JPS59176797A - Voice recognition equipment - 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 (1) Technical Field of the Invention The present invention relates to a speech recognition device equipped with means for recognizing input speech using registered standard patterns.

(2) Background of the technology Conventional speech recognition devices, especially speech recognition devices that recognize the voice of a specific speaker, generally convert the input voice from an analog signal to a digital signal using an AD converter, and then convert the input voice into a digital signal. Recognition is performed on signals, but in this case, the recognition device itself has a standard pattern of multiple sounds corresponding to frequently used phrases (or words), that is, a dictionary, and compares the standard pattern with the pattern of the input sound. By doing so, it searches for words from standard patterns and outputs them. In this case, the patterns to be compared are, for example, parameters representing the frequency characteristics of audio.

(3) Prior art and problems In this type of speech recognition device, it is necessary to create multiple different standard patterns depending on differences in speakers or frequently used words. Switch the device to the mode for creating standard patterns,
A standard pattern is created by having a speaker pronounce the voice for which a standard pattern is to be created and registering it. However, in such a method, each time a new voice that is not included in the standard pattern is input or a speaker is different, the device is switched to the standard pattern creation mode one by one.
After registering the voice and creating a standard pattern, the device must be switched to recognition mode, and the voice must be pronounced again for recognition, which requires extra operations.

(4) Purpose of the Invention In order to eliminate the above-mentioned drawbacks, the present invention has the purpose of eliminating the above-mentioned drawbacks.
When a new voice that is not registered in the standard patterns is input, a standard pattern for that voice can be automatically created and registered, and even registered standard patterns that are rarely used can be automatically deleted. It is an object of the present invention to provide a speech recognition device which the user of the device can use without worrying about whether the word ``has been'' registered in the standard pattern.

(5) Structure of the Invention To achieve the above object, the present invention provides a standard pattern creation dark value to automatically create and register a standard pattern of the input audio from the input audio. By providing a means for counting the frequency of pattern use, the configuration is such that standard patterns that are used less frequently can be automatically deleted.

(6) Examples of the invention An embodiment of the present invention will be described in detail below.

FIG. 1 shows an embodiment of a word speech recognition device using the present invention. First, the voice input from the microphone 1 is converted into a digital signal by the filter and AD converter 2, and then subjected to acoustic analysis in the acoustic analysis section 3, and a parameter M representing the characteristics of the voice is output as an output. . The parameter M is compared in the matching section 5 with a word standard pattern Ri stored in advance in a standard pattern memory 6 provided in the device via a signal line L1. There are multiple Ris, R1, R2, R3,...Rk, and each Ri (i=1.2°..., k) is 1.
One word represents a standard pattern, and one Ri represents a set of a plurality of parameters representing one word. Therefore, in this case, the standard Bakun memory 6 stores standard patterns of 2 words. In addition, the maximum value of patterns that can be stored in the standard pattern memory 6 is set to N (
k≦N). Now, in the matching section 5, a comparison is made between the input speech pattern M and the standard pattern Ri, and recognition is performed in the matching result determining section 7. This is done by setting data 4. To explain the operation of this part, first M is the standard pattern creation darkness value DL and DH.
If the input voice pattern M is between 1 and 2, it is determined that the input voice pattern M is not registered in the standard Bakun memory 6.

The parameter M is a new standard pattern Rk+1(k+
1≦N), it is written into the standard pattern memory 6 via the registration section 8 and the signal line L2, and the input voice is registered. At this time, this speech recognition device can recognize the new speech (word) by issuing a signal saying "Please speak again" and prompting the speaker to speak again.

A comparison is made between the input cover turn M and the standard pattern Ri.
If the input speech pattern M is a recognition target threshold R1,
If it falls between I, it is determined that the input speech pattern M is registered in the standard pattern memory 6, and a word corresponding to the standard pattern Re that is determined to be most similar to the input cover pattern M is output. is outputted to the recognition section as a recognition result, and the result is displayed on the character display section 12. At this time, each standard pattern R4 has frequency information indicating how many times it has been recognized, and for the recognized standard pattern Re, the frequency information of that Re is incremented by 1.

