JPH06186994A - Speech recognizing device - Google Patents

Abstract

PURPOSE:To precisely recognize an expression which is high in degree of freedom such as word order. CONSTITUTION:This device has a speech input part 1 which inputs a speech, a sentence templet storage part 2 which stores a sentence templet wherein words to be recognized, the meanings of the words, and the groups of the words are described in the appearance order in an input speech, a word dictionary 3 wherein the meanings and groups of words are described, a word detection part 4 which detects words in the input speech from a speech input part 1 by referring to the sentence templet and word dictionary 3, and a recognition part 5 which obtains the recognition result the input speech on the basis of a word string detected by the word detection part 4. Words which constitute the meaning structure of the speech input and other words are described in the sentence templet in the order of appearance in the speech input and the word detection part 4 handles chacteristics at the time of appearance in vocalization by separating between meaning characteristics and syntax characteristics to recognize even an expression which is high in the degree of freedom such as word order.

Description

Detailed Description of the Invention

The present invention relates to a voice recognition device for recognizing input voice.

[0002]

2. Description of the Related Art Conventionally, in speech recognition, it is not necessary to prepare in advance a wording or word order in which a user may input a voice, and moreover, a user's intention can be input in many kinds of wording or word order. In order to make it possible, for example, a voice recognition device disclosed in Japanese Patent Laid-Open No. 4-167176 has been proposed.

FIG. 15 is a diagram showing a configuration example of this voice recognition device. This speech recognition device includes a conceptual structure storage unit 51 that stores a plurality of conceptual structures represented by a plurality of words and a semantic relationship between these words, and a word detection unit 52 that detects words in an input speech. , A matching unit 53 that calculates the degree of coincidence between the conceptual structure and the word detected by the word detecting unit 52 for each conceptual structure, and outputs the conceptual structure having the maximum degree of coincidence among the plurality of conceptual structures as a recognition result. The recognition result determination unit 54 is included. As a result, even if a sentence has a different wording or word order, it is possible to express a sentence having the same meaning by a single conceptual structure so that even if various words or word orders are not prepared in advance, the sentence having the same meaning can be used. Is intended to be recognized as.

Furthermore, this publication also stores the possibility of the order of appearance in the input voice of words that are connected by a semantic relationship in the conceptual structure, and considers the order of appearance to detect the conceptual structure and word detection. There is also disclosed a voice recognition device that calculates the degree of coincidence with the word detected by the section, which aims to more accurately obtain the value of the degree of coincidence and select the correct conceptual structure as the recognition result. ing.

[0005]

However, if the conventional speech recognition apparatus described above is used to recognize the sentence of FIG. 16A and the sentence of FIG. In addition, the words "next month", "rest", and "meeting" are output as the recognition results, and the meaning of the sentence in FIG. 16 (a) and the meaning in FIG. 16 (b) are originally completely different. However, there is a problem that these are erroneously recognized as sentences having the same meaning.

Further, even when the word order is limited, there is a problem of the following erroneous recognition by recognizing only the words constituting the semantic structure. That is, for example, in the case where the word "tomorrow" is located at the back, when the description is made on the syntax network as shown in FIG.
In (a), “#” is a symbol indicating meaning, and the voice input is as shown in FIG. 17B, and “gaishai (gasi
When there was a sound similar to "tai)" and "tomorrow (asita)", the word strings were sometimes falsely detected, such as "Ginza", "going out", and "tomorrow".

There is also a problem with the unit for describing the word order. That is, in Japanese, verbs have the property that they appear after case elements, and in the above-described conventional speech recognition device, a method of describing the appearance order of words that make up the semantic structure is used based on this. There is. However, for example, in the sentences shown in FIGS. 18 (a) and 18 (b), “going out” is a semantic predicate.
In the sentence of 8 (b), it is not a predicate syntactically (word order) (the "tomorrow" is located behind it). From this, it can be seen that, when considering the word order of sentences, it is the syntactic verb that is positioned last, not the semantic verb that constitutes the conceptual structure. In other words, the semantic elements appearing in the conceptual structure are inappropriate as a unit for describing the word order,
If it is not a syntactic element, the correct description cannot be made.

