JPH03180897A - Voice recognition device - Google Patents

Abstract

PURPOSE:To improve the recognition rate by calculating the similarity between registered voice patterns, and making an annunciation and prompting a pattern change when the calculation result is larger than a specific value. CONSTITUTION:A CPU 3 calculates similarity between two of voice patterns out of plural voice patterns registered in a pattern memory 6 according to a command from a key matrix 7. The CPU 3 compares the calculated value with a threshold value and when the calculated value exceeds the threshold value, a combination of voice patterns with large similarity is displayed on a data display device 8 to prompt an alteration of the voice pattern of a voice 'hichi' or seven in English which is easily recognized as 'ichi' or one in English into 'nana' or seven in English. The voice pattern in the memory 6 is changed and the changed voice pattern is used as a standard pattern for recognition. Then there is no voice pattern whose similarity is larger than the specific value and the voice recognition rate is increased.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention is a method for recognizing word sounds by comparing a speech pattern generated by a user with a plurality of pre-registered speech patterns (standard patterns). Relating to a recognition device.

[Summary of the invention] It has an area to store voice registration patterns.

Speech patterns corresponding to the speech words to be recognized are stored in advance, and pattern matching is performed between the speech registration patterns in the registered pattern area and the input speech patterns to recognize the speech words corresponding to the input patterns. In the recognition device, the registered voice pattern A1 and the input voice pattern AJ (i=1.2.
・・・・・・eNt j= 1.2.・・・・・・
..., N) Mutual similarity S(i.

j) is calculated, and when this exceeds the threshold STH,
It has a display section that informs the user of the combination of voice-type Kabataan at that time, and an input section that allows the user to specify the pattern to be changed. A voice recognition device configured to allow the user to choose.

[Prior Art] In recognition by a word speech recognition device, generally, a matching process is performed one after another between a speech pattern generated by a user and a plurality of pre-registered speech patterns (standard patterns). In the recognition results, the similarity obtained during matching (
A voice pattern whose value (or distance) is larger (smaller) than the input voice pattern is selected. Causes of misrecognition include differences in the degree of generation of environmental noise 1, variations in speakers, etc., which pose a problem. One of the reasons is that the degree of similarity (distance M) between a plurality of registered voice patterns is very similar.

[Problem to be solved by the invention] In conventional systems, any word can be used to register, and the recognition result is often the one closest to the input speech pattern. No measures have been taken to prevent misrecognition caused by differences.

[Object of the Invention] In order to solve the above-mentioned problems, the object of the present invention is to obtain the degree of similarity (distance) between a plurality of registered voice patterns,
It is an object of the present invention to provide a speech recognition device for realizing a function of notifying a user of a combination of speech patterns that can be determined to cause erroneous recognition and prompting the user to re-register the speech patterns.

[Means for Solving the Problems] In order to achieve the above object, the present invention has an area for storing registered patterns of speech, stores speech patterns corresponding to speech words to be recognized in advance, and stores the speech patterns corresponding to the speech words to be recognized in advance. The speech recognition device recognizes the speech word corresponding to the input pattern by performing pattern matching between the registered speech pattern and the input speech pattern. a first means for calculating the degree of similarity (distance) of the above; a second means for comparing the value calculated by the first means with a predetermined value; When the degree of similarity is greater, the gist includes means for notifying this fact.

[Operation] Evaluate the degree of similarity or (distance M) between registered standard patterns, feed back the results to the user, and make it possible to configure standard patterns by combining voice patterns that are unlikely to be misrecognized.

[Examples] The present invention will be explained in more detail below using Examples with reference to the drawings, but these are merely illustrative and various modifications and improvements can be made without going beyond the scope of the present invention. Of course it is possible.

FIG. 1 is a block diagram showing the configuration of a speech recognition device according to the present invention, in which 1 is a microphone, 2 is a feature extraction section, 3 is a central processing unit (CPU), and 4 is an RO.
M, 5 is a RAM, 6 is a pattern memory, 7 is a key matrix, and 8 is a data display device.

As shown in FIG. 1, the voice is converted into an electrical signal by the microphone l, and then sent to the CPU 3. Program ROM4, RAM
Feature extraction unit 2 controlled by a microcomputer unit consisting of 5
The features are extracted and stored in the pattern memory 6.

As shown in FIG. 7, the feature extraction unit 2 includes a microphone, an AGC
Amplifier, high-frequency emphasis circuit, multiple bandpass filters (BP
F), a filter bank consisting of a plurality of rectifiers and a plurality of smoothing circuits, and an A/D converter.

The sound input from the microphone 1 is amplified by the AGC amplifier, and after the amplitude is limited to within a certain value, the high frequency emphasis circuit corrects the output drop in the high frequency range of the voice. Next, the sixth
It is input to a filter punk with the characteristics shown in the figure,
The output of each filter is rectifier 1~n, smoothing circuit 1~n
The signals are each rectified and smoothed and input to each channel of the A/D converter.

The operation of the above embodiment will be explained below.

While the registration mode is specified by the key matrix 7, the microcomputer section operates in the same manner as described above, and only the number of patterns that are less than or equal to the maximum number of registered words N in the system is input. The pattern memory 6 has an area where patterns corresponding to the voice at the time of registration are stored, and the voice pattern AI
(i==1.....,N), which corresponds to M and (i=1.....,N), respectively.
It is defined as

The similarity between the pattern Ml and the pattern M is expressed as S(i, j
), and calculate the similarity S for all combinations of ltJ. Therefore, the microcomputer section calculates the degree of similarity between certain patterns stored in the pattern memory 6, and the procedure is as follows.

