Time series signal analyzer including neural network having path groups corresponding to states of Markov chains

• US 5,381,513 A
• Filed: 03/16/1994
• Issued: 01/10/1995
• Est. Priority Date: 06/19/1991
• Status: Expired due to Term

• Current Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Original Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Inventors: Tsuboka, Eiichi
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Doerrler, Michelle
• Application Number:
  US08/213,480
• Time in Patent Office: 300 Days
• Field of Search: 395/2.11, 395/2.41, 395/2.64-2.66
• US Class Current: 704/232
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 15/16 using artificial neural networks

Abstract

A neural network with a high recognition rate when applied to static patterns is made applicable to dynamic time series patterns such as voice signals. Plural units with one or more inputs and outputs are interconnected, and a unique load coefficient is assigned to each connection to weight the signals flowing through that connection. The neural network includes an input unit group to which are input the components of plural vectors included in the input feature vector series {y(t)}; an output unit which outputs the converted vectors, which are produced by passing the input vectors through each unit and the associated connections; and J paths from input unit group to the output unit group. The units are connected to form a Hidden Markov Model wherein each signal path identified as j=1, 2, . . . , J corresponds to the same state.

Claims

1. A signal analyzer comprising a neural network of plural interconnected units with one or more inputs and outputs, and a unique weighting coefficient assigned to each connection to weight the signals flowing through, and comprising

an input unit group to which are input the components of plural vectors included in the input feature vector series {y(t)},

an output unit group which outputs the converted vectors, which are converted by passing the components of input vectors to the input unit through each unit and associated connections,

and the connections from the input unit group to the output unit group are grouped into a specific number of overlapping path groups, and each path group corresponds to the states or state transitions of a Markov chain.

2. A signal analyzer according to claim 1 where an output value of the neural network path group j (j=1, 2, . . . , J) is a predictive value of the path group j for y(t) or a value related to the predictive value.

3. A signal analyzer according to claim 2, comprising a Hidden Markov Model corresponding to recognition categories C₁, C₂, . . . , C_M, and computing equation (4) ##EQU64## where λ

_m is the set of Hidden Markov Model parameters for category C_m, for recognition pattern Y to return C_m as a recognition result.

4. A signal analyzer according to claim 2, comprising a parameter estimating means, which comprises a

likelihood calculation means for computing equation (6) ##EQU65## for E.sup.(r) (X.sup.(r)) for the r-th training pattern when the set of parameters of the circuit network is defined as the weighting coefficient w at a certain stage of the sequential estimating process during sequential estimating of the parameters of category C for the observation patterns Y.sup.(1), Y.sup.(2), . . . , Y.sup.(R) (where Y.sup.(r) ={Y.sup.(r) (t)}) associated with category C, and

a parameter correction quantity calculation means for computing the correction amount Δ

w that should increase the likelihood,

for estimating the parameters of category C by repeating the operation of substituting the value w+Δ

w for w and calculating the likelihood of the new w for the observation pattern Y.sup.(r+1) by means of the likelihood calculation means starting from a suitable value for w and repeating the operation for r=1, 2, . . . , R until a predetermined condition is satisfied.

5. A signal analyzer according to claim 2, comprising a dynamic programming means for inputting the input feature vector series Y to a model, and obtaining by a dynamic programming method the minimum cumulative value of the predictive error power based on a predictive model for the feature vector series Y.

6. A signal analyzer according to claim 1, further comprising a probability distribution parameter storage means for storing the parameters defining the probability distribution of a predictive error vector e_j (t) for the input feature vector series Y={y(t)} based on the path group j of the neural network, and

a likelihood calculation means for calculating the likelihood of model 1 for the input feature vector series Y from an observation density of each state or each state transition calculated using the parameters stored in the probability distribution parameter storage means by applying equation (1) ##EQU66## where X is a state series X=x{t} produced by the model and E(X)={e_x(t) (t)}.

7. A signal analyzer according to claim 1, further comprising a probability distribution parameter storage means for storing the parameters defining the probability distribution of a predictive error vector e_j (t) for the input vector series Y={y(t)} based on the path group j of the neural network, and

a likelihood calculation means for calculating the likelihood of model 1 for the input vector series Y from the probability density of each state or each state transition calculated using the parameters stored in the probability distribution parameter storage means by applying either equation (2) or (3) ##EQU67## wherein X is the state series X=x{t} produced by the model and E(X)={e_x(t) (t)}.