An object of the present invention is to improve the above-mentioned conventional drawbacks, and an object thereof is to provide a voice recognition apparatus capable of accurately recognizing expressions having a high degree of freedom such as word order.

[0009]

In order to achieve the above object, the invention according to claim 1 is such that a voice input means for inputting voice, a word to be recognized, a meaning of a word, and a group of words are input. Word detecting means for detecting a word from an input voice from the voice input means by referring to a sentence template described according to the order of appearance in the voice, a word dictionary in which word meanings and word groups are described, It is characterized in that it has a recognition means for obtaining a recognition result for the input voice based on the word string detected by the word detection means.

According to the second aspect of the invention, the time interval between words appearing in the input speech is described in the sentence template, and the word detecting means is described in the sentence template. It is characterized in that it refers to the time interval between words and recognizes the words constituting the semantic structure and the other words, and detects the words. By recognizing the words that make up the semantic structure and other words,
It is possible to understand the meaning more accurately and obtain a highly accurate recognition result.

According to the third aspect of the invention, a sentence template describes a plurality of words or meanings of words that co-occur, and the word detecting means uses the co-occurrence relation described in the sentence template. Therefore, the feature is that the word is limitedly detected. By using the co-occurrence relation, the words to be searched can be limited and efficient recognition can be performed. It is especially effective for expressions with a high rate of co-occurrence, such as response.

Further, in the invention according to claim 4, a word having a property of co-occurring with a specific word in a sentence template may be a word having co-occurrence information in addition to the word or the meaning of the word. It is described that the word detecting means is configured to detect a word having co-occurrence information by also referring to the co-occurrence information. In the sentence template,
By describing that the word has co-occurrence information, it is possible to reduce the description amount of the sentence template and separate the semantic property and the syntactic property to create more accurate data.

Further, in the invention according to claim 5, word repetition is described in the sentence template, and the word detecting means describes the word repetition in the sentence template when a certain word is detected. In this case, it is characterized in that a word having the same property is detected later in time than the detection position of this word. As a result, it is possible to efficiently process and create data such as a noun phrase that fills a case pattern and in which the order is quite free. It can be described separately from the syntactic properties such as.

According to the sixth aspect of the invention, the sentence template describes not only the word and the meaning of the word but also syntactic properties such as a part of speech, and the word detecting means is the syntax described in the sentence template. The feature is that the word is detected in consideration of the physical property. By adding annotations not only to meaning but also to syntactic properties such as part-of-speech to sentence templates, it is possible to narrow down the detected words to only grammatically correct sentences and improve recognition accuracy and efficiency. In addition, more accurate data can be created by separating the semantic and syntactic properties.

In the invention according to claim 7, the sentence template describes information on whether or not a word can be omitted in the input voice, and the word detecting means does not detect a certain word in the input voice. Also in this case, if this word can be omitted based on the sentence template, the processing is performed so as not to affect the recognition of the entire sentence with respect to this word. As a result, only basic data needs to be created, and the description amount of the sentence template can be reduced.

[0016]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a configuration example of a voice recognition device according to the present invention. Referring to FIG. 1, this speech recognition apparatus stores a speech input unit 1 for inputting speech, and a sentence template in which a word to be recognized, a meaning of the word, and a group of words are described in the order of appearance in the input speech. A sentence template storage unit 2, a word dictionary 3 in which word meanings and word groups are described, and a word detection unit that refers to the sentence template and word dictionary 3 to detect a word from an input voice from the voice input unit 1. Four
And a recognition unit 5 that obtains a recognition result for the input voice based on the word string detected by the word detection unit 4.

