JPH03230256A - Voice recognizing method - Google Patents

Abstract

PURPOSE:To enable a real-time processing and to secure the high rate of recognition by calculating the average of frequency characteristic in the respective blocks of hourly equally divided voice blocks and using a value differentiating the average between the block equipped with the maximum level and the other block as an input to a neural network. CONSTITUTION:The already known voice waveform of each recognizing word is passed through a 16 channel band pass filter 11 and the frequency characteristic of the input voice is calculated. In the respective bands of the band pass filter 11, respective voice blocks hourly equally dividing the voice waveform into eight are defined as one block, and an averaging circuit 12 calculates the average of the frequency characteristics in the respective blocks. Then, a differential value H between the block equipped with the maximum level and the other block is calculated and this calculated value is defined as the input to a neural network 20. Thus, the real-time processing is enabled and the high rate of recognition is secured.

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] Conventionally, Japanese Unexamined Patent Application Publication No. 1999-1-1 has been used as a speech recognition method that can ensure a high recognition rate despite the influence of noise, differences in input systems such as lines, etc.
A method as described in Japanese Patent No. 260490 has been proposed. This speech recognition method calculates the LPC cepstrum for each frame through LPG analysis of speech, and calculates the LPC cepstrum between frames.
Create the difference value of the PC cepstrum as an input parameter. On the other hand, a standard pattern for each voice is also created using the same difference value, and the degree of similarity between the input parameters and the standard pattern is calculated using a statistical measure, and the standard pattern with the maximum degree of similarity is selected. Use the voice as the recognition result.

[Problem to be solved by the invention] However, the prior art has the following problems (1) to (3).

(2) Since the LPC analysis is performed sequentially for each frame in which audio is divided into multiple segments in time series, the processing time is large.

(2) Calculation of similarity after calculation of input parameters is also performed sequentially for each frame, which requires a large amount of processing time.

(2) Due to the above (2) and (2), high-class and complicated processing is required to perform real-time processing.

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 as set forth in claim 1 is a word recognition method for recognizing a word from an input voice using a neural network, which calculates the frequency characteristics of the input voice and calculates the frequency characteristics of each band. In each of the blocks, 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 these averages are divided into the block with the maximum level in the corresponding band and the other blocks. The difference value between the block and the block is used as an input to the neural network.

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

[Function] According to the present invention as set forth in claim 1, there are the following effects (1) to (4).

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

■In principle, neural networks allow simple and quick calculation processing of 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 maintains a high recognition rate. This is because, as analyzed below, the average frequency characteristics of each block of the input audio are differentiated between the pro-v having the maximum level in the same band and the other pro-v. , due to the ability to eliminate spectral distortion. That is, ■ is the program number, k is the band number, Ak is the frequency transmission characteristic of the band, Smik is the audio signal of one block of the band at the learning stage, and Stik is the voice signal after passing through the telephone line at the evaluation stage. As shown in , when the audio signal of one block in the k band is spectrum-distorted due to the influence of the stationary frequency transmission characteristic Ak, Stik=AkIISmi. Then, each speech signal S tik at the evaluation stage is normalized by the entire power of the word, and then, by taking the logarithm of the above equation (1), for example, the j block having the maximum level in the same band Take the difference between This difference value H is: = log(S tik) - log(Σ ΣS tik
)−log(S tjk)+log(ΣΣS tik
)=log(S tik)−log(S tjk)
...(2)=log(A k-8
m1k)-1og(A k-8mjk)=log(A
k) + log(S m1k)−log(A k)−
log(S mjk)=log(S m1k)−lo
g(S mjk) (3).

Equation (2) of the above difference value H represents the difference value of the audio signal at the evaluation stage, and equation (3) represents the difference value of the audio signal at the learning stage. That is, the difference value of the audio signal in the evaluation stage of equation (2) has the frequency transmission characteristic Ak removed and becomes equal to the difference value in the learning stage of equation (3), that is, 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.

[Embodiment] Figure 1 is a schematic diagram showing an example of a speech recognition system to which the present invention is applied, Figure 2 is a schematic diagram showing a neural network, Figure 3 is a schematic diagram showing a hierarchical neural network, FIG. 4 is a schematic diagram showing the structure of the unit.

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
The present invention can be roughly divided into two types: the hierarchical network shown in FIG. 2(A) and the interconnected network shown in FIG. 2(B). Hierarchical networks are more useful because simple learning algorithms have been established as described below.

(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) Network learning means bringing the actual output closer to the target value (desired output). Learn by making changes.

Also, as a learning algorithm, for example, Rug+e
lhart, D.E., McClelland, J.
, L, and thepHl' Re5earch
Group, PARALLEL DISTRiBtl
TEDPROCESSING, the MIT Pr
ess, 1986. Pack propagation as described in .

゛Hereinafter, specific embodiments 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 block difference 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 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 Let S mik be the average frequency characteristic of the audio signal in the band l block in this learning stage.

■Find the block with the maximum level for each band.

Then, the difference value between the block with the maximum level and other blocks is determined. That is, the average of the frequency characteristics of each block obtained in each band in (■) above is calculated as the word power ΣΣS
It is normalized by dividing by mik, then the logarithm is taken, and the difference with block j having the maximum level within the same band is taken to calculate the difference value H as shown in equation (3) above.

(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 2o 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 a value of "O" (target value). In this way, a neural network 2o 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 the i-band block at this evaluation stage is defined as Stik.

■Find the block with the maximum level for each band.

Then, the difference value between the block with the maximum level and other blocks is determined. That is, the average of the frequency characteristics of each block obtained in each band in the above (■) is calculated as the total power of the word ΣΣS
Divide by tik to normalize as in equation (1) above, then take the logarithm, take the difference from block j having the maximum level within the same band, and calculate the difference value H as in equation (2) above. calculate.

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.

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

(Experiment 1) As an example of the present invention, the difference between the average frequency characteristics of the block having the maximum level and other blocks within the same 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 94.0%, and the processing speed was within 1 second (average recognition time for one word).

(Experiment 2) As a comparative example, an input parameter and a standard pattern were each created using the inter-frame difference values of the LPG correlation and the LPC cepstrum, and the degree of similarity between the two was calculated using a statistical scale. 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 93.2%, and the processing speed was more than 1 second (average recognition time for one word).

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.

(1) In principle, each unit constituting the neural network 20 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. As mentioned above in [Operation], this means that the difference value calculated in the evaluation stage as shown in equation (2) is the frequency transmission characteristic A.
This is because k is eliminated and the difference value becomes equivalent to the difference value as shown in equation (3) calculated in the learning stage, and spectral distortion caused by the influence of noise or differences in input systems such as lines can be eliminated.

For the 0-layer neural network 20, a simple learning algorithm (backpropagation) as described above has been established, and it is possible to easily form the neural network 20 that can achieve a high recognition rate.

[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... Block difference 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 As a block, the average frequency characteristics are calculated within each block, and the difference between the average and the block with the highest level in the corresponding band and other blocks is applied to the neural network. Speech recognition method used as input. (2) The speech recognition method according to claim 1, wherein the neural network is a hierarchical neural network.