If the recognized input cover M is further between the standard pattern correction darkness value Rish and Ls.

A command to correct the standard pattern Re recognized in the above operation is issued via the standard pattern correction section 11 and the signal line L3, and the recognized standard pattern Re is corrected under certain rules. This allows it to respond to subtle changes in the speaker's vocalizations. You can create better standard patterns.

When a new word not registered in the standard pattern memory 6 is uttered by the above operation, the pattern of the input voice is automatically stored and registered as a standard pattern in the standard pattern memory 6.

Furthermore, if the input cover pattern is registered in the standard pattern memory 6, that standard pattern is output as a recognition result,
The standard pattern is corrected in a timely manner.

However, if standard patterns of new words that are not registered in the standard pattern memory 6 are registered in the memory one after another, the storage capacity of the memory 1i (Nfilil) may be exceeded, so the standard pattern deletion circuit 10 is The time at which each standard pattern Ri stored in the pattern memory 6 was registered and the frequency at which it has been output as a recognition result up to now are read from the standard pattern R4, and less frequently used standard patterns are automatically output via the signal line L6. is configured to be deleted automatically. This allows standard patterns of frequently used words to be registered without exceeding the storage capacity of the standard pattern memory 6.

The speech recognition device is highly efficient (and easy to use), the two speakers do not have to worry about which words are registered in the standard pattern memory 6, and the labor of creating the standard pattern is saved. I can do it.

Next, the acoustic analysis section 3. A specific example of the matching section 5, the matching result determining section 7, and the standard pattern R4 will be described.

FIG. 2 shows a specific embodiment of the acoustic analysis section 3. Input audio is input from the microphone 13 and output as an electrical analog signal. Next, the frequency characteristic is band-limited (for example, 100Hz to 5KHz) by the bandpass filter 14.
) and then a bandpass filter 15. Zunawachi BPFI
~BPFn (for example, n-16). B.P.
FI to BPFn output a rectified output every fixed period of time 2, for example, 20 msec.

Each output signal is converted into a digital signal by the A/D converter 16, and these output signals (X+(jl.

X 2Ui ·+, Xn(jll is a parameter M representing the average frequency characteristic for 20 msec.

FIG. 3 (al is an example of the actual frequency envelope characteristic of the input voice for a certain period of time, 9, for example, 20 msec).

The horizontal axis is scaled in frequency from OHZ to a maximum of about 5 KHz. The vertical axis is the power (dB) at each frequency.
). Figure 3 (bl is 1001 as an example)
(z~500Hz, 5001(z~IKHz, I
Figure 3 shows a case where a group of band-pass filters with a passband width of KHz to 1.5K Hz, etc. For example, a set of outputs for 20 msec (X+ (j), X2 (j), . . .

Xn (j)], and each value is the power (dB) on the vertical axis. Although this figure shows the case where the passband width is constant, normally as the center frequency becomes higher, the passband width also becomes wider in proportion. in this case,
In the first 20 msec (Xl (1), X2 (1), =
・, Xn < 1)) is output, and (X
+ (2), X2 (2).

..., Xn (2)) is output. Since j lasts for the duration numerator, if the duration is, for example, 1 second, then T=1 second/20 msec-50.

For one word (X+ (1), X2 (1), -+, Xn
(1)) to (Xl (50), X2 (50)
, . . . up to Xn(50)l are output. In the end, for each word, as a parameter (X+ (j)
, X2 (j), ...Xn (j)) (j-L
2. . . , T) is output, which becomes the parameter output M of the acoustic analysis section 3 in FIG. Therefore, the standard pattern Ri stored in the standard pattern memory 6 in FIG.
(j), r2 (j), ....

rn (j)l (j=1.2....+,T)
It can be set as (i-1, 2, . . . , k).

Next, an example of a method for comparison in the matching section 5 of FIG. 1 using these parameters will be described.

