JPH0772891A - Device for recognizing voice - Google Patents

Abstract

PURPOSE:To provide a voice recognition device capable of correctly recognizing a sentence sound of an inverted expression where a predicative is omitted, and no predicative exists in the end of the sentence. CONSTITUTION:A sound signal is inputted to a sound analysis part 1, and a paragraph lattice is formed through a frame similarity calculation part 3 and a spotting part 5. In a key word estimation part 8, first of all, several paragraph hypotheses (hereafter, called as key word) with a high similarity score are found from the paragraph lattices, and by referring a relative/receptive rule of a construction information storage part 6, the predicative receiving the key word is estimated. In an input sentence estimation part 7, by using the estimated predicate, the key word and the relative/receptive rule of the construction information storage part 6, a paragraph column is retrieved from the paragraph lattices, and is made a recognition result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for automatically recognizing a human voice, particularly a continuous voice such as a sentence voice.

[0002]

2. Description of the Related Art Conventionally, as a method for realizing an apparatus for recognizing continuous speech such as sentence speech, data is actually input from a set data of hypotheses of existence of words or phrases that can be considered in an input speech section (hereinafter referred to as lattice). There is a method of estimating a specified word or phrase using syntactic information.

As a conventional technique, a speech recognition apparatus for estimating an input sentence by using a syntactic information consisting of a connection relation between phrases after obtaining a lattice for the phrase will be described. FIG. 6 shows the configuration of a conventional voice recognition device. In FIG. 6, reference numeral 1 denotes a voice analysis unit that analyzes input voice for each frame that is a unit of analysis and obtains a characteristic parameter.
Reference numeral 2 denotes a standard pattern storage unit that stores a standard pattern that represents a feature of a "basic unit of recognition" (for example, a phoneme, which will be described below using a phoneme). 3 is a frame indicating the similarity between the feature parameter and the standard pattern. A frame similarity calculation unit for each calculation, 4 is a pronunciation information storage unit in which information about pronunciation of a phrase, which is a “linguistic unit”, is represented by a phonetic symbol that represents the basic unit of recognition, and 5 is the presence of a phrase in a voice section. A spotting unit that creates a phrase lattice, which is a set of hypothesis data, 6 is a syntax information storage unit that stores the connection information between the phrases in the sentence to be recognized, and 7 is input from the phrase lattice created by the spotting unit 5. It is an input sentence estimation unit that estimates a sentence.

The operation of the speech recognition apparatus configured as described above will be described below. First, the voice signal is input to the voice analysis unit 1. The voice analysis unit 1 analyzes an input voice and outputs a time series of feature parameters representing the feature of voice for each frame which is a basic unit of analysis. The frame similarity calculation unit 3 obtains the frame similarity obtained by calculating the similarity between the feature parameter time series obtained by the speech analysis unit 1 and the standard pattern prepared for each phoneme in the standard pattern storage unit 2. The degree is output for all voice sections. The spotting unit 5 assumes the existence of conceivable phrases in the entire speech section, and uses the phoneme notation of each phrase stored in the pronunciation information storage unit 4 and the frame similarity calculated by the frame similarity calculation unit 3. The pattern matching can be done freely at the beginning and end of the phrase.

FIG. 7 shows an example of the phrase lattice thus obtained. For each bunsetsu hypothesis in the bunsetsu lattice, its existence interval and similarity score are required. The similarity score is a measure of the possibility of existence of the bunsetsu hypothesis, and is calculated by optimally adding the frame similarities in the bunsetsu hypothesis section. The syntax information storage unit 6 stores syntax information that is connection information between clauses based on a framework such as a context-free grammar, an automaton, and a dependency structure rule. FIG. 8 shows an example of syntax information based on the dependency structure rule. The syntax information in FIG. 8 relates to the verbs "want to drink" and "want to eat" that are the receiving phrases. There are two clauses related to "I want to drink", one is a clause consisting of an independent word in the category of "drink" and the adjunct word "wo", and the other is only an independent word such as "one more cup" or "one cup". It means that the clause consists of. Also, there are two clauses related to "I want to eat", one consisting of an independent word in the category of "food" and an adjunct word of "wo", the other is "another piece", "one", etc. It means that it is a bunsetsu consisting only of independent words.