1) S(11j) (i = 1.2....
...1 N.

J　　　　　　　　　ly　2e...,N).

2) Store and display the combination lyJ such that S(i,j)>STH. STH is a preset similarity threshold.

Similarity and distance have a dual relationship; when the similarity between pattern A and pattern > B is large, the distance becomes small,
When the similarity is small, the distance is large. What is similarity?
Pattern A and pattern B obtained by DP matching
means the reciprocal of the cumulative distance of

(Similarity)=(cumulative distance)-1 The concept of DP matching and cumulative distance is described in the following document.

Fliroaki, 5AKOE and 5eibi
“Dynamic Programmin” by CHIBA
gA1goritha+ 0ptiIIlizatio
nfor 5poken Word Recognit
ion"IEEE TRANSACTION ON A
COUSTIC3,5PEECH,ANDSTGNAL
PROCESSING, Vol, ASSP-26
, No. l.

Feb. 1978. PP, 43-49 threshold STH
The combinations of ltJ that exceed the above are displayed on the display section 8 to inform the user of which registered patterns have a greater degree of similarity.

The user now selects the number (k) of the pattern that must be changed from among the number combinations displayed on the display section
) is specified by the key matrix 7, the microcomputer section stores the change pattern input from the microphone 1 into the memory area M11 of the pattern memory 6 via the feature extraction section 2.

When a new pattern is input to the memory area Mk, the microcomputer section calculates the similarity S as described above as follows.

3) S (k, j) (j = 1921...
... Find N).

4) Store and display the combination of k and j such that S(k, j)>5tI4.

Here, if the pattern input into the memory area Mk has a sufficiently small degree of similarity with other patterns compared to the previously stored pattern, it will not exceed the threshold STH for the process in 4). 1'i.

Even if the process 2) is repeated again, k will not be included in the combination of numbers (codes) output to the data display device 8.

By repeating the above processing procedure between the user and the system, no combination of similarities S(x, j) between patterns finally stored in the pattern memory 6 will exceed the threshold STH. . Therefore, a certain standard (ST) can be given to the difference in the degree of similarity between the created lR11I patterns, so that it is possible to prevent misrecognition caused by the degree of similarity of the registered patterns being very similar. can.

The above processing will be explained in detail with a specific example using FIGS. 2 and 3.

FIG. 2 shows a table of similarity combinations when N=10. 5(xtj)=S(j,x), so
It is not necessary to obtain 5 (ITJ) for all combinations of 1+J, and the combinations shown in FIG. 2 may be used. S(i, 1)
(i=1, ..., 10) is the similarity about itself, so S(i+j)(j =
L...10), and STH
(ST□ is S(i,i)
).

Figure 3 shows the pattern memory area M, (i==1,...
... 10) stores a voice pattern in which the numbers l to 10 are read aloud. ), if calculation is performed using the method of calculating the similarity described above, 1M1(
Ichi) and M? The similarity of (Sichi) is large and S(L,
Suppose that 7)>5T)l. At this point, without changing the pattern, say + A = “ichi’”,
When voice recognition is performed, Ml (ichi) and M? (Sichi)
Since the similarity of is large, originally S(A, 1) is S(
A, 1 )...... S(A, 10 ), and as conceptually shown in Figure 4, the recognition code corresponding to the pattern Ml should be returned, but S( A.7
) may be determined.

Therefore, in order to prevent such a situation, when S(1,7)>STH is outputted from the data display device 8 shown in Fig. 3 in Fig. 3, the user performs the previously explained action. If we continue to change the pattern of M7 to ``Nana'', the similarity between the patterns of Ml and M7 will decrease, so
S(A.

The result of comparing 1) and S(A, 7) is S(A.

1) will be bigger and will be wiser and less likely to make mistakes.

FIG. 5 is a flowchart showing the operation of the microcomputer section. In the figure, N is the number of registered words.

S(i,j) represents the similarity between the i-th pattern and the j-th pattern, STH represents the similarity to S(i,j), and MIl represents the th audio storage area. It is assumed that N voice patterns are stored in the pattern memory Mk before starting this operation.

[Effects of the Invention] As explained above, according to the present invention, the matching algorithm used in conventional speech recognition devices can be used as is.
If matching is applied to a standard pattern, the present invention can be constructed. Therefore, this function can be added to the current system relatively easily, and the cost can be reduced and the recognition rate can be increased.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of the speech recognition device according to the present invention, FIG. 2 is a diagram showing the similarity between pattern memories when N=10, and FIG. 3 is a diagram showing an example of changing the i2 record pattern. , FIG. 4 is a conceptual diagram of recognition code determination, FIG. 5 is a flowchart showing the operation of the microcomputer section, and FIG.
The figure is a block diagram showing the configuration of the characteristic oil outlet, and Figure 6 is a block diagram showing the configuration of the characteristic oil outlet.
FIG. 3 is a frequency characteristic diagram of the filter bank shown in the figure. 1...Microphone, 2...
...Feature extraction part. 3...Central processing unit (CPU),
４・・・・・・・・・ROM 、５・・・・・・・・・
...RAM, 6...Pattern memory, 7...Key matrix, 8...
・・・・・・Data display device.

Claims (1)

[Claims] It has an area for storing registered speech patterns, stores speech patterns corresponding to speech words to be recognized in advance, and performs pattern matching between the registered speech patterns in the registered pattern area and the input speech patterns. In a speech recognition device that recognizes speech words corresponding to an input pattern by performing a second means for comparing the value calculated by the first means with a predetermined value; and a means for notifying when the degree of similarity is greater according to the comparison result of the second means. A voice recognition device featuring: