JPH06202688A - Speech recognition device - Google Patents

Abstract

PURPOSE:To increase the degree of freedom of vocalization and to improve the recognition rate of a speech by providing a case frame generating means which generates a new hypothesis for the vocalization contents of the speech on the basis of the analysis result of an analyzing means. CONSTITUTION:A purser 2 analyzes the recognition result of a speech recognition part 1 in vocalization units such as sentence units on the basis of a case frame representing the meaning of the hypothesis for the vocalization contents of the speech supplied from an interaction control part 3. The purser 2 repeats this process as to plural concept frames supplied from the interaction control part 3 and outputs the concept frame having the largest total scores described in an attribute value to the interaction control part 3 as the analysis result of the recognition result of the speech recognition part 1 once the description of the attribute values of respective slots of respective concept frames ends. The interaction control part 3 sets a hypothesis for the vocalization contents of a speech inputted to the speech recognition part on the basis of interaction control information stored previously in a domain knowledge dictionary 4 and generates a concept frame representing the meaning of the hypothesis.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus suitable for recognizing speech.

[0002]

2. Description of the Related Art In a conventional voice recognition device, for example, an input voice is word-spotted, and a word candidate string obtained as a result is parsed on a sentence-by-sentence basis using syntax information prepared in advance. Has been done.

Then, based on the result of the syntactic analysis, a correct sentence (sentence) is obtained from the word candidate sequence by removing erroneous word candidates.

[0004]

By the way, in spoken language, non-grammatical sentences are often used (less often spoken in correct grammar), and unnecessary words (for example, between clauses) are used. "Eh" is spoken to take
Is often used, which makes parsing difficult and
There was a problem that the recognition rate of voice deteriorates.

Therefore, there is a method of registering all non-grammatical sentences and sentences in which unnecessary words are inserted in the syntax information.

However, this method has a problem in that the amount of syntax information is enormous and the syntax analysis takes time.

Therefore, there is a method of limiting the utterance pattern of the utterer, but in this case, the degree of freedom of the utterance of the utterer becomes small, and there is a problem that the utterer feels annoyance.

The present invention has been made in view of such circumstances, and improves the degree of freedom of speech and the recognition rate of voice.

[0009]

A speech recognition apparatus according to claim 1, wherein a speech recognition unit 1 as a recognition means for recognizing a speech and a hypothesis of speech utterance content are set to express the hypothesis semantically. The dialog management unit 3 as a generation unit that generates a frame, and the parser 2 as an analysis unit that analyzes the recognition result of the voice recognition unit 1 based on the case frame generated by the dialog management unit 3 are provided. It is characterized in that the unit 3 generates a new hypothesis of the utterance content of the voice based on the analysis result of the parser 2.

A speech recognition apparatus according to a second aspect of the present invention is characterized in that the case frame expresses the hypothesis generated by the dialogue management unit 3 in a semantic relationship centered on a predetermined keyword.

A speech recognition apparatus according to a third aspect of the present invention is characterized in that the case frame is represented by a set of an attribute name and an attribute value.

[0012]

In the voice recognition device having the above structure, the parser 2
Then, the voice recognition result of the voice recognition unit 1 is analyzed based on the case frame that expresses the hypothesis of the utterance content of the voice generated by the dialogue management unit 3. Then, the dialogue management unit 3 is caused to generate a new hypothesis of the utterance content of the voice based on the analysis result of the parser 2. Therefore, since the voice is analyzed regardless of the word order, the degree of freedom of speech can be increased. Furthermore, since meaningless words such as unnecessary words contained in the voice are ignored, the voice recognition rate can be improved.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a schematic diagram of a voice recognition device according to the present invention.
It is a block diagram which shows the structure of one Example of a V system control apparatus. This AV system controller uses audio
It is configured so that the V device system 7 can be operated.

That is, the voice recognition unit 1 performs word spotting processing (word spotting processing) on the input voice, and from the spotting result of the word output-requested from the parser 2, that is, from the input voice. , If the word output requested by the parser 2 cannot be spotted, the information that the word does not exist is output, and the word output requested by the parser 2 is spotted from the input voice. If you can, you can spot the word name, score,
In addition, the start point and the end point (hereinafter, referred to as a detection section) of the vocalization section in the voice of the word that can be spotted are output to the parser 2.