The input sentence estimation unit 7 searches for a phrase hypothesis forming a sentence from the phrase lattice and uses it as a recognition result.
At this time, the input sentence estimation unit 7 searches for a phrase sequence forming a sentence by using the temporal continuity between the phrase hypotheses and the connection information between the phrases stored in the syntax information storage unit 6 as a constraint condition. Below, the P clause of FIG. 9 will be used for the phrase sequence search process using the dependency syntax rule as a clue to the predicate that is the dependent phrase located at the end of the sentence.
This will be described with reference to the AD diagram. First, a bunsetsu hypothesis that can be a predicate that exists near the end of the speech section is searched from the bunsetsu lattice (step S1). A clause hypothesis that can serve as a found predicate is set as a set of partial sentences (step S2). The following processing is performed until the set of partial sentences becomes empty (step S4 to
S12) is repeated (step S3). A bunsetsu hypothesis (hereinafter referred to as a pre-adjacency hypothesis) temporally adjacent to each sub-sentence is searched from the bunsetsu lattice (step S5),
It is checked whether or not there is a bunsetsu that can receive those pre-adjacent hypotheses in a dependency relationship (step S).
7). Specifically, the dependency information structure rule regarding each clause forming the partial sentence is inquired to the syntax information storage unit 6.
If a subsentence has a clause that can receive the preceding adjacency hypothesis in a dependency relationship, the preceding adjacency hypothesis connected to that subsentence is added to the set of new subsentences (step S
8). For each sub-sentence of the set of new sub-sentences created in this way, it is checked whether the sub-sentence covers the voice section from the front to the back (step S10). The covering partial sentence is removed from the set of new partial sentences and added to the set of sentence candidates (step S1).
1). Finally, the set of new sub-sentences is the set of sub-sentences. Here, if the set of sub-sentences is not empty, the processes of steps S4 to S12 are repeated. As a result of the above processing, a set of sentence candidates is obtained. In most cases, multiple sentence candidates are found. In order to select one of the sentence candidates as the recognition result, an evaluation function that can be calculated from the similarity score of the bunsetsu hypothesis forming the sentence candidate is prepared, and the one having the highest value is set as the recognition result. Further, in order to reduce the amount of sentence candidate search processing, in step S12 using the above evaluation function, only new sub-sentences having a high evaluation function value are targeted for the next processing (step S4). Leave as. This is a search method generally called a beam search (branching method), and the remaining number is called a beam width.

[0007]

However, the above-mentioned conventional speech recognition apparatus does not operate correctly when the predicate at the end of a sentence is not found in the phrase lattice as a phrase hypothesis with a high similarity score due to the fact that it is not uttered clearly. It had a problem that it could not be recognized. Moreover, since the predicate at the end of the sentence is used as a direct clue for the phrase lattice search, there is a problem that it is not possible to recognize a sentence in which the predicate at the end of the sentence is omitted or an inverted expression sentence in which there is no predicate at the end of the sentence. Furthermore, in the conventional speech recognition system that connects the pre-adjacent bunsetsu hypotheses that satisfy the dependency relation rule from the end of the sentence to the beginning of the sentence, the search process will not continue unless the pre-adjacent bunsetsu hypotheses that satisfy the dependency relation rule are found. There was a problem that results could not be obtained.

The present invention is to solve the above-mentioned conventional problems. As a predicate at the end of a sentence, a sentence voice that is not required in the phrase lattice as a phrase hypothesis having a high similarity score or a predicate at the end of a sentence is omitted. An object of the present invention is to provide a speech recognition apparatus that can correctly recognize a sentence speech that does not have an inverted expression and that can continue search processing even when a pre-adjacent bunsetsu hypothesis that satisfies the dependency rule is not found.

[0009]

In order to achieve the above object, a speech recognition apparatus of the present invention comprises a speech analysis section for analyzing input speech for each frame which is a unit of analysis and obtaining a characteristic parameter, a phoneme and a syllable. And the like, a standard pattern storage unit that stores a standard pattern that represents a feature of a “basic unit of recognition”, a frame similarity calculation unit that calculates the similarity between the feature parameter and the standard pattern for each frame, a word or A pronunciation information storage unit in which information about pronunciation of "linguistic units" such as phrases is represented by phonetic symbols that represent the basic units of recognition,
Between all the language units stored in the pronunciation dictionary, between the spotting unit that calculates the similarity in an arbitrary section of the input speech to create a lattice and the language units such as words and clauses in the sentence to be recognized. The sentence information input unit that estimates the sentence input by voice from the lattice information storage unit that stores the combined information and the lattice created by the spotting unit, and uses this as the recognition result, and the input sentence by the input estimation unit. The keyword estimation unit estimates a keyword to be used as a clue from the lattice created by the spotting unit.

