Abstract: A method, computer program product, and computing system for inserting a spectral pooling layer into a neural network of a speech processing system. An output of a hidden layer of the neural network is filtered using the spectral pooling layer with a non-integer stride. The filtered output is provided to a subsequent hidden layer of the neural network.

Abstract: A method, computer program product, and computer system for processing one or more portions of an input sequence to generate one or more candidate output sequences, thus defining a plurality of prediction scores for the candidate output sequences. One or more specialized entities may be identified from the candidate output sequences. A first scoring methodology may be applied on the candidate output sequences based upon the portions of the input sequence, thus defining a first set of prediction scores for the one or more candidate output sequences. A second scoring methodology may be applied on the specialized entities from the candidate output sequences based upon the portions of the input sequence, thus defining a second set of prediction scores for the specialized entities. The plurality of predictions scores for the specialized entities may be at least partially modified based upon the first set and the second set of prediction scores.

Abstract: Methods and apparatus for training a multi-layer artificial neural network for use in speech recognition. The method comprises determining for a first speech pattern of the plurality of speech patterns, using a first processing pipeline, network activations for a plurality of nodes of the artificial neural network in response to providing the first speech pattern as input to the artificial neural network, determining based, at least in part, on the network activations and a selection criterion, whether the artificial neural network should be trained on the first speech pattern, and updating, using a second processing pipeline, network weights between nodes of the artificial neural network based, at least in part, on the network activations when it is determined that the artificial neural network should be trained on the first speech pattern.

Abstract: Training a neural network is a time consuming and computationally expensive task. Embodiments provide efficient methods and systems for neural network training One example embodiment is implemented by a plurality of agents, where each agent performs a pipelined gradient analysis to update respective local models of the neural network using respective subsets of data from a common pool of training data. In turn, a common global model of the neural network is updated based upon the local models.

Abstract: Methods and apparatus for training a multi-layer artificial neural network for use in speech recognition. The method comprises determining for a first speech pattern of the plurality of speech patterns, using a first processing pipeline, network activations for a plurality of nodes of the artificial neural network in response to providing the first speech pattern as input to the artificial neural network, determining based, at least in part, on the network activations and a selection criterion, whether the artificial neural network should be trained on the first speech pattern, and updating, using a second processing pipeline, network weights between nodes of the artificial neural network based, at least in part, on the network activations when it is determined that the artificial neural network should be trained on the first speech pattern.

Abstract: A neural network in a speech-recognition system has computing units organized in levels including at least one hidden level and one output level. The computing units of the hidden level are connected to the computing units of the output level via weighted connections, and the computing units of the output level correspond to acoustic-phonetic units of the general vocabulary. This network executes the following steps: determining a subset of acoustic-phonetic units necessary for recognizing all the words contained in the general vocabulary subset; eliminating from the neural network all the weighted connections afferent to computing units of the output level that correspond to acoustic-phonetic units not contained in the previously determined subset of acoustic-phonetic units, thus obtaining a compacted neural network optimized for recognition of the words contained in the general vocabulary subset; and executing, at each moment in time, only the compacted neural network.

Abstract: A method of optimizing the execution of a neural network in a speech recognition system provides for conditionally skipping a variable number of frames, depending on a distance computed between output probabilities, or likelihoods, of a neural network. The distance is initially evaluated between two frames at times 1 and 1+k, where k is a predetermined maximum distance between frames, and if such distance is sufficiently small, the frames between times 1 and 1+k are calculated by interpolation, avoiding further executions of the neural network. If, on the contrary, such distance is not small enough, it means that the outputs of the network are changing quickly, and it is not possible to skip too many frames. In that case, the method attempts to skip remaining frames, calculating and evaluating a new distance.

Abstract: A method of optimizing the execution of a neural network in a speech recognition system provides for conditionally skipping a variable number of frames, depending on a distance computed between output probabilities, or likelihoods, of a neural network. The distance is initially evaluated between two frames at times 1 and 1+k, where k is a predetermined maximum distance between frames, and if such distance is sufficiently small, the frames between times 1 and 1+k are calculated by interpolation, avoiding further executions of the neural network. If, on the contrary, such distance is not small enough, it means that the outputs of the network are changing quickly, and it is not possible to skip too many frames. In that case, the method attempts to skip remaining frames, calculating and evaluating a new distance.

Abstract: A method for accelerating neural network execution (4) in a speech recognition system, specifically for recognition of words contained in one or more subsets of a general vocabulary, involves the following steps.—at the recognition system initialisation phase, calculating the union of vocabulary subsets and determining the acoustic-phonetic units required for recognising the words contained in that union; re-compacting the neural network eliminating all the weighted connections afferent to computation output units corresponding to unnecessary acoustic-phonetic units;—executing unnecessary acoustic-phonetic units;—executing only the re-compacted network at each instant of time.

Abstract: A method of speeding up the execution of a wide class of neural networks for processing input signals evolving slowly through time, such as, for instance, voice, radar, sonar, video signals, and which requires no specialized, costly or hard-to-find hardware. The method requires storing, for the neurons in at least one level of the network, the activation value at a certain instant and comparing it with the one computed at the subsequent instant. If the activation is equal, the neuron carries out no activity, otherwise it propagates the difference in activation, multiplied by the interconnection weights, to the neurons it is connected to.

Abstract: A speech recognition apparatus in which the speech signal is digitalized and subjected to special analysis, word end detection is effected by energy analysis of the speech signal and the recognition system utilizes a Markov model in combination with a neural network learning by specific training steps.