JPH01116700A - Voice recognition control system - 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] [Summary] This invention relates to a speech recognition control method that selects recognition words from a plurality of candidate word groups selected based on similarity information to input speech words. The purpose is to select safe recognition words with a simple structure that are free from the risk of abnormal behavior, and is a voice system that selects recognition words from a group of candidate words selected based on similarity information to input speech words. In the recognition control method, if the first candidate word with the highest degree of approximation is not included in the fail-safe processing target word group, the first candidate word is set as a recognition word;
If the word is included in the safe processing target word, and the candidate word group does not include any other fail-safe processing target word, the first
The candidate words are recognized words, and other fail words are added to the candidate word group.
If a word subject to safe processing is included, the configuration is configured to perform fail-safe processing. , [Industrial Application Field] The present invention relates to a speech recognition control method, particularly a speech recognition control method that selects a recognized word from a group of candidate words selected based on similarity information to an input speech word. Regarding.

[Conventional technology]

Speech recognition is widely used for voice control of machines and devices, voice data input, and the like. In this case, erroneous voice recognition manifests itself as an error in control input or data input, and the effects thereof are extremely serious. In particular, when there is a high possibility that a dangerous situation will occur due to abnormal operation due to erroneous recognition, it is necessary to increase the recognition rate of voice recognition to avoid erroneous recognition.

When recognizing speech, a dictionary that stores the speech information of many standard words is usually prepared in advance, and multiple standard words are selected as candidate words to be recognized based on their similarity to the input speech word. A so-called pattern matching method is used in which the first candidate word with the highest degree of approximation is used as the recognized word. In other words, when the distance between the input speech word and the standard word is used as similarity information, a plurality of standard words within a predetermined distance range from the input speech word are considered as candidate words, and the first candidate word with the smallest distance among them is the candidate word. It is considered a recognized word. In this case, the recognition rate can be increased by reducing the distance, which is the similarity information for selecting a candidate word, and selecting a word to be recognized only when the recognition result is more reliable.

In this way, in conventional speech recognition control methods, the effective boundary of the similarity information that is considered to be recognized is strengthened, and recognition is treated only when the recognition result is more certain, thereby improving the recognition rate and preventing false recognition. was decreasing.

Furthermore, if there is a possibility that a dangerous situation may occur due to abnormal operation due to erroneous recognition, an emergency stop switch is provided, and in the event of abnormal operation, the switch is operated to stop the operation to prevent danger.

[Problem that the invention seeks to solve]

As described above, the conventional speech recognition control method increases the recognition rate by strengthening the effective boundary of the similarity information and using the first candidate word as the recognition result.

However, it is difficult to reliably prevent erroneous recognition by simply strengthening the effective boundaries of the similarity information, and the recognition rate is limited because the first candidate among the candidate words to be recognized is immediately recognized as the recognition result. was there. Therefore, there is a problem in that it is not possible to eliminate the possibility of abnormal operation occurring due to misrecognition.

Furthermore, this limits the range in which speech recognition devices can be used in fields where abnormal operations occur due to misrecognition, and
This has the disadvantage of requiring separate measures to prevent danger.

SUMMARY OF THE INVENTION An object of the present invention is to provide an improved speech recognition control method that is capable of selecting safe recognition words that are accurate and free from the risk of abnormalities from input speech words with a simple configuration.

[Means for solving problems]

First, the principle of the present invention will be explained. When performing speech recognition based on the similarity information to the input speech word and selecting a plurality of candidate words including the first candidate word, it is not necessarily guaranteed that the first candidate word matches the input speech word, but the candidate word Looking at the group as a whole, the probability that there is one that matches the input speech word is extremely high, and in reality, the probability can be considered to be 1.

On the other hand, some spoken words are likely to behave abnormally if they are mistakenly recognized as other spoken words, but they may also behave abnormally even if other spoken words are mistakenly recognized as the aforementioned spoken word. There are things that are unlikely to occur.