In the voice recognition apparatus having such a configuration, the input voice input from the voice input unit 1 is subjected to feature extraction processing in the word detection unit 4 and spotted by using a standard pattern stored in advance. Words are detected. That is, the word detection unit 4 refers to the sentence template of the sentence template storage unit 2 when detecting a word included in the input voice, and refers to the word of the condition of each item of each sentence template in the input voice. To detect from. Then, when the term of the sentence template is the meaning of a word or the group name of the word, the word to be recognized by searching the word dictionary 3,
And its position and score (degree of coincidence with standard pattern). In this way, when the word detection unit 4 detects a word string, its position, and a score, the recognition unit 5 outputs the word string having the highest score among the detected word string scores as a recognition result. To do.

By the way, in the first embodiment of the present invention, the word detecting unit 4 is adapted to recognize words other than the words constituting the semantic structure. 2 and 3 are diagrams showing an example of a sentence template and a word dictionary, respectively, when the word detection unit 4 is adapted to recognize words other than words constituting a semantic structure. In addition, in FIG. 2, “#” is a symbol representing the meaning of a word, and the rest is the word notation itself. 2 as shown in FIG.
When the type of sentence template, “#place go out #I want to go out date”, and “#place go out” is stored and there is a voice input such as “I want to go out in Ginza”, the word detection unit 4 , Sentence template storage 2
Each sentence template of is sequentially examined from the first item.

Specifically, first, the first term of the first sentence template "# place out # want to date" is checked, and since this first term is "# place", a word whose meaning is a place. Is extracted from the word dictionary 3 in FIG. 3, and it is checked whether or not each word is in the input voice. In the present case, "Ginza", "Roppongi", and "Akasaka" are extracted from the word dictionary 3 in FIG. 3, and it is checked whether or not these words are in the input voice. As a result, the word "Ginza" is detected, and the position of the word in the input voice (in this case, the head position) and the score (degree of coincidence with the standard pattern) are obtained. Next, the second term of the first sentence template is examined. Since this second term is "going out", the word "going out" is behind the word "Ginza" in the input voice. Check whether or not, that is, whether or not the position is behind "going out". As a result, "going out" is detected immediately after "Ginza", and the position and score of this word are obtained. Next, the third term of the first sentence template is examined, and the word having the meaning of date is extracted from the word dictionary 3 of FIG. 3 because the third term is “#date”.
In the present case, "Tomorrow", "July 7th", "September 5th",
... are extracted from the dictionary 3 and it is checked whether or not these words are in a position after "going out" in the input voice. In the above input voice, "I want to do it" (gasitai)
Since there is a part similar to "tomorrow" (asita), "tomorrow" is detected in this case. Next, examine the fourth term of the first sentence template, and because this fourth term is "I want to", it is difficult to find "I want" in the fourth term after the position of "tomorrow", but it has already been detected. Since "Tomorrow" is located at the end of the voice section, "I want to" in the fourth term is either not detected or has a very low score.

In this way, after detecting the word corresponding to each term of the first sentence template, and obtaining its position and score, the second sentence template "#place" is obtained.
Similar detection processing is performed for "I want to go out". As a result, the words “Ginza”, “go out”, and “want to” are detected, and the positions and scores of these words are obtained.

The recognizing unit 5 recognizes, for each sentence template, the total score of each term of the sentence template as the degree of similarity between the input voice and this sentence template. In this case, the total score of each term of the first sentence template is the similarity between the input speech and the first sentence template, and the total score of each term of the second sentence template is , The similarity between the input speech and the second sentence template, and of these two similarities, the word string corresponding to the sentence template having the greater similarity is output as the recognition result. In the above example, the degree of similarity with the second sentence template is the highest, and therefore the word strings “Ginza”, “go out”, and “want” are obtained as recognition results.

As described above, in the first embodiment, by recognizing words other than the words forming the semantic structure, a more accurate recognition result can be obtained.

In a second embodiment of the present invention, the sentence template stores information about the time interval between words, that is, the time interval between words that appear in the input speech. Thus, the words that make up the semantic structure and the other words are recognized.

