JPH03230200A - Voice recognizing method - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition method suitable for recognizing words from input speech using online terminals such as electric locks and IC cards.

[Prior Art] The present applicant has proposed, in Japanese Patent Application No. 1-98376, a method for recognizing words from input speech using a neural network as a speech recognition method that can be easily processed in real time. In this speech recognition method, the frequency characteristics of the input speech are calculated as input to the neural network, and each speech interval divided equally in time in each band is treated as one block. The average of the frequency characteristics is calculated, and the average is normalized by the power of the word as a whole.

[Problem to be solved by the invention] However, in the case of the above-mentioned conventional technology, it is difficult to create a standard input cover pattern (learning input cover pattern) in order to construct a neural network, and to use the constructed neural network. If the conditions for creating evaluation cover patterns for speech recognition differ due to the introduction of stationary noise or differences in input systems such as lines, the recognition rate will drop.

This reduction in recognition rate is due to the inability to eliminate spectral distortion due to normalization across word powers, as analyzed below. That is, i is the block number, k is the band number, Ak is the frequency transmission characteristic of the band, Smik is the audio signal of the band i block in the learning stage, and Stik is the sound signal after passing through the telephone line in the evaluation stage. , when the speech signal of k-band i-block is spectrum-distorted due to the influence of the stationary frequency transmission characteristic Ak, 5tik=Ak-3■i, and each speech signal S tik at the evaluation stage is expressed as a word. Normalized across the power of S tik
A k S *ikΣΣS tik Σ
Since ΣA k S aik, the frequency transmission characteristic Ak on the right side cannot be eliminated, that is, the spectral distortion cannot be eliminated.

SUMMARY OF THE INVENTION An object of the present invention is to provide a speech recognition method that can easily perform real-time processing and ensure a high recognition rate.

[Means for Solving the Problems] The present invention according to claim 1 provides a word recognition method for recognizing a word from an input voice using a neural network, calculating the frequency characteristics of the input voice, and Each of the speech intervals divided equally in time is treated as one block, and the average frequency characteristics are calculated within each block, and the average is normalized for each corresponding band and sent to the neural network. It is designed to be used as an input.

According to a second aspect of the present invention, the neural network is a hierarchical neural network.

[Action] According to the present invention as set forth in claim 1, there are the following effects (1) to (2).

■Since "frequency characteristics" are used as the feature parameters input to the neural network, frequency analysis is simpler and more parallel than complex feature extraction such as LPC correlation or LPC cepstrum. I can do it,
The time required for the preprocessing is short.

■Neural networks are, in principle, capable of simple and quick arithmetic processing for the entire network.

■In principle, each unit that makes up a neural network operates independently, and parallel arithmetic processing is possible. Therefore, calculation processing is quick.

(2) With the above (2) to (4), voice recognition processing can be easily performed in real time without using a complicated processing device.

■It is resistant to constant spectral distortion and can maintain a high recognition rate. This is because, as will be analyzed below, the average of the frequency characteristics of each block of input audio is normalized within the same band, so that spectral distortion can be eliminated. That is, as mentioned above, l is the block number, k is the band number, and A
k is the frequency transmission characteristic of the band, Sik is the audio signal of the band i block in the learning stage, and Stik is the voice signal of the band i block in the evaluation stage due to the influence of the steady frequency transmission characteristic Ak, such as after passing through a telephone line. When the spectrum is distorted and it is a k-band i-block audio signal, S tik = A k-3mi
...(1). Then, each audio signal S tik at the evaluation stage is normalized for each band as follows: Σ 5tik Ak Σ S■ik Σ
The frequency transmission characteristic Ak can be eliminated, that is, the spectral distortion can be eliminated.

According to the present invention as set forth in claim 2, there is the following effect (2).

Currently, in a zero-layer neural network,
A simple learning algorithm (pack propagation) as described below has been established, and a neural network that can achieve a high recognition rate can be easily formed.

[Example] Fig. 1 is a schematic diagram showing an example of a speech recognition system to which the present invention is applied, Fig. 2 is a schematic diagram showing a neural network, and Fig. 3 is a schematic diagram showing a hierarchical neural network. , FIG. 4 is a schematic diagram showing the structure of Uni-Soto.

Prior to describing specific embodiments of the present invention, the configuration of the neural network and the learning algorithm will be described.

(1) Due to its structure, neural networks
It can be roughly divided into two types: a hierarchical network shown in FIG. 2(A) and an interconnected network shown in FIG. 2(B). Although the present invention may be constructed using either of these networks, the hierarchical network is more useful because a simple learning algorithm as described below has been established.

(2) Network Structure A hierarchical network has a hierarchical structure consisting of an input layer, an intermediate layer, and an output layer, as shown in FIG.

Each layer is composed of one or more units. The connections are forward connections from the input layer to the middle layer to the output layer, and there are no connections within each layer.

(3) Structure of the unit The unit is a model of a neuron in the brain and has a simple structure as shown in FIG. It receives input from other units, sums it up, transforms it using a certain rule (conversion function), and outputs the result. Each connection with another unit is given a variable weight that represents the strength of the connection.

(4) Learning (pack propagation) Learning of a network is to bring the actual output closer to the target value (desired output), and generally changes the conversion function and weight of each unit shown in Figure 4. Let them learn.

Also, as a learning algorithm, for example, Ru+5e
lhart, D.E., McClelland, J.
, L, and thePDP Re5earch G
roup, PARALLEL DISTRIBLIT
EDPROCESSING, the MIT Pre
ss, 1986. Pack propagation as described in .