Here, the score means the likelihood of the spotted word, and the larger the value, the more likely the spotted word is.

For example, as shown in FIG. 2, the parser 2 registers a utterance hypothesis pattern table 11 in which a case frame expressing the hypothesis of the utterance content of the voice supplied from the dialogue management unit 3 is registered, and information about words. A word dictionary 12 (FIG. 3), a word thesaurus 13 in which word concepts are described in a hierarchical structure (FIG. 4), and a priority queue 14 in which a concept frame (FIG. 5) as a case frame under analysis is registered. And a result queue 15 in which the concept frame whose analysis has been completed is registered.

Here, the concept frame means a case frame, for example, as shown in FIG. 5, which expresses a hypothesis of speech utterance content in a semantic relationship centered on a predetermined keyword such as a verb. The case frame shown in FIG.
That is, in the concept frame, the hypothesis of the utterance content of the voice is an action target centering on the verb as a keyword, the action target that exerts the action that the verb means, a means for performing that action, a method for performing that action, and that action The action starting point that starts the action and the action ending point that ends the action are described.

In other words, in the concept frame, the hypothesis of the utterance content of the voice is centered on the action (action expressed by a verb), what the action is, what, by what, where and from where. Will be expressed as follows.

Further, the concept frame is represented by a slot in which an attribute name and an attribute value are paired.

That is, in the conceptual frame shown in FIG. 5, "verb", "action target", "means", "method", "action start point", or "action end point" as attribute names, and It is expressed by a set (slot) of a name as an attribute value, a score, and a detection section.

That is, in the conceptual frame shown in FIG. 5, for example, the slot having the attribute name "verb" is the name as the verb described in the slot (for example, "record").
, "Play", etc., and its "name" is the voice recognition unit 1.
Then, the input voice has a score output by word spotting processing and an attribute value including a detection section (start point and end point).

Further, the attribute name of each slot is divided into a noun part and a particle or auxiliary verb part, and the attribute value is described for each part.

That is, for example, when "record" is described in the slot of the attribute name "verb", "record"
Is divided into "record" as a noun part and "do" as an auxiliary verb part, and the attribute value is described for each.

In FIG. 5, what is surrounded by <> means that an actual value (or word) is described.

In the slot having the attribute name "id", the id of the conceptual frame is described as an attribute value, and
In the slot with the attribute name "score", the attribute name "verb",
The total score of the score described in the attribute value “score” of each “action target”, “means”, “method”, “action start point”, or “action end point” is described as the attribute value. .

Further, in the slot of the attribute name "detection section", the attribute value "verb", "action target", "means", "method", "action start point", or "action end point" is set to the attribute value " All the detection sections described in "Detection section" are described as attribute values, and the attribute name "verb", "action target", "means",
The length of the detection section described in the attribute value “detection section” of each of “method”, “action start point”, or “action end point” (=
The total value of (end point-start point) is described as an attribute value.

Further, in the conceptual frame of FIG. 5, both particles and auxiliary verbs are shown as particles.

The parser 2 (FIG. 2) uses the case frame (conceptual frame), which is supplied from the dialogue management unit 3 and expresses the hypothesis of the utterance content of the voice, as a result of recognition by the voice recognition unit 1, for example, a sentence. Analyze in a predetermined utterance unit such as a unit.

That is, the parser 2 causes the voice recognition unit 1 to input the word described in the attribute value “name” of each slot of the conceptual frame that expresses the hypothesis of the utterance content of the voice supplied from the dialogue management unit 3. It outputs sequentially and from the input voice,
Request spotting processing for the word.

The parser 2 receives the word-spotted word, the score of the word, and the detection section (start point and end point) of the word output from the voice recognition unit 1, and outputs the score or the detection section as Described in the attribute value “score” or “detection section” of the slot having the word spotted word as the attribute value “name”, respectively.

The parser 2 repeats the above processing for a plurality of conceptual frames supplied from the dialogue management unit 3,
When the description of the attribute value of each slot in each concept frame is completed, for example, the attribute value of the slot with the attribute name "score", that is, the attribute name "verb", "action target", "means", "method", "action" The concept frame having the largest total score, for example, of the scores described in the attribute value “score” of each of “start point” or “action end point” is output to the dialogue management unit 3 as the analysis result of the recognition result of the voice recognition unit 1. To do.