FIG. 4 is a diagram showing an example of a sentence template in the second embodiment. This sentence template
It describes the meaning and notation of the words that appear and the interval between each word. In FIG. 4, "+" is a flag that is added to the annex that has no time interval from the immediately preceding word. That is, in the second embodiment, for example, unnecessary words and the like do not generally enter between the independent particle and the particle that follows the independent word, but can enter between independent words. Since the target interval is different, this information is also described in the sentence template.

Further, in this case, the word detection unit 4 detects not only the independent word but also the word of the unit in which the auxiliary word is connected to the independent word as the recognition candidate when the word is detected by the sentence template. Has become.

The sentence template storage unit 2 stores two kinds of sentence templates "# date + # person name # ticket # buy" and "# date + # person name # ticket # buy" as shown in FIG. In this case, for example, "July 7
On the day, when there is a voice input such as "I want to buy a ticket for an erbrecker concert," the word detection unit 4
Each sentence template stored in the sentence template storage unit 2 is sequentially checked from the first item.

Specifically, first, with respect to the first term of the first sentence template, it is checked whether or not there is "+" in the immediately following term (second term). If there is "+", the second term has no time interval from the first term, so the word to be detected is the first term,
It is assumed that the first term and the second term are two connected words. That is,
Two words "July 7th" and "July 7th" are detected from the voice input.

Next, regarding "#personal name" in the third term, since there is no "+" in the immediately following term, only this term will be recognized.
That is, words whose meaning is a person's name are extracted from the word dictionary 3,
The words extracted at the position behind the preceding word in the input voice are detected. As a result, "Breaker" is detected in the input voice. Then, in the same way, "ticket",
"I want to buy" can be detected.

In this way, after detecting the word corresponding to each term of the first sentence template and obtaining the position and score thereof, the same processing is performed for the second sentence template. Calculate the total score for each. In the case of the present example, the total of the first score is the highest, and “on July 7”, “Breaker”, “ticket”, and “want to buy” are obtained as recognition results.

As described above, in the second embodiment, the words constituting the semantic structure and the other words can be recognized by defining the time interval between the words to be recognized, so that the meaning can be understood more correctly. You can

Further, in the third embodiment of the present invention, a sentence template stores a plurality of co-occurring words or meanings of the words and the co-occurrence relation is used to limit the words to be searched. It is designed to be recognized.

FIG. 5 is a diagram showing an example of a sentence template in the third embodiment. This sentence template
The words or the meanings of the words are arranged in the order in which they appear during voice input. The co-occurring words have a flag "@" and a flag "@".
"@" Is added. In particular, "@" is set for the person who is the center of the meaning. Further, in this case, the word detecting unit 4 determines that the input voice and the sentence template match only when all the co-occurring words in the sentence template are detected, and one of them is not detected. In some cases, it is determined that they do not match.

When two types of sentence templates "@ when # go out @@ do you want to go?" And "# date # go out?" Are stored in the sentence template storage 2 as shown in FIG. When there is a voice input such as "When do you want to go out?"
Examine each sentence template in turn.

Specifically, the word "when" having "@" in the item in the first sentence template is first detected, and its position and score are obtained. If the score is higher than a predetermined value, the word "Masuka" of the flag "@@" that co-occurs with it next
To detect later than "when". If it can be detected, draw the word dictionary for the remaining term "#outgoing",
The word whose meaning is out is detected at a position between “when” and “masu”. Regarding the sentence template with the co-occurrence relation specified, it is judged as coincident only when all the terms with the co-occurrence relation are recognized.

In the two types of sentence templates shown in FIG. 5,
The first sentence template has the highest total score,
As a recognition result, the word strings “when”, “go out”, and “do you do” can be obtained.

When the voice input is, for example, "I'm going out on September 5th", the sounds are similar to "When" and "When", but "Masuka" is not recognized. It does not match the sentence template "@ when # go out @ @ masashi", but does match the second sentence template "# date # go out." As a recognition result, the word string "September 5",
You can get "go out" and "do you".

As described above, in the third embodiment, by using the co-occurrence relation, the words to be searched can be limited and efficient recognition can be realized. It is especially effective for expressions with a high rate of co-occurrence, such as response.

