JPS59133599A - 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] (b) Industrial application field The present invention relates to a speech recognition device that recognizes speech.

1) Prior Art Figure 1 shows the configuration of a conventional speech recognition device. In the figure, +13 is a microphone that converts voice into an electrical voice signal, and (2) is a voice pattern extraction circuit that extracts a voice pattern indicating voice characteristics from the voice signal obtained by the microphone (1). , (3) is an input voice pattern memory that stores the voice obtained in the voice pattern extraction circuit (21), and (4) is an input voice pattern memory that stores the voice patterns of the IT voice to be recognized in numbered form and stores them in advance as previously registered voice patterns. (5) is a recognition processing unit that performs pattern recognition, which compares the voice pattern of the input voice pattern memo 1c3) with each voice pattern of the registered voice pattern memo 4), respectively; Most similar registered pattern memo January 4)
The number of the voice pattern is output. That is, the recognition process of the voice input to the microphone (1) is performed by the basic configuration consisting of the microphone (1) to the recognition processing section (5). Further, 16) is a comparison circuit, 17) is a kakunta, (8) is a display, and (9) is a correction section. The comparison circuit 16) to the correction unit +91 perform processing for correcting each voice pattern in the registered voice pattern memory (4) to a more accurate pattern.

Next, the operation will be described by taking as an example a case where the conventional speech recognition device as described above is used to recognize a place name speech.

First, in the registration mode, the speaker uses the microphone (1
) to input place name sounds sequentially. At this time, the kakunta circuit (7) sequentially increments the numerical display on the display (8) from 1", and when the Z numerical display is 1", the speaker inputs the voice "tokukiyoku" into the microphone (1). do. At this time, the voice pattern "Tokukiyoku" extracted by the pattern extraction circuit (2) is temporarily stored in the input cover turn memo (January 3), and then this voice pattern is registered in the voice pattern memo (January 4). is stored in association with the number ``1''.Next, when the number display on the display (8) is incremented to ``2#'', the speaker uses the full voice microphone (ll) called ``Nagoya''.
- Register the voice pattern "Nagoya" as in the case of "Tokukiyoku" mentioned above Voice pattern memo January 4
) is stored in association with the number "2#".In the same way, depending on the speaker's utterance, the registered voice pattern memo is stored (January 4).
II]@
It is stored in association with the next # numbers "3"', "4", "5#, II 6I+, and puff"'.

However, when using the registration mode described above, it is not always possible for the speaker to pronounce the sound with accurate pronunciation and accent, and there is a possibility that erroneously "stopped" speech patterns may be stored in the registered speech pattern memo (January 4). Therefore, a correction mode is provided to correct this erroneous voice pattern.In the prayer correction mode, when the number displayed is 1'', the speaker says ``
The voice "Tokiyoku" is input into the microphone (1). At this time, the pattern extraction circuit (2) extracts "
The voice pattern "Tokukiyoku" is temporarily stored in the input voice pattern memory (3). Zoshi C, recognition processing city) is the input voice pattern memo January 31, 1 special key yoku'' registered voice pattern voice pattern memo January 4) each 7
Pattern recognition based on individual speech patterns. As a result, if it is accurately recognized as the voice pattern "Tokukiyoku" associated with the number Pa1# of the registered voice pattern memo January 4), the recognition processing unit (5) outputs the voice pattern. The comparator circuit 16 detects that the number "1" and the above-mentioned kakunta circuit (force count value "1#" match) and outputs a match signal S.

The concordance signal S increments the count value of the kakunta circuit (7) and changes the numeric display on the display (8) to Q271. Thereafter, in the same way, according to the numerical display on the display (8), place names corresponding to the numbers are inputted into the microphone (1) one after another. When the voice "Osaka" is input into the microphone tl), the recognition processing unit (5) recognizes that the input voice pattern at this time is 1 Okayama corresponding to the number "5" in the registered voice pattern memo January 4). '', and outputs the number II 5 II, the comparison circuit 16) compares this number ゛5''' with the count value ``4'' of the kakunta circuit (7). outputs a discrepancy signal S. This discrepancy signal S operates the correction section +91 without causing the kakunta circuit 1 to step forward.