The dialogue management unit 3 (FIG. 1) refers to the dialogue management information stored in advance in the domain knowledge dictionary 4, manages the flow of dialogue, and analyzes the parser 2 (concept output from the parser 2). The AV device system 7 is controlled based on the semantic representation of the frame).

That is, the dialogue management unit 3 determines, based on the dialogue management information stored in advance in the domain knowledge dictionary 4, or the semantic representation of the conceptual frame output from the parser 2.
A hypothesis of the speech content of the voice input to the voice recognition unit 1 is set, and a concept frame that expresses the hypothesis is generated.

Further, the dialogue management unit 3 refers to the dialogue management information stored in the domain knowledge dictionary 4 in advance, and replies to the semantic representation of the conceptual frame output from the parser 2, that is, is input to the voice recognition unit 1. A meaning expression of a reply to the utterance content of the voice is generated and output to the natural language generation unit 5.

The dialogue management section 3 also outputs to the AV equipment system 7 a command (control command) for causing the AV equipment system 7 to perform an operation corresponding to the semantic expression of the conceptual frame output from the parser 2.

The domain knowledge dictionary 4 stores in advance dialogue management information for managing dialogue (for example, a reply pattern to a question and a next utterance pattern following the current utterance).

The natural language generation unit 5 generates text data based on the semantic expression of the conceptual frame output from the dialogue management unit 3 and outputs it to the text-to-speech synthesis unit 6. The text-to-speech synthesis unit 6 generates synthetic speech based on the text data output from the natural language generation unit 5, and outputs it from a built-in speaker (not shown). The AV device system 7 is composed of at least one AV device and operates in response to a command output from the dialogue management unit 3.

In the AV system control device configured as described above, the dialogue management unit 3 first makes a hypothesis of the utterance content of the input voice, and the concept frame ( The attribute value of each slot in FIG. 5) is described (each slot is filled) and output to the parser 2. In the parser 2 (FIG. 2), the conceptual frame from the dialogue management unit 3 is stored in the utterance hypothesis pattern table 11.

In the dialogue management unit 3, all hypotheses that are likely to be uttered are set, and all concept frames that express the hypotheses are output sequentially. Therefore, in the parser 2, all concept frames output from the dialogue management unit 3 are sequentially stored in the utterance hypothesis pattern table 11.

Further, in the dialogue management unit 3, not all slots of the concept frame are filled (although all of them may be filled), but only slots for expressing the hypothesis made are expressed.

When a voice is first input to the voice recognition unit 1, that is, when utterance is started, the dialogue management unit 3
In this case, only the attribute value of the slot having the attribute name "verb" is described, and the conceptual frames (hereinafter referred to as blank conceptual frames) in which other attribute values are not described are sequentially output to the parser 2. ing.

Hereafter, including the blank conceptual frame,
All concept frames output from the dialogue management unit 3 to the parser 2 will be referred to as an initial concept frame group.

When the initial concept frame group is output from the dialogue management unit 3, the parser 2 refers to the word dictionary 12 and fills the slots of the initial concept frame in which the attribute values are not described.

Here, words are registered in the word dictionary 12 as shown in FIG. 3, for example.

That is, in the word dictionary 12, for example, ""
A "laser disk" that can be the target of the action (act) of "play" (which may be described as the attribute value of the slot with the attribute name "action target") is "L".
The word "L" as it is possible to say "D", "laser", "video disc", or "disc".
It is a representative term for "D", "laser", "video disc", and "disc".

All the words registered in the word dictionary 12 are described at any one of the ends of the word thesaurus 13 in which the concept of the word as shown in FIG. 4 is described in a hierarchical structure. In the word dictionary 12, the words "LD", "laser", "video disk", and the word "laser disk" as a representative word of "disk" are classified as read-only video media. "Laser disc" is the word thesaurus 13 (Fig. 4)
Means that it is described at the end of the playback-only video medium.

Therefore, the words "LD", "laser", "video disk", and "disk" typified by the word "laser disk" are also added to the end of the read-only video medium of the word thesaurus 13 (FIG. 4). It has been described.

