Method for accelerating the execution of speech recognition neural networks and the related speech recognition device

• US 20050171766A1
• Filed: 02/12/2003
• Published: 08/04/2005
• Est. Priority Date: 02/28/2002
• Status: Active Grant

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: Loquendo SpA (Microsoft Corporation)
• Inventors: Albesano, Dario, Gemello, Roberto
• Granted Patent: US 7,827,031 B2
• Time in Patent Office: Days
• US Class Current: 704/202
• CPC Class Codes: G10L 15/16 using artificial neural networks

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.

Claims

1. Method for accelerating neural network (4) execution in a speech recognition system, for recognising words contained in a subset of a general vocabulary of words that the same system is capable of recognising, said neural network (4) comprising a number of computing units organised in levels, among which at least one hidden level (12) and one output level (14), the computing units (H_j) of said hidden level (12) being connected to the computing units (N_i) of said output level (14) via weighted connections (W_ij), said computing units (N_i) of said output level (14) corresponding to acoustic-phonetic units (2) of said general vocabulary, characterised in that it comprises the following steps:

determining a subset of acoustic-phonetic units necessary for recognising all the words contained in said general vocabulary subset;

eliminating from the neural network (4) all the weighted connections (W_ij) afferent to computing units (N_i) of said output level (14) that correspond to acoustic-phonetic units not contained in said previously determined subset of acoustic-phonetic units, thus obtaining a compacted neural network (4′

) optimised for recognition of the words contained in said general vocabulary subset;

executing, at each moment in time, only said compacted neural network (4′

).

2. Method according to claim 1, in which said acoustic-phonetic units (2) comprise stationary units (2a, 2c) and transition units (2b), and said step of determining a subset of acoustic-phonetic units consists of determining the stationary units (2a, 2c) and transition units (2b) present in said general vocabulary subset.

3. Method according to claim 1, in which said acoustic-phonetic units (2) comprise stationary units (2a, 2c) and transition units (2b), and said step of determining a subset of acoustic-phonetic units consists of selecting all the stationary units (2a, 2c) and determining the transition units (2b) present in said general vocabulary subset.

4. Method according to claim 2 or 3, in which said general vocabulary subset is the union of several subsets of vocabularies active at any given moment.

5. Speech recognition system, comprising a neural network (4) with a number of computing units organised in levels, including at least one hidden level (12) and one output level (14), the computing units (H_j) of said hidden level (12) being connected to the computing units (N_i) of said output level (14) via weighted connections (W_ij), said computing units (N_i) of said output level (14) corresponding to acoustic-phonetic units (2) of a general vocabulary of words to be recognized, characterised in that it comprises means (18, 16) for accelerating neural network (4) execution, for recognising words contained in a subset of said general vocabulary, said means (18, 16) comprising:

a first module (18) for determining the subset of acoustic-phonetic units necessary for recognising all the words contained in said general vocabulary subset;

a second module (16) for selecting, from among the weighted connections (W_ij) connecting the computing units (H_j) of hidden level (12) with those of output level (14), the weighted connections afferent to computing units (N_i) corresponding to acoustic-phonetic units contained in said subset of acoustic-phonetic units determined by said first module (16), thus obtaining a compacted neural network (4′

) optimised for recognition of the words contained in said general vocabulary subset.

6. System according to claim 5, in which the acoustic-phonetic units (2) comprise stationary units (2a, 2c) and transition units (2b), and said first module (18) comprises the means for determining both the stationary units (2a, 2c) and transition units (2b) present in said general vocabulary subset.

7. System according to claim 5, in which said acoustic-phonetic units (2) comprise stationary units (2a, 2c) and transition units (2b), and said first module (18) comprises means for selecting all the stationary units (2a, 2c) and determining the transition units (2b) present in said general vocabulary subset.

8. System according to claim 6 or 7, in which said general vocabulary subset is the union of several subsets of vocabularies active at any given moment.