EP0318858A2
Expired / Lapsed
EP
⏱️ 0.0 years remaining
Connected word recognition system including neural networks arranged along a signal time axis
Assignee: NEC Corp
•
Filing: 1988-11-25
•
Granted: —
•
Expires: 2008-11-25
Abstract
Abstract For use in recognizing an input pattern consisting of feature vectors positioned at a first, "" an i-th, ..., and an I-th pattern time instant along a pattern time axis, a connected word recognition system comprises first through N-th neural networks B(1) to B(N) which are assigned to reference words identified by a first,..., an n-th, ..., and an N-th word identifier and are arranged along a signal time axis divisible into a first, ..., a j-th, ..., and a J-th signal time instant. The n-th word identifier n corresponds to consecutive ones of the pattern time instants by a first function n(i). The time axes are related to each other by a second function j(i). Among various loci (n(i), j(i)) in a space defined by the word identifiers and the time axes, an optimum locus (n̂(i), ĵ(i)) is determined to maximize a summation of output signals of the respective neural networks when the consecutive pattern time instants are varied between the first and the I-th pattern time instants for each word identifier. The input pattern is recognized as a concatenation of word identifiers used as optimum first functions n(i) in the optimum locus. Preferably, the optimum locus is determined by a dynamic programming algorithm. If possible, use of a finite-state automaton is more preferred.
Inventors
- 👤 Hiroaki C/O Nec Corporation Sakoe
Bibliographic Details
Application Number
—
Jurisdiction
EP
Data Source
google_patents
Last Synchronized
2026-10-30 07:08
CPC Classifications
G; G06; G06N; G06N3/00; G06N3/02; G06N3/06; G06N3/063; G06F; G06F18/00; G06F18/20; G06F18/29; G10; G10L; G10L15/00; G10L15/08; G10L15/16
Claims Analysis
Total Claims: 9
Claim 1
Independent Claim
344 words
A speech recongition system comprising first through N-th neural networks for receiving an input pattern representative of connected words, said first through the N-th neural networks representing first through N-th reference words including an n-th reference word, where N represents a first predetermined natural number, n being variable between 1 and said first predetermined natural number, both inclusive, said input pattern having a pattern time axis divided into first through I-th pattern time instants including an i-th pattern time instant, where I represents a second predetermined natural number, i being variable between 1 and said second predetermined natural number, both inclusive, each of said neural networks comprising input neuron units of an input layer, intermediate neuron units of an intermediate layer, and an output neuron unit, the input and the intermediate neuron units of each of said neural networks having a time sequential structure, the time sequential structures of said first through said N-th neural networks having a signal time axis divided into first through J-th signal time instants including a j-th signal time instant, where J represents a third predetermined natural number, j being variable between 1 and said third predetermined natural number, both inclusive, said n-th reference word being related to a portion of said input pattern by a first function n(i) defined by a plurality of consecutive ones of said first through said I-th pattern time instants, said pattern and said signal time axes being related to each other by a second function j(i) for each of said first through said N-th reference words, said speech recognition system further comprising determining means for determining an optimum locus (n(i), j(i)) in a space defined by said first through said N-th reference words and said pattern and said signal time axes so that said optimum locus maximizes a summation of output signals of the output neuron units of said first through said N-th neural networks when said plurality of pattern time instants are varied between said first through said I-th pattern time instants for each of said first through said N-th reference words.
Claim 2
Dependent Claim
344 words
A speech recongition system comprising first through N-th neural networks for receiving an input pattern representative of connected words, said first through the N-th neural networks representing first through N-th reference words including an n-th reference word, where N represents a first predetermined natural number, n being variable between 1 and said first predetermined natural number, both inclusive, said input pattern having a pattern time axis divided into first through I-th pattern time instants including an i-th pattern time instant, where I represents a second predetermined natural number, i being variable between 1 and said second predetermined natural number, both inclusive, each of said neural networks comprising input neuron units of an input layer, intermediate neuron units of an intermediate layer, and an output neuron unit, the input and the intermediate neuron units of each of said neural networks having a time sequential structure, the time sequential structures of said first through said N-th neural networks having a signal time axis divided into first through J-th signal time instants including a j-th signal time instant, where J represents a third predetermined natural number, j being variable between 1 and said third predetermined natural number, both inclusive, said n-th reference word being related to a portion of said input pattern by a first function n(i) defined by a plurality of consecutive ones of said first through said I-th pattern time instants, said pattern and said signal time axes being related to each other by a second function j(i) for each of said first through said N-th reference words, said speech recognition system further comprising determining means for determining an optimum locus (n(i), j(i)) in a space defined by said first through said N-th reference words and said pattern and said signal time axes so that said optimum locus maximizes a summation of output signals of the output neuron units of said first through said N-th neural networks when said plurality of pattern time instants are varied between said first through said I-th pattern time instants for each of said first through said N-th reference words.
Claim 3
Dependent Claim
37 words
A speech recognition system as claimed in Claim 1, wherein said second function defines for said i-th pattern time instant one of a predetermined number of consecutive ones of said first through said J-th signal time instants.
Claim 4
Dependent Claim
39 words
A speech recognition system as claimed in Claim 2, said predetermined number being equal to three, wherein said second function defines for said i-th pattern time instant one of the j-th, the (j-1)-th, and the (j-2)-th signal time instants.
Claim 5
Dependent Claim
46 words
A speech recognition system as claimed in Claim 1, wherein said second function is an inverse function of a mapping function which defines for said j-th signal time instant one of a predetermined number of consecutive ones of said first through said I-th pattern time instants.
Claim 6
Dependent Claim
39 words
A speech recognition system as claimed in Claim 4, said predetermined number being equal to three, wherein said mapping function defines for said j-th signal time instant one of the i-th, the (i-1)-th, and the (i-2)-th pattern time instants.
Claim 7
Dependent Claim
26 words
A speech recognition system as claimed in Claim 1, wherein said determining means determines said optimum locus with said first function controlled by a finite-state automaton.
Claim 8
Dependent Claim
27 words
A speech recongition system as claimed in Claim 1, wherein said determining means is put into operation of determining said optimum locus by a dynamic programming algorithm.
Claim 9
Dependent Claim
79 words
A speech recognition system as claimed in Claim 7, wherein said first through said N-th neural networks are collectively implemented by a combination of a coefficient memory and a coefficient processor, said coefficient processor being for receiving said input pattern, said determining means comprising a combination of a work area and a dynamic programming algorithm processor, said dynamic programming algorithm processor being connected to said coefficient processor to carry out said dynimac programming algorithm by using said work area.
References Cited (Backward Citations)
No citation records indexed.
USPTO Annuity & Maintenance Windows
Maintenance schedule is calculated for granted US utility patents.