Therefore, if such a spoken word exists in the candidate word group, if the candidate word is used as the recognition result, no abnormal operation will occur, so fail-safe operation can be performed. Hereinafter, this audio word or another word that is likely to be misrecognized is referred to as a first type fail-safe processing target word (first type F-3 word).

In addition, the device that uses the recognition results instructs the device to perform actions or processes that cannot be performed simultaneously, or contradictory actions or processes, such as “right” and “left” or “front” and “back”. There is a group of spoken words that

If such a spoken word group exists in the candidate word group, there is a high possibility that no matter which candidate word is recognized as a result, it will be misrecognized and cause an abnormal operation. Fail-safe can be achieved by disabling the .

Hereinafter, such spoken words will be referred to as second-type fail-safe processing target words (second-type F-3 words).

In addition, words that are subject to fail-safe processing due to malfunctions, abnormal operations, etc., including the first and second types of fail-safe processing, are referred to as fail-safe processing target words.

Based on the above-mentioned idea, the present invention eliminates fail-safe processing target words from recognized words to obtain accurate and safe recognition results, and eliminates fail-safe processing target words from candidate word groups. In this case, fail-safe processing is performed to ensure the safety of the equipment used.

Hereinafter, the solution adopted by the present invention will be explained with reference to FIG. FIG. 1 is a flowchart showing the basic configuration of the present invention.

In FIG. 1, Sl and S2 are processing steps, and in step S1, when the first candidate word is not included in the word group to be subjected to fail-safe processing, the first candidate word is Processing is performed to obtain a recognition result.

In step S2, if the first candidate word is included in the word group subject to fail-safe processing candidates, if the candidate word group does not include other words subject to fail-safe processing, the first candidate word is・If the candidate word group includes other fail-safe processing target words, fail-safe processing is performed.

[For production]

Speech recognition is performed based on the similarity information for input sounds and voice words, and multiple candidate words to be recognized are selected in descending order of similarity, and the word with the highest similarity among these candidate words is selected as the It is selected as one candidate word.

First, it is determined whether the first candidate word is included in a group of words that are candidates for fail-safe processing. If it is not included, there is no risk of abnormal behavior even if the first candidate is used as the recognition result, so the first candidate with the highest degree of approximation is output as the recognized word (step S+) - first candidate If the word is included in the fail-safe processing candidate word group, it is determined whether the candidate word group includes other fail-safe processing candidate words. If it is not included, there is no risk of abnormal operation occurring even if the first candidate word is used as the recognition result, so the first candidate word is output as the recognized word.

If it exists, fail-safe processing is performed (step SZ).

The fail-safe processing candidate word group and the contents of the fail-safe processing are appropriately determined in accordance with the contents of the device using the speech recognition results (the specific contents will be explained in the next example section).

By doing the above, it is possible to obtain safe recognition words that are accurate and free from any risk of abnormal operation. Furthermore, if there is any doubt about the safety of the recognition results, fail-safe processing is performed, so the safety of the device used can be guaranteed without the need for special safety measures. Thereby, the field of use of the speech recognition device can be expanded.

〔Example〕

Each embodiment of the present invention will be described with reference to FIGS. 2 to 6. FIG. 2 is an explanatory diagram of the voice recognition control device used to implement each embodiment of the present invention, FIG. 3 is an explanatory diagram of the first type table used in the first embodiment, and FIG. 4 is an explanatory diagram of the first type table used in the first embodiment. FIG. 5 is an explanatory diagram of the second type table used in the second embodiment, and FIG. 6 is a processing flowchart of the second embodiment.

(A) Structure of Voice Recognition Control Apparatus In the voice recognition control apparatus shown in FIG. 2, reference numeral 11 denotes a voice input microphone, which converts input voice words into electrical voice word signals.