Further, in the fourth embodiment of the present invention, in the sentence template of the third embodiment, a word having the property of co-occurring with a specific word is not limited to the word and the meaning of the word. , A word with co-occurrence information is also described so that the word can be detected. In the fourth embodiment, the configuration of FIG. 1 is modified as shown in FIG. 6, and for words having co-occurrence information, this co-occurrence information is stored in the word co-occurrence information storage unit 6 of FIG. It has become so. FIG. 7 is a diagram showing an example of the word co-occurrence information stored in the word co-occurrence information storage unit 6. This word co-occurrence information is a pair of co-occurring words or meanings. In the example of FIG. 7, the meanings of words or words having a case pattern and the meanings of words filling the case are related to the words having a case relationship. It has been described.

FIG. 8 is a diagram showing an example of the sentence template in the fourth embodiment. Similarly to the sentence template shown in FIG. 5, this sentence template also has the words or the meanings of the words arranged in the order in which the voice appears.
In the sentence template of FIG. 8, a flag “&” of a word having information in the word co-occurrence information storage unit 6 and a flag “&&” are attached to a word co-occurring with the flag. In particular, "&" is set for words that have a case pattern or words that are semantically central (such as verbs). Further, in this case, the word detection unit 4 also detects the word having the co-occurrence information by also referring to the co-occurrence information in the word co-occurrence information storage unit 6.

For example, when there is a voice input such as "go out tomorrow", the word detection unit 4 sequentially checks the sentence templates in the sentence template storage unit 2. Now, when the sentence template storage unit 2 stores a sentence template “&& item & out” as shown in FIG.
A word (“going out”) having a flag “&” in the item in the sentence template is detected first in the input voice. When a word is detected with a score equal to or higher than a certain score, the word co-occurrence information storage unit 6 is searched for a word (meaning) that co-occurs with “going out”. When the word co-occurrence information storage unit 6 stores the word co-occurrence information as shown in FIG. 7, the word co-occurrence information is “outgoing # date # place”.
Will be searched. Next, the term of the "&&" flag that co-occurs with this "going out" is detected. Since the term of "&&" flag is "term"("&&term" is the meaning of the term of case pattern), it detects the one that is located in time before "going out" and the meaning is date and place. . "Tomorrow" corresponding to "#date" is detected before "going out". As a result, the word strings “tomorrow” and “going out” are recognized. For example, if the score of the word having the word co-occurrence information (the word with the flag “&”) is lower than a preset value, it is determined that the sentence templates do not match, and the co-occurring word is not detected and the next Process sentence templates.

As described above, efficient word detection can be performed by using the co-occurrence relation. In the past, for a word that co-occurs with a specific word (for example, case slot), the meaning of the word was written directly on the sentence template or syntactic network, so sentence templates and networks are prepared for the patterns of the co-occurring meaning. However, according to the fourth embodiment, by describing that the word has co-occurrence information, the description amount of the sentence template is reduced and the semantic property and the syntactic property are reduced. More accurate data can be created by cutting.

In addition, the fifth embodiment of the present invention is designed to store word repetitions in a sentence template for recognition. That is, when a plurality of words having the same property can be arranged, the information is described in the sentence template and word detection is performed.

FIG. 9 is a diagram showing an example of a sentence template in the fifth embodiment. This sentence template
The words or the meanings of the words are arranged in the order in which they appear in time, and “*” is given as a flag for repeatedly appearing words of the same nature. In this case, the word detector 4
When a word is detected, if a plurality of pieces of information that can be arranged are described in the sentence template, the word having the same property is detected later in time than the detected position. In addition, in the fifth embodiment, the structure of the voice recognition device may be any one of FIG. 1 and FIG. 6, but in particular, by providing the word co-occurrence information storage unit 6 as shown in FIG. , It becomes easy to detect words that have a case relationship.
In other words, the word detection unit 4 can repeatedly detect the words to be detected by narrowing down to words that co-occur with a specific word.

