Neural network, a method of learning of a neural network and phoneme recognition apparatus utilizing a neural network

• US 6,026,358 A
• Filed: 12/21/1995
• Issued: 02/15/2000
• Est. Priority Date: 12/22/1994
• Status: Expired due to Fees

• Current Assignee: Justsystems Corporation
• Original Assignee: Justsystems Corporation
• Inventors: Tomabechi, Hideto
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/576,585
• Time in Patent Office: 1,517 Days
• Field of Search: 395/2.41, 395/2.31, 395/2.28, 395/2.32, 395/2.12, 395/2.58, 395/23, 395/24, 395/2.63, 704/232, 704/222, 704/219, 704/223, 704/203, 704/249
• US Class Current: 704/232
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G10L 15/16 using artificial neural networks

Abstract

A neuron device network is provided with a speech input layer, a context layer, a hidden layer, a speech output layer and a hypothesis layer. A phoneme to be learned is spectral-analyzed by an FFT unit and a vector row at a time point t is input to a speech input layer. Also, a vector state of the hidden layer at a time t-1 is input to the context layer, the vector row at a time t+1 is input to the speech output layer as an instructor signal, and a code row for hypothesizing the phoneme, or the code row, is input to the hypothesis layer. The time series relation of the vector rows and the phoneme are hypothetically learned. Alternatively, a spectrum, a cepstrum or a speech vector row based on outputs from the hidden layer of an auto-associative neural network is input to the speech input layer, and the code row is output from the hypothesis layer, taking into account the time series relation. The speech is recognized when a CPU reads the stored output values of the hidden layer and the connection weights of the hidden layer and the hypothesis layer from a memory of the neuron device network and calculates output values of the respective neuron devices of the hypothesis layer based on the output values and the connection weights. The corresponding phoneme is determined by collating the output values of the respective neuron devices of the hypothesis layer with the code rows in an instructor signal table.

Claims

1. A learning method of a neural network comprising:

inputting first vector rows representing data to a data input layer;

inputting second vector rows as a first instructor signal to a first output layer, the first instructor signal comprising vector rows input to the data input layer at a time after inputting the first vector rows; and

inputting a definite meaning as a second instructor signal to a second output layer, the second instructor signal being determined in accordance with vector rows input to the data input layer at times before and after the first vector rows are input to the data input layer, and

performing learning for the data by having a plurality of first vector rows represent the definite meaning.

2. The learning method of a neural network according to claim 1 wherein learning is performed in accordance with error back-propagation.

3. A learning method of a neural network comprising:

inputting first vector rows representing data to a data input layer;

inputting second vector rows as a first instructor signal to a first output layer; and

inputting a definite meaning as a second instructor signal to a second output layer, the definite meaning is phoneme data representing a speech element and the plurality of first vector rows represent characteristics of the definite meaning analyzed in a time series; and

performing learning for the data by having a plurality of first vector rows represent the definite meaning.

4. A learning method of a neural network according to claim 3, wherein one of spectral data, cepstrum data of the speech, and an output value data of the hidden layer of an auto-associative neural network is used as the plurality of first vectors representing the characteristics of the definite meaning.

5. A learning method of a neural network comprising:

inputting, to a feedback input layer, output vector values of a hidden layer having a plurality of neuron devices, the plurality of neuron devices corresponding to first vector rows of the feedback input layer, the feedback input layer connected with the hidden layer and having a number of neuron devices equal to a number of neuron devices of the hidden layer;

inputting second vector rows representing data to a data input layer;

inputting third vector rows as a first instructor signal to a first output layer, the first instructor signal comprising vector rows input to the data input layer at a time after inputting the first vector rows; and

inputting a definite meaning as a second instructor signal to a second output layer, the second instructor signal being determined in accordance with vector rows input to the data input layer at times before and after the first vector rows are input to the data input layer; and

performing learning for the data by having a plurality of second vector rows represent the definite meaning.

6. The learning method of a neural network according to claim 5, wherein learning is performed in accordance with error back-propagation.

7. A learning method of a neural network comprising:

inputting to a feedback input layer, output vector values of a hidden layer having a plurality of neuron devices, the plurality of neuron devices corresponding to first vector rows of the feedback input layer, the feedback input layer connected with the hidden layer and having a number of neuron devices equal to a number of neuron devices of the hidden layer;

inputting second vector rows representing data to a data input layer;

inputting third vector rows as a first instructor signal to a first output layer, and

inputting a definite meaning as a second instructor signal to a second output layer; and

performing learning for the data by having a plurality of second vector rows represent the definite meaning, wherein the definite meaning is phoneme data representing a speech element and the plurality of second vector rows represent characteristics of the definite meaning analyzed in a time series.