At this time, the correction section (9) is connected to the input voice pattern memory (3).
The voice pattern ``Osaka'' from this time stored in ``Osaka'' and the voice pattern ``Osaka'' associated with the number ``4'' in the registered voice pattern memo January 4) are averaged. '' is obtained, and the voice pattern "Osaka" in the registered voice pattern memory (4) is changed to this new averaged voice pattern. At this time, the number displayed on the display (8) is '4'', and the speaker again inputs the voice "Osaka" into the microphone (1), and the recognition processing unit (5) accurately The operation for adding (+5) to the voice pattern "Osaka" in the registered voice pattern memory (4) as described above is repeated until the voice is recognized as "Osaka" and the number "4#" is output. When this correction is completed, the process moves on to the above-mentioned correction process for the voice ``Okayama'' in the next registered f voice pattern memo January 4), and this process continues sequentially until the last voice ``...kata''. It is done.

In the recognition mode, the voice pattern input to the microphone 111 is extracted; 2) and stored in the input voice pattern memo order 3), and registered voice pattern memo January 4)
The recognition processing unit (5) performs pattern recognition based on each voice pattern. That is, the recognition processing unit 15) has the number °'4.
”, the voice uttered by the speaker at this time was “Osaka”.

However, when it comes to new and conventional speech recognition devices,
For example, if the pronunciation and accent of the voice "Okayama" is incorrect in the registration mode, or if noise is mixed in when pronouncing this voice, it may correspond to the scent "5'" of the registered voice pattern memo January 4). This sound pattern of “Okayama” is “Osaka” which corresponds to the number “4″′.
In some cases, the voice pattern becomes similar to that of the other person. In this case, in the correction mode, the corresponding numerical order is followed, and the 16th correction of the voice pattern "Osaka" in the registered voice pattern memory 14) is performed first. Even if the voice pattern of ``Osaka'' is quite accurate, the f-voice pattern of ``Osaka'' uttered by the speaker is more likely to be a number.''
The sound pattern may resemble the inaccurate "okayama" sound pattern that corresponds to 5". At this time, the recognition processing unit 1
5) outputs the number "5" and registers the voice pattern (4)
However, it is not possible to perform a modification process that can sufficiently distinguish the voice pattern ``Osaka'' from the erroneously similar voice pattern ``Okayama.'' Therefore, not only was it impossible to complete the correction process for the voice pattern "Osaka" in the C1 registered voice pattern memo (January 4), but it also became impossible to correct the voice pattern thereafter.

(C) Object of the Invention The present invention aims to eliminate the above-mentioned drawbacks and provides a speech recognition device that can reliably modify speech patterns in a registered speech pattern memory.

B) Structure of the Invention The speech recognition device of the present invention comprises a modification means for modifying each speech pattern in the registered speech pattern memory, and an instruction means for instructing the speaker to utter the registered speech to be modified by the modification means. The user sets the registered voice to the registered voice specified by the speaker in accordance with the instruction means, and the instruction means instructs the speaker to use the new registered voice changed by the correction means. .

(e) Embodiment FIG. 2 shows an embodiment of the speech recognition device of the present invention. In this figure, parts having the same functions as those of the conventional device shown in FIG. 1 are given the same numbers as in FIG.

Note that +69, (8), and (9) indicate a comparison circuit, a display, and a correction unit as in the conventional device, but as will be described later, the comparison target, display target, and correction target are the same as in the conventional device. different from. +10) selectively outputs the count value of the kakunta circuit (7) and the output number of the recognition processing section 15), and transmits this output value to the comparison circuit + (3j, display 18) and the correction section (9). It is a selection circuit, and when it receives a match signal S from the comparator circuit 16), it outputs the count value of the force from the kakunta circuit (force count value).On the other hand, when it receives a mismatch signal S from the comparison port 1F1+e'+, it performs recognition processing. The output number from the circuit 15) is output, and then, at the time when either the coincidence signal S or the disagreement signal S is received from the comparator circuit 16), this output value becomes the count value of the catank circuit (7). will be reinstated. Uυ is an opening/closing circuit that opens and closes the connection between the Kakunta circuit (7) and the coincidence signal S line that causes the Kakunta circuit (7) to step; , the coincidence signal S is connected to the Katank circuit (force) only when the selection circuit outputs the count value of the Kakunta circuit (force).Therefore, the comparison port [165 is the output number from the recognition processing section (5) and the selection When the output value from the circuit +1 (l) matches, the -interrogation circuit (ll) is connected, thereby incrementing the counter circuit (7), and using the mismatch signal S to change the correction unit (9).
), and furthermore, these two
, S causes the selection circuit 11O) to make the selection 11. In addition, the display section (8j selection circuit l (kakunta times 1 from υ)
The count value of 171 or the output number of the recognition processing section +51 is displayed. Furthermore, the correction section +95 is the above comparison circuit +6
When the discrepancy signal S from 5 is received, the voice pattern of the registered voice pattern memo 1/4) corresponding to the output number of the recognition processing section (5) obtained from the selection circuit is corrected.

