초록 ▼

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

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.

대표청구항 ▼

The invention claimed is: 1. A method for accelerating neural network execution in a speech-recognition system, for recognizing words contained in a subset of a general vocabulary of words that the same system is capable of recognizing, said neural network comprising a number of computing units org

The invention claimed is: 1. A method for accelerating neural network execution in a speech-recognition system, for recognizing words contained in a subset of a general vocabulary of words that the same system is capable of recognizing, said neural network comprising a number of computing units organized in levels including at least one hidden level and one output level, the computing units of said hidden level being connected to the computing units of said output level via weighted connections, said computing units of said output level corresponding to acoustic-phonetic units of said general vocabulary, said acoustic-phonetic units comprising stationary units and transition units, the method comprising the following steps: determining a subset of the acoustic-phonetic units to always include all of said stationary units and only include those of said transition units that are necessary for recognizing all the words contained in said general vocabulary subset; eliminating from the neural network all the weighted connections afferent to computing units of said output level that correspond to acoustic-phonetic units not contained in said previously determined subset of said acoustic-phonetic units, thus obtaining a compacted neural network; and executing, at each moment in time, only said compacted neural network. 2. The method according to claim 1 further comprising determining the transition units present in said general vocabulary subset. 3. The method according to claim 2 wherein said general vocabulary subset is the union of several subsets of vocabularies active at any given moment. 4. A method comprising: determining, from a set of acoustic-phonetic units connected to the output of a neural network and configured to model a set of sounds, a subset of the set of acoustic-phonetic units configured to model a subset of the set of sounds, wherein the neural network comprises at least one output neuron connected to each acoustic-phonetic unit of the set, each output neuron receives a plurality of weighted inputs, the set of acoustic-phonetic units includes stationary units and transition units, and the subset always includes all of the stationary units and only those of the transition units necessary for recognizing all the words contained in the subset; executing the neural network such that only the weighted inputs that are connected to the output neurons that are connected to acoustic-phonetic units in the subset are computed. 5. The method of claim 4, wherein the executing the neural network further comprises computing only the output neurons that are connected to acoustic-phonetic units in the subset. 6. The method of claim 4, wherein the executing occurs in a speech recognition application.