For example, when there is a voice input such as "go out to Ginza tomorrow", the word detection unit 4 first checks the sentence templates in the sentence template storage unit 2 in order. The voice recognition device is configured as shown in FIG. 6, and the sentence template storage unit 2 stores the sentence template “*” as shown in FIG.
In the case where “&& item & go out” is stored, the word detection unit 4 checks whether “go out” having information in the word co-occurrence information storage unit 6 is included in the input voice. As a result, if “going out” can be detected, “going out” is detected in the word co-occurrence information storage unit 6.
Search for words that co-occur with. When the word co-occurrence information storage unit 6 stores the word co-occurrence information as shown in FIG. 7, it is possible to search “go out #date #place” as the word co-occurrence information. Next, the word detection unit 4 detects the word “&& term” that co-occurs with “going out” from the sentence template. That is, it is checked whether "#date" and "#place" that co-occur with "going out" are before the position where "going out" is detected from the beginning of voice input. As a result, “Ginza” corresponding to “#place” is detected before “going out”. "&& In process
In the “term”, there is a flag “*” in which words of the same nature appear repeatedly, so there is the remaining co-occurrence word “#date” after the detected position “Ginza” and before the “going out” position. Find out if As a result, the word "tomorrow" is detected. As a result, the word strings "Ginza", "tomorrow",
Can recognize "going out". If other sentence templates have the same word order, such as "# place # date # outgoing", the total score will be the same, but since noun phrases are generally arranged in the order of date and place, Figure 10
Often prepare a sentence template as shown in. However, with the sentence template as shown in FIG.
Since the word order of "# place" and "# date" is different from that of the input voice, the total score is low and it is difficult to recognize the input voice. On the other hand, in the fifth embodiment, it is possible to correctly recognize even if the word order is different.

As described above, in the fifth embodiment, it is possible to efficiently perform processing and data creation for a plurality of noun phrases that line up the case pattern and whose order is considerably free. , It is possible to separate and describe semantic properties such as case patterns and syntactic properties such as word order.

In addition, in the sixth embodiment of the present invention, even if words having the same meaning, the form appearing on the utterance is often due to the syntactical property rather than the semantic property. Focusing attention, the sentence template stores not only the words and the meanings of the words but also the syntactic properties such as the part of speech appearing therein, so that the detected words are limited to only the grammatically correct sentences.

FIG. 11 is a diagram showing an example of a sentence template in the sixth embodiment. This sentence template describes not only the word and the meaning of the word but also syntactic properties such as a part of speech that appears therein. In the example of FIG.
After ":", syntactic properties such as a part of speech are described.
In this case, the word dictionary 3 is composed of word notations, meanings, and parts of speech, as shown in FIG. 12, for example.

In the sentence template storage unit 2, three types of sentence templates shown in FIG.
When "I want to date", "# date # requirement: sahenon", and "# date # requirement: verb" are stored, if there is a voice input such as "go tomorrow", the word detection unit 4 sequentially examines the sentence templates stored in the sentence template storage unit 2 from the first item.

Specifically, first, the first item of the first sentence template is examined. In this case, the first item of the first sentence template is "# requirement", and its part of speech is Sahen noun, so the word whose meaning is the subject and whose part of speech is Sahen noun is extracted from the word dictionary 3. To do. In the present case, "go out"
Is extracted from the dictionary 3, and it is checked whether or not this word “going out” is present in the input voice. When the word “going out” is detected in the input voice, the position and the score are obtained. In this way, each item of each sentence template is sequentially detected by the word detection unit 4, and the detected word, score, and detection position are obtained. In the above example, the third sentence template “# Date # Message: Verb” best matches, and as a result, “Tomorrow” and “Go out” can be obtained.

As described above, in the sixth embodiment, not only the meaning but also the annotation based on the syntactic property such as the part of speech is added to narrow down the detected words to only the grammatically correct sentences, thereby improving the recognition accuracy, The efficiency can be improved. Also,
More accurate data can be created by separating the semantic and syntactic properties.