The operation when the speech recognition device of the present invention having such a configuration is used to recognize a place name speech will be described. Note that the registration mode and the recognition mode operate in the same manner as the conventional device, so their explanation will be omitted.

Therefore, similar to the explanation in the conventional example, the sounds "Tokukiyoku", ~, and "No Kata" are stored in the registered voice pattern memory 4) in association with the numbers "1" and "'7". The positive mode will be explained below.

First, in the initial state of deafness, input is made to the display (6) via the counter circuit (the power kakunt IMB "i" is forcibly selected by the selection circuit QO), and this display (8) shows the number °゛1.
'' is displayed.The speaker follows this and inputs the voice "tokukiyoku" associated with the number u 1 # into the microphone (1), and this voice pattern is extracted as a nokai turn (2) and becomes the input voice pattern. The memo is temporarily stored in January 3) and pattern recognized in the recognition processing city).

At this time, if the recognition processing unit (5) correctly determines that the voice pattern is "tokukiyoku" which is associated with the number uI11 of the registered voice Bakun memo Iha4), then the number 111
The comparator circuit (6) inputs the coincidence signal S to the counter circuit (force) via the open/close switch circuit Uυ and transmits it to the selection circuit 11tll.

Therefore, the correction unit 19) does not perform any correction operation, and the selection circuit Ut) increments the count value of the counter circuit (7).
'' is selected and displayed on the display 18).Then, the numerical display on the display 18) is sequentially changed in the order of the corresponding numbers until the discrepancy No. 8 S is obtained from the comparator circuit 16)'.
According to this, the speaker sequentially inputs the place name sounds into the microphone (1).

On the other hand, for example, when the number displayed on the display (8) is "4"' and the corresponding voice "Osaka" is input into the microphone (1), the recognition processing unit (5) Pattern registered voice pattern memo January 4) number “
If the number "5#" is determined to be most similar to the voice pattern "Okayama" which is inaccurate due to some reason corresponding to "5", the comparison circuit 16) will output a mismatch signal.
Outputs 1. In this case, the opening/closing circuit t11) is opened by this discrepancy signal S, and the count value II 478 is maintained at the kakunta cycle b@1.7). on the other hand,
The selection circuit (11) selects the number "5" from the recognition processing section (5) at this time and displays it on the display (8), and the modification section (9j) performs a modification operation. , i6
The main part 191 is the registered voice pattern memory corresponding to the count value °°4# held by the counter circuit 17), which is the number of the direct Afr whose output from the selection circuit +ll is changed to the number "5". After changing the sound pattern "Osaka" into an averaged sound pattern in the same way as the correction unit (9) of the conventional device shown in FIG. The voice pattern "1 OKAYAMA" in the registered voice pattern memory associated with "5" becomes a new target for correction.

Then, when the speaker inputs the voice "Okayama" into the microphone (1) according to the numerical display pattern 5'' on the display (8), the registered voice pattern memo is created by the operation of the correction unit (9). The voice pattern "Okayama" from January 4) is corrected to the averaged pattern, but at this time, the output number as the recognition result of the recognition processing unit (5) is not necessarily 5". In other words, the registration Because the voice pattern "Okayama" in voice pattern memo January 4) is inaccurate, the output number is often not "5". Therefore, the comparison circuit
Either a match signal S or a mismatch signal S is obtained from the signal S, and depending on this signal S and S, the switching circuit Uυ is closed and the selection circuit +lQl that has been held in the counter circuit (7) is The count value "4" is output again.