First, as a measure to compare the degree of similarity between the cover pattern and the standard pattern, use 2, for example.

A vector distance of (Xn (j)-rn (N)"J can be defined. Then, for this value S, the standard pattern creation darkness values (Dv and Dshi) explained in Fig. 1, the recognition target possible darkness value (R and Rs), as well as the standard pattern correction darkness value (L.

and Ls), the matching result determination unit 7
recognition becomes possible. FIG. 4 shows an example in which the darkness value is set for the magnitude of the value of S. In this case, the smaller the hector distance l1ilts is, the smaller the input cover turn M can be said to be similar to the standard pattern Ri, so calculate S for all standard patterns Ri, and calculate the standard pattern creation threshold DL and DI-1. If the word is included in the standard pattern, it will be newly registered in the standard pattern as a word that is not in the standard pattern.

If there is a value of S that falls between the recognition target possible thresholds R1, . . . and R8, the standard pattern R1 corresponding to the value of S closest to R5 is the recognition result.

And if it is between the standard Bakun correction threshold and Ll-1, then the standard pattern Ri is corrected. Further, if the value of S does not fall within the above range, that is, if S>D, l, or S<RL, it can be set to 1, for example, to issue a signal of "Please pronounce it again." FIG. 5 is an example of a format representing one word of information stored in the standard pattern memory 6 of FIG. In addition to the standard pattern R4, the information for one word in FIG. 4 includes information Tt indicating when the standard pattern was first registered and information C indicating the frequency of use, as explained in FIG.
etc. can be set.

The above explanation is only an example of implementing the method shown in FIG. 1, and it goes without saying that the input cover pattern M and the standard pattern R1 can be obtained by changing the analysis method of the acoustic analysis section 3 shown in FIG. 1 to a different method. Since the standard pattern Ri and the input cover pattern M can be set to different values, different methods can be considered for comparing the standard pattern Ri and the input cover pattern M in the collation section 5, so that various implementation methods are possible.

(7) Effects of the present invention According to the present invention, since there is no mode for creating a standard pattern, the effort of the two speakers is saved.

Furthermore, since standard patterns are automatically added and deleted, and frequently used words remain, one speaker can use them without worrying about whether or not the word is registered in the standard pattern.

[Brief explanation of the drawing]

Fig. 1 is an overall block diagram of an embodiment of the present invention, Fig. 2 is a block diagram showing the specific configuration of the acoustic analysis section, and Fig. 3 (a) shows an example of the frequency envelope characteristic of input audio. Characteristic diagram, Figure 3 (bl is a power frequency characteristic diagram showing the output of the acoustic analysis section, Figure 4 is a diagram showing an example of setting each darkness value data in the matching result judgment section, and Figure 5 is a configuration of a standard pattern. It is a pattern diagram showing an example. 1.13...Microphone 2...Filter and A/
D converter 3...Acoustic analysis section 4...Each threshold value data 5...Verification section 6...Standard pattern memory
7... Verification result determination section 8... Registration section
9...Recognition section 10...Standard pattern deletion circuit 11...Standard Bakun Shobu section 12...Character display section Ll, L2. L3. La..., each signal line 14...bandpass filter 15...
Bandpass filter group 16...A/D converter patent applicant Fujitsu Limited

Claims (3)

[Claims] (1) A standard pattern creation dark value is provided for matching during voice recognition, and a means for automatically creating and registering a standard pattern for that voice from voice input results that fall between the dark values, and a recognition target at the same time. 1. A speech recognition device comprising: a means for providing a dark value and determining a target/-IL pattern corresponding to a voice input result falling between the dark values as a recognition result. (2) Speech recognition according to claim 1, comprising means for counting the frequency of use of the standard patterns used as recognition targets and means for automatically deleting the standard patterns that are used less frequently. Device. (3) A patent characterized in that a standard pattern update darkness value is provided by further subdividing the recognition target possible darkness values, and means is provided for correcting the standard pattern corresponding to the audio input that falls between the darkness values. A speech recognition device according to claim 1.