In the seventh embodiment of the present invention, a sentence template is written with information as to whether or not a word can be omitted in the input voice, and a word such as a case element or an adjunct word is easily recognized. Intended to be possible.

FIG. 13 is a diagram showing an example of a sentence template in the seventh embodiment. In this sentence template, each term is a word to be detected during voice input, or a meaning of the word, a group of words, and a flag "%" is attached to an optional term. In this case, the word detection unit 4 detects a word, a position, and a score that satisfy the conditions of each term of the sentence template in the input voice, but if the sentence template has an optional symbol, the word is not detected. Sometimes the score doesn't get worse.

A sentence template "% # want to go out of place" as shown in FIG. 13 is stored in the sentence template storage unit 2.
("%" Is the meaning of the word; otherwise, this notation itself)
If there is a voice input such as “I want to go out” when is stored, the word detection unit 4 sequentially selects the sentence templates stored in the sentence template storage unit 2.
Check from that one item. In the sentence template of FIG. 13, since the first term is “#place”, words whose meaning is place are extracted from the word dictionary 3 and it is checked whether or not each word is being input by voice. If the word dictionary 3 is as shown in FIG. 3, the word that is the place is not detected, but since there is a flag “%” and it can be omitted, it affects the score and recognition when the scores do not match. A score that does not affect the recognition, for example, "0" is set, instead of a non-existent score, for example, "-0.5".

Next, the word corresponding to the second item "going out" of the sentence template can be examined in the input voice to detect "going out". Next, since the third term of the sentence template is "I want to", the corresponding word in the input voice is detected at a position after "going out". As a result of such processing, the result as shown in FIG. 14 is obtained, and the word strings “go out” and “want to” are obtained as recognition results.

As described above, in the seventh embodiment, the sentence template including the words and expressions that are easily omitted does not significantly reduce the score even if these words are not detected. , It becomes possible to recognize words that are easily omitted such as attached words. Further, conventionally, it was necessary to create at least two kinds of data depending on whether the data was omitted or not. According to the seventh embodiment, only basic data needs to be created. The amount of template description can be reduced.

[0056]

As described above, according to the present invention, when words that make up the semantic structure of voice input and other words are described based on the order in which they appear in voice input, and appear in utterances. Since the properties of are treated separately as semantic and syntactic, expressions with a high degree of freedom such as word order can be accurately recognized.

In particular, according to the second aspect of the invention, the sentence template describes the time interval between the words appearing in the input voice, and the word detecting means uses the sentence template. Since the words constituting the semantic structure and the other words are recognized by referring to the time intervals between the written words and the words are detected, the correct meaning can be understood.

According to the third aspect of the present invention, the sentence template describes a plurality of co-occurring words or meanings of the words, and the word detecting means includes the co-occurrence relation described in the sentence template. By using, the words can be detected in a limited manner, which enables efficient recognition, and is particularly effective for expressions such as co-occurrence that have a high co-occurrence ratio.

Further, according to the invention of claim 4, in the sentence template, not only the word or the meaning of the word but also the word having the co-occurrence information is included in the word having the property of co-occurring with the specific word. It is also described that the word detection means performs word detection for a word having co-occurrence information by also referring to the co-occurrence information. More accurate data can be created by separating the semantic and syntactic properties.

According to the invention described in claim 5, the sentence template describes the repetition of words, and the word detecting means repeats the words in the sentence template when a certain word is detected. Is described, since it is designed to detect a word of the same property temporally after the detection position of the word, a plurality of words are arranged like a noun phrase that fills a case pattern, and It is possible to efficiently process and create data whose order is fairly free, and to describe the semantic properties such as case pattern and the syntactic properties such as word order separately.

According to the sixth aspect of the invention, the sentence template describes not only the word and the meaning of the word but also syntactic properties such as a part of speech, and the word detecting means is described in the sentence template. Since the words are detected in consideration of the syntactic properties that have been obtained, it is possible to improve the recognition accuracy and efficiency by limiting the detected words to only the grammatically correct sentences. In addition, more accurate data can be created by separating the semantic and syntactic properties.