Reference numeral 12 denotes a speech recognition processing unit, which includes a processor (not shown) therein, processes speech recognition for input speech words based on similarity information, and selects a group of candidate words to be recognized and the first word with the highest similarity. Output one candidate.

13 is a main memory, which stores programs, data, and tables (131 or 13) for executing the voice recognition control of the present invention;
2) is stored. 131 is the first type table, 132 is the second type table, and they are shown in Figures 3 and 5.
This will be explained with a diagram.

A main processing unit 14 includes a processor (not shown) therein and performs voice recognition control according to the program stored in the main memory 13.

Reference numeral 15 denotes an external communication processing unit that performs processing to transfer the recognition result or fail-safe processing contents to a host (not shown).

(B) First Embodiment A first embodiment of the present invention will be described with reference to FIGS. 3 and 4.

The first example is an example that targets the first type of fail-safe processing target words.

First, the first type table 131 shown in FIG. 3 will be explained. In FIG. 3, the left column is a word column in which words to be recognized are shown. "Right, left, up, down, stop" indicates the right, left, up, down, and stop movements, respectively, and rl
, 2J indicate the moving distance.

The right column is a constraint instruction column, which indicates which of the first type fail-safe processing target words are constraint instruction words. A constraint instruction word is likely to cause an abnormal operation if it is erroneously recognized as another word, but there is no risk of abnormal operation occurring even if another word is erroneously recognized as a constraint instruction word.

In the center is a constraint column in which constraint words among the words subject to type 1 fail-safe processing are indicated. The constraint word is
The above-mentioned constraint instruction word is another word that is likely to be misrecognized.

These constraint condition words and constraint instruction words are selected in accordance with the device using the recognition result. When selecting each of these words, the judgment is made not only based on the recognition rate, but also based on factors such as how close they are to other words, and how easy they are to pronounce.

In the case of FIG. 3, "stop" is selected as the constraint instruction word, and "right, left, top, bottom" are selected as the constraint condition words. This first
The seed table 131 is created in advance and stored in the main memory 13.

Next, the operation of the first embodiment will be explained according to the steps with reference to the processing flowchart of FIG. 4. In the processing flowchart of FIG. 4, steps S and S2
The details are as explained in FIG.

Steps 81□ to S and 14 indicate the internal processing of step S1, and steps 5t11 and 32+□ indicate the internal processing of step S2, respectively. Note that the Sth step 4 is shared by steps S and S2.

■Step 5ill l szz The main processing section 14 instructs the speech recognition processing section 12 to start accepting voice input (Step 5111).

According to this instruction, the speech recognition processing section 12 accepts and accepts the speech input, and performs recognition processing on the input speech words. That is, the input speech word data is compared with the reference speech data (for example, standard pattern data), and a plurality of candidate words are selected in descending order of similarity and are sent back to the main processing unit 14 (
Step 5 IIz) O When selecting a plurality of candidate word groups, judgment is made based on the distribution state of the degree of approximation using, for example, a statistical method, and usually 3 to 8
candidate words are selected.

The candidate word with the highest degree of approximation is the first candidate word.

■ Step Sll: l , 5114 The main processing unit 14 refers to the first type table 131, and the speech recognition processing unit 1
It is determined whether the first candidate word returned from No. 2 is a constraint word (step Sll: l).

If the first candidate word is not a constraint word, there is no risk of abnormal operation even if the first candidate word is used as the recognition result, so the first candidate word with the highest degree of approximation is used as the recognition result, and The data is sent to a host (not shown) (step S4).

For example, if the first candidate word is "1", "1" is the recognition result.

■ Step S'z++ y S++a p Sz
+z step "If the first candidate word is a constraint word in S113, the main processing unit 14 further refers to the first type table 131 and determines whether a constraint instruction word is included in the candidate word group. (Step S2□).

If the constraint instruction word is not included, there is no risk of abnormal operation occurring even if the first candidate word is used as the recognition result, so a process is performed in which the first candidate word is used as the recognition word (step Sl'+4).