Hereinafter, specific examples of the present invention will be described.

The recognition system 1 is composed of a 16-channel bandpass filter 11, an averaging circuit 12, a normalization circuit 13, a neural network 20, and a determination circuit 30 (see FIG. 1).

In this recognition system 1, the recognized words were the names of 47 prefectures, and the specific speaker was one person. The learning operation and evaluation operation of the recognition system 1 will be described in detail below.

(Learning) 1. Input creation ■ Pass the known input speech wave/form of each recognition word through the 16-channel bandpass filter 11 to calculate the frequency characteristics of the input speech.

■In each band of the band pass filter 11, each of the voice sections obtained by temporally dividing the voice waveform into 8 equal parts is treated as one block, and the frequency characteristics obtained in the above (■) are determined in each block by the averaging circuit 12. Calculate the average of Let Sik be the average frequency characteristic of the audio signal in one band block at this learning stage.

■The average of the frequency characteristics of each block obtained in each band in the above ■is the sum of the levels of all blocks in the corresponding band ΣS
Divide by mik and normalize as follows for each corresponding band.

(2) The value obtained in (2) above is input to the neural network 20. Number of inputs is 16 channels x 8 blocks =
There will be 128 pieces.

2. Learning ■ A neural network 20 with 128 input layers and 48 output layers is used.

Each of the 047 recognized words is numbered and made to correspond to the 47 output layers, and for each recognized word, the output layer corresponding to that word is "1" for the input obtained in step 1 above.
Learning is performed 5000 times by pack propagation so that the other output layers have the value rOJ (target value). In this way, a neural network 20 that can guarantee a constant recognition rate is constructed.

(Evaluation) 1. Input Creation - Pass the unknown input speech waveform of each recognized word through a 16-channel bandpass filter 11 to calculate the frequency characteristics of the input speech.

■In each band of the band pass filter 11, the audio waveform is temporally divided into eight equal parts, each of which is treated as one block, and the averaging circuit 12 calculates the frequency characteristics obtained in the above (■) within each block. Calculate the average of The average frequency characteristic of the audio signal in one band block at this evaluation stage is defined as Stik.

■The average of the frequency characteristics of each block obtained in each band in the above ■is the sum of the levels of all blocks in the corresponding band ΣS
Divide by tik and normalize as tik (4) Σ S tik for each corresponding band.

2. Learning ■ Input the values obtained in (■) above to the neural network 20.

(2) The determination circuit 30 determines the input word based on the value of the output layer of the neural network 20.

Hereinafter, experimental results of the present invention will be explained.

(Experiment 1) As an example of the present invention, the average of frequency characteristics normalized for each band was input to the neural network 20.

The words to be recognized were the names of 47 prefectures, and the specific speaker was one person.

As a result, the recognition rate was 94.4%.

(Experiment 2) As a comparative example, the average frequency characteristics were calculated and normalized by the entire word power, and then the neural network 20
It was used as an input. The words to be recognized were the names of 47 prefectures, and the specific speaker was one person.

As a result, the recognition rate was 57.0%.

Hereinafter, the operation of the above embodiment will be explained.

■Since "frequency characteristics" are used as the feature parameters input to the neural network 20, the preprocessing to obtain the input is simpler and parallel to the frequency characteristics than complex feature extraction such as LPG correlation or LPC cepstrum. can be analyzed, and the time required for preprocessing is short.

(2) In principle, in the neural network 20, the calculation processing of the entire network is simple and quick.

(2) In principle, each unit constituting the neural network 20 operates independently, and parallel arithmetic processing is possible. Therefore, calculation processing is quick.

■With the above ■～■, voice recognition processing becomes complicated! It can be easily processed in real time without having to rely on

■It is resistant to constant spectral distortion and maintains a high recognition rate. This is due to the fact that the value of equation (4) normalized at the evaluation stage is canceled out by the frequency transmission characteristic Ak as analyzed using equation (2), as mentioned above in [effect]. This is because it becomes equivalent to the value of equation (3) normalized in the learning stage, and spectral distortion caused by the influence of noise or differences in input systems such as lines can be eliminated.

(2) Regarding the hierarchical neural network 20, a simple learning algorithm (pack propagation) as described above has been established at present, and a neural network 20 that can achieve a high recognition rate can be easily formed.

[Effects of the Invention] As described above, according to the present invention, real-time processing is easily possible.
Moreover, it is possible to obtain a speech recognition method that can ensure a high recognition rate.

[Brief explanation of drawings]

Fig. 1 is a schematic diagram showing an example of a speech recognition system to which the present invention is applied, Fig. 2 is a schematic diagram showing a neural network, Fig. 3 is a schematic diagram showing a hierarchical neural network, and Fig. 4 is a schematic diagram showing a unit. FIG. DESCRIPTION OF SYMBOLS 1... Recognition system, 10... Band pass filter, 12... Averaging circuit, 13... Normalization circuit, 20... Neural network, 30... Judgment circuit.

Claims (2)

[Claims] (1) A word recognition method that uses a neural network to recognize words from input speech, in which the frequency characteristics of the input speech are calculated, and each speech interval divided equally in time in each band is divided into one A speech recognition method in which the average frequency characteristics of each block are calculated, and the averages are normalized for each corresponding band and used as input to a neural network. (2) The speech recognition method according to claim 1, wherein the neural network is a hierarchical neural network.