If C is greater than Z, the decoder emits an MPS for the current position in the decoded symbol string and sets the lower bound A to the test variable Z for use during decoding of the next position in the decoded symbol string. If C is less than Z, the decoder emits an LPS and computes a new lower bound A and a new binary fraction C for use during decoding of the next position in the decoded symbol string. The encoder operates according to analogous techniques to compose coded data from original data.

A binary arithmetic coder and decoder provides improved coding accuracy due to improved probability estimation and adaptation. They also provide improved decoding speed through a &ldquo;fast path&rdquo; design wherein decoding of a most probable symbol requires few computational steps. Coded data represents data that is populated by more probable symbols (&ldquo;MPS&rdquo;) and less probable symbols (&ldquo;LPS&rdquo;). In an embodiment, a decoder receives a segment of the coded data as a binary fraction C. It defines a coding interval of possible values of C, the interval extending from a variable lower bound A to a constant upper bound 1. For each position in the decoded symbol string, the decoder computes a test value Z that subdivides the coding interval into sub-intervals according to the relative probabilities that an MPS or an LPS occurs in the position. A first sub-interval extends from the lower bound A to the test value Z; the second sub-interval extending from the test value Z to 1. If C is greater than Z, the decoder emits an MPS for the current position in the decoded symbol string and sets the lower bound A to the test variable Z for use during decoding of the next position in the decoded symbol string. If C is less than Z, the decoder emits an LPS and computes a new lower bound A and a new binary fraction C for use during decoding of the next position in the decoded symbol string. The encoder operates according to analogous techniques to compose coded data from original data.

Z-coder: fast adaptive binary arithmetic coder

A machine learning paradigm called Graph Transformer Networks extends the applicability of gradient-based learning algorithms to systems composed of modules that take graphs as inputs and produce graphs as output. Training is performed by computing gradients of a global objective function with respect to all the parameters in the system using a kind of back-propagation procedure. A complete check reading system based on these concept is described. The system uses convolutional neural network character recognizers, combined with global training techniques to provides record accuracy on business and personal checks.

Growing community of inventors

Montreal, Canada

Yoshua Bengio