If the constraint instruction word is included, the constraint instruction word is used as the recognition word instead of the first candidate word, and the external communication processing unit 15
(Step Sol□). for example,
"Right" is the first candidate word, but if the constraint instruction word "stop" is included in the candidate word group, the first candidate word "stop" is included in the candidate word group.
Instead of "right", the constraint instruction word "stop" is used as the recognition result. This allows fail-safe operation.

In addition, 'Step 5I13 t sz++ and 511
The process of 4 is equivalent to the process of using the constraint instruction word as a recognition result when the constraint instruction word is included in the candidate word group regardless of the presence or absence of the constraint condition word. Therefore, such processing is also included in the first embodiment.

(C) Second Embodiment A second embodiment of the present invention will be described with reference to FIGS. 5 and 6. The second example is an example that targets words to be subjected to the second type of fail-safe processing. .

First, the second type table 132 shown in FIG. 5 will be explained. In FIG. 5, the left column is a word column in which words to be recognized are shown. Similar to the first type table 131 in FIG. 3, "right, left, up, down, stop" means right, left,
It instructs the up, down and stop movements, and rl and 2J instruct the moving distance.

The right column is a 2nd type F-3 word column, in which 2nd type fail-safe processing target words (2nd type F-3 words) are shown among the recognition target words.

This second type F-3 word is also selected in accordance with the device using the recognition result, similar to the constraint condition word and constraint instruction word of the first type F-3 word. When selecting each of these words, the judgment is made not only based on the recognition rate, but also based on their mutual similarity to other words, ease of pronouncing, and other factors.

In the case of Fig. 5, "right, left, top, bottom" are Type 2 F-
It has been selected as a 5fI-word. This second type table 13
2 is created in advance and stored in the main memory 13.

Next, the operation of the second embodiment will be explained according to the steps with reference to the processing flowchart of FIG. In the processing flowchart of FIG. 6, steps S1 and S2
The details are as explained in FIG. Steps SI□ and -5I24 indicate the internal processing of step S1, and steps S2□1 and S2° indicate the internal processing of step S2, respectively.

Note that step SI□4 is shared by steps S and S2.

■ Step SI□ItSI□2 The main processing unit 14 instructs the voice recognition processing unit 12 to start accepting voice input (Step S1□).

In accordance with this instruction, the speech recognition processing section 12 accepts the speech input, performs recognition processing on the input speech words, selects a plurality of candidate words in descending order of degree of approximation, and replies to the main processing section 14 . The candidate word with the highest degree of approximation is the first candidate word (Step 5Izz). ■ Step 5I23 23124 The main processing unit 14 refers to the second type table 132 and selects the first candidate word replied from the speech recognition processing unit 12. Determine whether the first candidate word is a type 2 F-3 word (step 5IZ3). If the first candidate word is not a type 2 F-3 word, abnormal behavior is detected even if the first candidate word is the recognition result. Since there is no risk of this occurring, the first candidate word with the highest degree of approximation is used as the recognition result.
Send to the host through the external communication processing unit 15 (step 5
124).

For example, if the first candidate word is rlJ, "1" is the recognition result.

■ Step 3221) Slz<t 5zzz
In step S1□3, the first candidate word is type 2 F-3.
If it is a word, the main processing unit 14 further refers to the second type table 132 and determines whether the second type F-3 word is included in the candidate word group (step Stt+).

If the second type F-3 word is not included, there is no risk of abnormal operation even if the first candidate word is used as the recognition result, so processing is performed that uses the first candidate word as the recognition result (non-chip 5I24 ).

If Type 2 F-3 words are included, there is a high possibility that any candidate word will be misrecognized and cause an abnormal operation. The host is notified of the rejection through the processing unit 15.

This allows fail-safe operation.

(D) Third Example If either the first type or the second type F-3 word is included in the candidate word group, there is a risk that an abnormal operation may occur.