Due to this, the correction department (9j's correction target is "Osaka")
9f&, and the display (the number display of 81 was changed from 5# to "4", indicating 1-Osaka to the speaker)
Instructs to input the voice into the microphone (11) again.

However, the incorrect voice pattern for the voice "Okayama" in the registered voice pattern memo January 4), which was subject to misrecognition for the voice "Osaka", was first corrected, and then again for the voice "Osaka". ” correction processing is started for the voice pattern. However, the modification process at this time is difficult because the voice pattern "1 Okayama" in the registered voice pattern memo January 4) has already been modified to the extent that it can be sufficiently distinguished from the voice pattern "Osaka". The voice pattern "Osaka" has also been corrected by -K (IS), so if this correction process is repeated at most two or three times, the effect of inaccurate voice patterns other than the voice pattern "Osaka" will be disappears, and the modification of the "Osaka" voice pattern can be completed.

Therefore, the comparison circuit +65 outputs the matching signal S, and it is possible to quickly move on to the correction process for the next l "5" voice "Okayama", and in the same way, the last number "7" can be corrected.
The correction process is performed sequentially until the ``Hakata'' sound is reached.

If you are reluctant to read the above explanation, please take a look at the registered voice pattern memo (January 4).
As a correction means for correcting the registered voice pattern of , a correction unit (
9) has been shown, but it may also be a modification means that simply changes this registered voice pattern to the input voice pattern.

Unfortunately, as a means of instructing the speaker which voice to utter, I used a display (8) that displays a # symbol associated with the voice, but it does not display the voice at all. Alternatively, it is also possible to provide a voice synthesis circuit and output the voice using this circuit.

(F) Effects of the Invention As is clear from the above description, the speech recognition device of the present invention includes a modification means for modifying each speech pattern in the registered speech pattern memory, and a registered speech that is modified by the modification means. If the user inputs the registered voice instructed by the speaker according to the instruction means by setting it in the correction mode and misrecognizing it as another registered voice,
The correction means changes the target of the registered voice to be corrected to the registered voice that has been misrecognized, and the instruction means instructs the speaker to use the new registered voice changed by the correction means. Even if the speech pattern to be corrected is relatively accurate, and other speech patterns are erroneously similar to the previous speech pattern for some reason, after correcting this other inaccurate speech pattern, , the previous voice pattern can be modified. Therefore, it is possible to correct the previous speech pattern accurately to the extent that it can be sufficiently distinguished from other speech patterns, and the correction process can be completed reliably and quickly, improving the operability of this type of speech recognition device. Not only can this be a significant improvement, but it also contributes to reducing misrecognition in recognition mode.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a conventional speech recognition device, and Fig. 2 is a block diagram of an embodiment of the speech recognition device of the present invention. IIJ is a microphone, (2) is a speech pattern extraction circuit, and 13)
is the input voice pattern memo 1c14) is the registered voice pattern memo 1c1fil is the recognition processing unit, +6++6J is the comparison circuit, (power is the kakunta circuit, +5) (sj ke display, +5
)t9j is a modification section, Cl0) is a selection circuit, and Q, ]) are open/open circuits, respectively.

Claims (1)

[Claims] 1) A speech recognition device in which a recognition processing unit performs pattern recognition on a speech pattern obtained by a pattern extraction circuit from a speech signal obtained by inputting speech into a microphone, based on a plurality of speech patterns in a registered speech pattern memory. In addition to a modification means for sequentially modifying each voice pattern in the registered voice pattern memory, an instruction means for instructing a speaker to utter the registered voice to be modified by the modification means is provided. When modifying the voice pattern in the pattern memory, the speaker inputs the registered voice instructed by the instruction means into the microphone, and the registered voice input at this time is distinguished from other registered voices by the recognition processing section. In the case of erroneous recognition, the correction means changes the target of the registered voice to be corrected to the registered voice that has been erroneously recognized by the recognition processing section, and the instruction means changes the target of the registered voice to be corrected to the registered voice that has been erroneously recognized by the recognition processing section, and the instruction means changes the target of the registered voice to be corrected to the registered voice that has been erroneously recognized by the recognition processing section. A speech recognition device characterized by instructing a speaker.