8. A learning method of a neural network according to claim 7, wherein one of spectral data, cepstrum data of the speech, and an output value data of the hidden layer of an auto-associative neural network is used as the plurality of second vector rows representing the characteristics of the definite meaning.

9. A learning method of a neural network comprising:

inputting output vector values of a first output layer corresponding to first vector rows of a feedback input layer, the feedback input layer connected to a hidden layer that has neuron devices whose number is equal to that of neuron devices of the first output layer;

inputting second vector rows to a data input layer;

inputting third vector tows as a first instructor signal to the first output layer, the first instructor signal comprising vector rows input to the data input layer at a time after inputting the first vector rows; and

inputting a definite meaning as a second instructor signal to a second output layer, the second instructor signal being determined in accordance with vector rows input to the data input layer at times before and after the first vector rows are input to the data input layer,

wherein learning is carried out for data having a plurality of the second vector rows representing the definite meaning.

10. A learning method of a neural network according to claim 9, wherein learning is performed in accordance with error back-propagation.

11. A learning method of a neural network comprising:

inputting output vector values of a first output layer corresponding to first vector rows of a feedback input layer, the feedback input layer connected with a hidden layer and having neuron devices whose number is equal to that of neuron devices of the first output layer;

inputting second vector rows to a data input layer;

inputting third vector rows as a first instructor signal to the first output layer; and

inputting a definite meaning as a second instructor signal to a second output layer, wherein the definite meaning is phoneme data representing a speech element and the plurality of second vector rows represent characteristics of the definite meaning analyzed in a time series,

wherein learning is carried out for data having a plurality of the second vector rows representing the definite meaning.

12. A learning method of a neural network according to claim 11, wherein one of spectral data, cepstrum data of the speech, and an output value data of the hidden layer of an auto-associative neural network is used as the plurality of second vector rows representing the characteristics of the definite meaning.

13. A neural network comprising:

a neuron device network having a data input layer, a hidden layer connected to said data input layer, and an output layer connected to said hidden layer, said output layer comprising a first output layer and a second output layer;

learning means in said neuron device network for learning about data having a plurality of first vector rows representing a definite meaning, said learning means inputting said plurality of first vector rows to said data input layer, inputting second vector rows as a first instructor signal to said first output layer and inputting said definite meaning as a second instructor signal to said second output layer;

inputting means for inputting said plurality of first vector rows to said data input layer of said neuron device network; and

outputting means for outputting output signals of said second output layer based on input of said plurality of first vector rows by said inputting means.

14. The neural network according to claim 13, further comprising a plurality of hidden layers at least equal to a number of definite meanings;

wherein said data input layer and said output layer are connected with each other, and values corresponding with said output signals relative to a plurality of third vector rows are fed back and input to respective hidden layers.

15. The neural network according to claim 13, wherein said data input layer, said hidden layer and said second output layer each have connection weights between respective neuron devices.

16. A neural network comprising:

a neuron device network comprising an input layer having a data input layer and a feedback input layer;

a hidden layer connected to said input layer; and

an output layer connected to said hidden layer, said output layer having a first output layer and a second output layer;

learning means in said neuron device network for learning about data having a plurality of first vector rows representing a definite meaning by;

(1) inputting a plurality of second vector values of said hidden layer or said first output layer, to said input layer,

(2) inputting said plurality of first vector rows to said data input layer of the input layer,

(3) inputting a plurality of third vector rows as a first instructor signal to said first output layer, and

(4) inputting said definite meaning as a second instructor signal to said second output layer;

inputting means for inputting said plurality of first vector rows to said data input layer of said neuron device network such that said learning means performs said learning; and

outputting means for outputting output signals of said second output layer based on input of said plurality of first vector rows by said inputting means,

said input layer having a plurality of neuron devices equal in number to a number of neuron devices of said hidden layer or said first output layer.

17. The neural network according to claim 16, wherein said data input layer, said feedback input layer, said hidden layer and said second output layer each have connection weights between respective neuron devices.

18. A speech recognition apparatus comprising:

a neural network, said neural network including an input layer comprising a speech input layer and a context layer, a hidden layer, and an output layer comprising a speech output layer and a hypothesis layer;

speech inputting means for inputting speech;

analyzing means for analyzing in a time-series, vector rows representing characteristics of the speech input by said speech inputting means;

vector row inputting means for successively inputting said vector rows analyzed by said analyzing means to said input layer of said neural network; and

phoneme specifying means for specifying a phoneme in accordance with outputs of said output layer of said neural network by successively inputting said vector rows to said data input layer by said vector row inputting means.

19. The neural network according to claim 18, wherein said analyzing means uses one of spectral data, cepstrum data of the speech, and an output value data of a hidden layer of an auto-associative neural network as said vector rows representing said characteristics of the speech.