Further, in the word dictionary 12 (FIG. 3), a semantic relationship is registered with respect to words that can be attribute values of the slot of the attribute name "verb", such as the word "record".

That is, for the word "record", for example, "
It describes the semantic relation of the act of "recording" (recording) with respect to what kind of action target, by what means, by what method, and from which action start point to which action end point.

Further, even in words such as the word "record" that can be the attribute value of the slot of the attribute name "verb", as in the case of the above-mentioned word "laser disk",
A word having that as a representative word is described, and which hierarchy of the word thesaurus 13 (FIG. 4) belongs to is described.

Information (grammar information) on the use of words such as the word "record" that can be the attribute value of the slot of the attribute name "verb" is also described.

Therefore, when the initial concept frame is output from the dialogue management section 3, the parser 2 outputs the word dictionary 12
(FIG. 3) is referred to, and the word described in the attribute value of the slot of the attribute name “verb” in the initial concept frame (FIG. 5) is detected. Then, based on the semantic relationship of the word (the attribute value of the slot of the attribute name "verb") described in the word dictionary 12, the attribute value of the slot in which the attribute value is not described among the slots of the initial concept frame. Is described.

That is, when the initial concept frame is a blank concept frame, that is, when only the word "record" is described in the attribute value of the slot of the attribute name "verb" of the initial concept frame, the parser 2 6, the description of the word “record” in the word dictionary 12 shown in FIG.
As shown in, the slots in which the attribute value is not described among the slots of the initial concept frame are filled.

On the other hand, when a voice is input to the voice recognition unit 1, word spotting processing is performed there. Then, when the word spotting process is completed in the voice recognition unit 1, a control signal for notifying that the word spotting process for the input voice is completed is output to the parser 2 together with the utterance time of the voice input therein. It

In the parser 2 (FIG. 2), the voice recognition unit 1
When the control signal from is received, first, the word described in the attribute value of the slot of the attribute name "verb" of the initial concept frame is output to the voice recognition unit 1, and from the input voice, An output request is made to output the spotting processing result of the word.

In the voice recognition unit 1, when a word is output from the parser 2 and an output request for the spotting result of the word is made, the word name, the score, and the word can be detected (spotting). The detection section (start point and end point) is output to the parser 2.

In the parser 2, the word-spotted word output from the speech recognition unit 1, the score of the word, and the detection section (start point and end point) of the word.
Is received, and the score or the detection section has the attribute value “score” or “detection section” of the slot having the word-spotted word as the attribute value “name” (in this case, the slot of the attribute name “verb” (FIG. 5)). ], Respectively.

As described above, the concept frames in which the attribute values "score" and "detection section" of the slot having the attribute name "verb" are described are sequentially transferred to and stored in the priority queue 14.

In the priority queue 14, the concept frame has the attribute value of the slot having the attribute name "score" (FIG. 5), that is, the attribute name "verb", "action target", "means",
It is sorted and stored in ascending order of the total score of the scores described in the attribute value “score” of each of “method”, “action starting point”, and “action end point”.

When the above processing is completed for all initial concept frames supplied from the dialogue management unit 3, the concept frames stored at the head of the priority queue 14 are sequentially read.

In the parser 2, the word thesaurus 13 is referred to, and a specific word is described in the attribute value of the highest priority slot among the slots in which the attribute value is not described in the concept frame.

That is, for example, as shown in FIG. 6, when a specific word is currently described only in the attribute value of the slot of the attribute name "verb" of the concept frame, the specific word is not described. Of the slots, the highest priority,
For example, instead of (video medium video display device), the word thesaurus 13
A word (for example, "laser disk") described at the end of the hierarchy of the video medium and the video display device (Fig. 4).
Or "video tape", "laser disc player", "video tape device" (neither shown), etc. are described.

Then, in the parser 2, the word described in place of (video medium video display device) in the attribute value of the slot of the attribute name "action target" is output to the voice recognition unit 1 and the input voice An output request is made to output the spotting processing result of the word from.

In the voice recognition unit 1, when a word is output from the parser 2 and an output request for the spotting result of the word is made, the word name, the score, and the word can be detected (spotting). The detection section (start point and end point) is output to the parser 2.