Further, according to the invention of claim 7, the sentence template describes information on whether or not a word can be omitted in the input voice, and the word detecting means detects a word in the input voice. Even if it is not detected in, if this word can be omitted based on the sentence template, processing that does not affect the recognition of the entire sentence with respect to this word is performed.
The user only needs to create the basic data and can reduce the description amount of the sentence template.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration example of a voice recognition device according to the present invention.

FIG. 2 is a diagram showing an example of a sentence template.

FIG. 3 is a diagram showing an example of a word dictionary.

FIG. 4 is a diagram showing an example of a sentence template.

FIG. 5 is a diagram showing an example of a sentence template.

FIG. 6 is a diagram showing another configuration example of the voice recognition device according to the present invention.

FIG. 7 is a diagram showing an example of word co-occurrence information.

FIG. 8 is a diagram showing an example of a sentence template.

FIG. 9 is a diagram showing an example of a sentence template.

FIG. 10 is a diagram showing an example of a sentence template.

FIG. 11 is a diagram showing an example of a sentence template.

FIG. 12 is a diagram showing an example of a word dictionary.

FIG. 13 is a diagram showing an example of a sentence template.

FIG. 14 is a diagram showing a specific example of word detection processing.

FIG. 15 is a configuration diagram of a conventional voice recognition device.

FIG. 16 is a diagram showing an example of a sentence to be recognized.

FIG. 17 is a diagram for explaining an example of a case where an erroneous recognition is made in a conventional voice recognition device.

FIG. 18 is a diagram for explaining an example of a case where an erroneous recognition is made in a conventional voice recognition device.

[Explanation of symbols]

 1 voice input unit 2 sentence tempo rate storage unit 3 word dictionary 4 word detection unit 5 recognition unit 6 word co-occurrence information storage unit

Claims (7)

[Claims] 1. A voice input means for inputting a voice, a word to be recognized, a meaning of the word, a sentence template in which groups of words are described in the order of appearance in the input voice, and the meaning of the word,
A word detection unit that detects a word from the input voice from the voice input unit by referring to a word dictionary in which a group of words is described, and a recognition result for the input voice based on the word string detected by the word detection unit. A speech recognition apparatus having a recognition means for obtaining the speech recognition apparatus. 2. The voice recognition device according to claim 1, wherein
The sentence template describes the time interval between words that appear in the input speech, and the word detecting means refers to the time interval between words described in the sentence template. A voice recognition device characterized in that it recognizes a word constituting a semantic structure and a word other than the word and detects the word. 3. The voice recognition device according to claim 1, wherein
In the sentence template, a plurality of co-occurring words or meanings of the words are described, and the word detecting means detects the words by using the co-occurrence relation described in the sentence template. A voice recognition device characterized in that 4. The voice recognition device according to claim 3,
In the sentence template, it is described that a word having a property of co-occurring with a specific word is not only a word or a meaning of a word, but also a word having co-occurrence information. A speech recognition apparatus characterized in that a word having co-occurrence information is also detected by referring to the co-occurrence information. 5. The voice recognition device according to claim 1,
In the sentence template, word repetition is described, and when a certain word is detected, the word detection means detects a word repetition from the detection position of the word when the word repetition is described in the sentence template. A voice recognition device characterized in that it is designed to detect words of the same nature later in time. 6. The voice recognition device according to claim 1, wherein
The sentence template describes not only words and meanings of words, but also syntactic properties such as part of speech. The word detecting means detects a word in consideration of the syntactic properties described in the sentence template. A voice recognition device characterized by being adapted to perform. 7. The voice recognition device according to claim 1, wherein
The sentence template describes information on whether or not a word can be omitted in the input voice, and the word detecting means uses the word template based on the sentence template even when a word is not detected in the input voice. A voice recognition device characterized by performing a process that does not affect the recognition of the entire sentence for this word when can be omitted.