In the third embodiment, the first type or the second type F is included in the candidate word group.
-S Fail-safe processing is performed when the word is included.

Specifically, steps S1□1 and Sll of the first embodiment
□ (or step 3121 and S1□2 of the second embodiment
), each process after step 31.3 of the first embodiment and each process after step 5123 of the second embodiment are performed in parallel.

That is, in the third embodiment, the fail-safe processing target word in step Sl is a constraint word or the second type of fail-safe processing target word, and in the former case, another fail-safe processing target word in step S2 The processing target word is a constraint instruction word. In the latter case, the other fail-safe processing target words in step S2 are the second type of fail-safe processing target words.

The processing contents of each step in the third embodiment are clear from the description of the first and second embodiments, so a description thereof will be omitted.

〔Effect of the invention〕

As explained above, according to the present invention, the following effects can be obtained.

(1) Accurate and safe recognition words that are free from abnormal behavior can be obtained for input speech.

(2) If there is any doubt about the safety of the recognition result, fail.
Since safe processing is performed, the safety of the device that uses the voice recognition results can be guaranteed without providing any special safety measures.

(3) Due to (1) and (2) above, this device can be widely used in fields where its use was previously hindered due to the possibility of abnormal operation due to problems with the recognition rate of speech recognition. became.

[Brief explanation of the drawing]

Fig. 1 is an explanatory diagram of the basic configuration of the present invention, Fig. 2 is an explanatory diagram of a speech recognition device used in implementing each embodiment of the present invention, and Fig. 3 is a diagram of a type 1 speech recognition device used in the first embodiment. An explanatory diagram of the table, FIG. 4 is a processing flowchart of the first embodiment, FIG. 5 is an explanatory diagram of the second type table used in the second embodiment, and FIG. 6 is a processing flowchart of the second embodiment. It is. In FIG. 2, 11... voice input microphone, 12... voice recognition processing section, 13... main memory, 14... main processing section, 15...
External communication processing unit, 131... first type table, 132
...Second type table. Patent Applicant: Fujitsu Ltd. Basic Structure of the Invention Fig. 1 Speech recognition device used to implement each embodiment Fig. 2 Type 1 F/S Table Fig. 3 Type 2 table Fig. 5

Claims (4)

[Claims] (1) In a speech recognition control method that selects recognition words from a group of candidate words selected based on similarity information to input speech words, (A) the first candidate word with the highest similarity is fail-safe; (B) If the first candidate word is included in the fail-safe processing target word group, the first candidate word is set as a recognized word (step S_1); (B) If the first candidate word is included in the fail-safe processing target word group, other candidate words are If the fail-safe processing target word is not included, the first candidate word is used as the recognition word, and if the candidate word group includes other fail-safe processing target words, fail-safe is performed (step S
_2) A voice recognition control method characterized by: (2) There is a high possibility that abnormal behavior will occur if the word targeted for fail-safe processing in step S_2 is misrecognized by another word, but abnormal behavior will occur even if the word is misrecognized by another word. The constraint instruction word is a fearless constraint instruction word, and the other fail-safe processing target word in step S_1 is a constraint condition word with a high possibility that the constraint instruction word will be misrecognized. The voice recognition control system according to claim 1, characterized in that fail-safe processing is performed. (3) The fail-safe processing target word is a type 2 fail-safe processing target word group whose contents include actions or processes that cannot be performed simultaneously or contradictory actions or processes, and the fail-safe processing in step S_2 is performed. 2. The voice recognition control system according to claim 1, wherein the input voice word is invalidated as . (4) The fail-safe processing target word in step S_1 is a constraint condition word or the second type of fail-safe processing target word, and in the former case, the other fail-safe processing target word in step S_2 is a constraint instruction word. In the latter case, the other fail-safe processing target words in step S_2 are also second-type fail-safe processing target words, the voice recognition control method according to claim 1. .