In the parser 2, the word-spotted word output from the voice recognition unit 1, the score of the word, and the detection section (start point and end point) of the word.
Is received, and the score or detection interval has the attribute value "score" or "detection" of the slot having the word-spotted word as the attribute value "name" (in this case, the slot having the attribute name "action target" (FIG. 5)). Section ”.

As described above, the conceptual frames in which the attribute values "score" and "detection section" of the slot having the attribute name "action target" are described are sequentially transferred to and stored in the priority queue 14.

Thereafter, the same processing is repeated, and all attribute values of the slots of the concept frame attribute names "verb", "action target", "means", "method", "action start point", and "action end point". “Score” and “detection section” are described.

In the speech recognition unit 1, when the word output from the parser 2 cannot be spotted in the input speech, the attribute values "score" and "slot" of the corresponding slot in the concept frame. Detection section "
Indicates that.

All attribute values "score" and "detection section" of the slot of the attribute name "verb", "action target", "means", "method", "action start point", and "action end point" are described. The concept frame is transferred from the priority queue 14 to the result queue 15 and stored therein.

In the result queue 15, as in the case of the priority queue 14, the concept frame has the attribute value of the slot with the attribute name "score" (FIG. 5), that is, the attribute name "verb", "action target", " It is sorted and stored in ascending order of the total score of the scores described in the attribute value “score” of each of “means”, “method”, “action starting point”, or “action end point”.

When all the concept frames stored in the priority queue 14 are transferred to and stored in the result queue 15 in the parser 2 by repeating the above process, the attribute names "verb" and "action target" of each concept frame are stored. ",
Particles or auxiliary verbs attached to each word described in the attribute value of the noun part (FIG. 5) of the slot of “means”, “method”, “starting point of action”, or “end point of action” (hereinafter, the particle part including both) Is output) (FIG. 5) is issued to the voice recognition unit 1.

Then, in the voice recognition unit 1, the noun part of the slot of the attribute name "verb", "action target", "means", "method", "action start point", or "action end point" of the concept frame (Fig. 5) The spotting result (score and detection section) of the particle part (Fig. 5) attached to each word described in the attribute value of 5) is output to the parser 2, and the spotting result is the attribute name of the conceptual frame. Described in the attribute values "score" and "detection section" of the particle part (Fig. 5) of the slot of "verb", "object of action", "means", "method", "starting point of action", or "end point of action". It

After the above processing, in the result queue 15, the concept frame shows that the attribute value of the slot of the attribute name "score" (FIG. 5), that is, the attribute name "verb", "action target", "means", " It is sorted and stored in the ascending order of the total scores of the scores described in all the attribute values “score” of the noun part and particle part of “method”, “action starting point”, or “action ending point”.

Then, the concept frame stored at the head of the result queue 15 is used as the analysis result of the recognition result of the voice recognition unit 1 based on the hypothesis of the utterance content of the voice generated by the conversation management unit 3, as a result of the dialogue management unit. 3 is output.

When the parser 2 outputs the conceptual frame as the analysis result, the dialog management unit 3 refers to the dialog management information stored in advance in the domain knowledge dictionary 4, and the meaning of the conceptual frame output from the parser 2 is referred to. A semantic expression of a reply to the expression, that is, a reply to the utterance content of the voice input to the voice recognition unit 1 is generated and output to the natural language generation unit 5 (FIG. 1).

In the natural language generation unit 5 (FIG. 1), text data is generated based on the semantic expression of the conceptual frame output from the dialogue management unit 3 and output to the text-to-speech synthesis unit 6. In the text-to-speech synthesis unit 6, synthetic speech is generated based on the text data output from the natural language generation unit 5, and output from the built-in speaker.

At the same time, in the dialogue management unit 3, the parser 2
A command (control command) that causes the AV equipment system 7 to perform an operation corresponding to the semantic expression of the conceptual frame output from
Is output to the AV device system 7.

In the AV equipment system 7, the operation corresponding to the command output from the dialogue management section 3 is performed.

That is, when the parser 2 outputs, for example, the conceptual frame of the meaning expression “CD reproduction” to the dialogue management unit 3, the AV device system 7 starts the reproduction of the CD and the text-to-speech synthesis unit 6 Will output a synthesized sound such as "CD playback has started" as a reply to "CD playback".

After that, in the dialogue management unit 3,
Along with the domain knowledge dictionary 4, the semantic representation of the concept frame output from the parser 2 (in the above case, "C
Based on "D reproduction"), the hypothesis of the utterance content of the next uttered voice can be established.

That is, in this case, the dialogue management unit 3 selects "C".
A hypothesis of the utterance content of the voice uttered after "D reproduction" is set, for example, "CD stop" or "CD fast forward".

The case where the voice recognition device of the present invention is applied to the AV system control device has been described above, but the present invention can be applied to any device that recognizes voice in addition to the AV system control device.

Although the method of word spotting in the voice recognition unit 1 is not mentioned in this embodiment, in the voice recognition unit 1, for example, the DP matching method, the HMM method, or the Japanese Patent Laid-Open No. 60-249198 is used. , JP-A-60-249199, or JP-A-60-2523.
The word spotting process can be performed based on any voice recognition algorithm such as the voice recognition algorithm of the voice recognition device disclosed in H.96.

Further, in the present embodiment, the concept frame in which the hypothesis of the speech utterance content of the voice is expressed by the semantic relationship centering on the verb is used, but the present invention is not limited to this, and the speech utterance content of the voice is not limited to this. It is possible to use the concept frame expressed by the semantic relationship centering on a predetermined keyword.

Further, in the present embodiment, the attribute values of the slot of the attribute name "score" of the concept frame (FIG. 5) are set to the attribute names "verb", "action target", "means", "method", "action". The total score of the scores described in the attribute value "score" of each of the "starting point" or the "action end point" is used, but the present invention is not limited to this. For example, the attribute name "verb", "action target", "means", "method" ], “Action start point”, or “action end point”, and the attribute value “score” and the length of the attribute value “detection section” (= end point−start point) can be summed.

Further, the attribute value of the slot of the attribute name "score" in the concept frame (FIG. 5) is changed to the attribute name "verb",
The product sum of the attribute value "score" of each "action target", "means", "method", "action start point", or "action end point" and the length of the attribute value "detection section" (= end point-start point) Then, the product sum value is the length of the attribute value “detection section” (= end point−start point)
It can be a value divided by the sum of.

[0087]

As described above, according to the voice recognition apparatus of the present invention, the analysis unit is configured to recognize the voice of the recognition unit based on the case frame representing the hypothesis of the utterance content of the voice generated by the generation unit. Let the recognition result be analyzed. And
The generation means is caused to generate a new hypothesis of the speech content of the voice based on the analysis result of the analysis means. Therefore, since the voice is analyzed regardless of the word order, the degree of freedom of speech can be increased. Furthermore, since meaningless words such as unnecessary words contained in the voice are ignored, the voice recognition rate can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of an AV system control device to which a voice recognition device of the present invention is applied.

2 is a more detailed block diagram of the parser 2 of the embodiment of FIG.

FIG. 3 is a diagram showing a configuration example of a word dictionary 12.

FIG. 4 is a diagram showing a configuration example of a word thesaurus 13.

FIG. 5 is a diagram showing a conceptual frame.

FIG. 6 is a diagram showing a conceptual frame.

[Explanation of symbols]

 1 Speech recognition unit 2 Parser 3 Dialog management unit 4 Domain knowledge dictionary 5 Natural language generation unit 6 Text-speech synthesis unit 7 AV device system 11 Speech hypothesis pattern table 12 Word dictionary 13 Word thesaurus 14 Priority queue 15 Result queue

Claims (3)

[Claims] 1. A recognition unit for recognizing a voice, a generation unit for generating a case frame which expresses the hypothesis of the utterance content of the voice and expresses the hypothesis, and a case frame generated by the generation unit. Based on, based on the analysis result of the analysis unit, and the analysis unit for analyzing the recognition result of the recognition unit,
A voice recognition device, which generates a new hypothesis of the utterance content of the voice. 2. The voice recognition device according to claim 1, wherein the case frame expresses the hypothesis generated by the generating means in a semantic relationship centered on a predetermined keyword. 3. The voice recognition device according to claim 1, wherein the case frame is represented by a set of an attribute name and an attribute value.