[0010]

With this configuration, a bunsetsu hypothesis with a high similarity score in a bunsetsu lattice is used as a keyword, and this keyword and a predicate estimated from this keyword are used as clues for a bunsetsu lattice search, and a predicate at the end of a sentence has a high similarity score. As a hypothesis, it is possible to recognize sentence speech which is not required in the phrase lattice and sentence predicates at the end of the sentence are omitted, and sentence speech of an inverted expression where there is no predicate at the end of the sentence. Further, even if a part of the bunsetsu in the input sentence is not caught as the bunsetsu lattice's bunsetsu hypothesis, the search process is advanced and the bunsetsu string forming a part of the input sentence can be obtained.

[0011]

【Example】

(First Embodiment) A first embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a voice recognition device in this embodiment. In the present embodiment, the same components as those of the conventional speech recognition apparatus shown in FIG. 6 are designated by the same reference numerals and detailed description thereof will be omitted.

In FIG. 1, reference numeral 1 denotes a speech analysis unit that analyzes input speech for each frame that is a unit of analysis and obtains a characteristic parameter. Reference numeral 2 denotes a "basic unit of recognition" (for example, a phoneme. Hereinafter, description will be made using phonemes. ) A standard pattern storage unit for storing a standard pattern representing a feature, 3 is a frame similarity calculation unit for calculating the similarity between the feature parameter and the standard pattern for each frame, and 4 is a clause of a "language unit". A pronunciation information storage unit in which information about pronunciation is represented by phonetic symbols that represent the basic unit of recognition, 5 is a spotting unit that creates a phrase lattice that is a set of hypothesis existence data of a phrase in a voice section, and 6 is a recognition target. A syntax information storage unit that stores connection information between clauses in a sentence, and 7 is an input that estimates a sentence that is input from the phrase lattice created by the spotting unit 5. Estimator, 8 keywords that clue in estimating the input sentence in the input estimating section 7 is a keyword estimation unit that estimates the lattice created by the spotting unit.

The operation of the speech recognition apparatus configured as above will be described. First, the speech signal is input to the speech analysis unit 1 and the same as in the conventional example up to the point where the phrase lattice is created through the frame similarity calculation unit and the spotting unit. The keyword estimation unit 8 first finds some bunsetsu hypotheses (hereinafter referred to as keywords) having a high similarity score from the bunsetsu lattice. Next, by referring to the dependency relation rule of the syntax information storage unit 6, a predicate that receives the keyword is estimated, and it is checked whether or not the predicate exists near the end of the speech section of the phrase lattice. When the estimated predicate exists in the phrase lattice, the input sentence estimation unit 7 searches the phrase string so as to connect the predicate and the preceding keyword as a clue. At this time, the dependency relation rule of the syntax information storage unit 6 is used.

FIG. 2 shows an example of the above operation. The example of FIG. 2 is a schematic diagram of the phrase lattice obtained for the utterance "I want to drink another beer", and the similarity score of the phrase hypothesis "I want to drink" that should be correct is low. It is assumed that the search is performed by a beam search with a beam width of 1. First, the phrase hypothesis "beer" having the highest similarity score in the phrase lattice is found as a keyword. Next, in the dependency relation rule of the syntax information storage unit 6, a predicate that can have “beer” as a dependency clause is searched for. Using the dependency relation rule of FIG. 8, "I want to drink" is estimated as a predicate. It is present if the hypothesis of this predicate exists near the end of the speech section of the bunsetsu lattice. It exists, so using the predicate and the preceding keyword as a clue, while connecting the preceding bunsetsu hypotheses using the dependency relation rules in FIG. If you search for the phrase series, you'll see
The phrase string "I want to drink" can be correctly searched and is output as a recognition result.

On the other hand, when the search is performed by the beam search with the beam width of 1 in the conventional speech recognition apparatus, "I want to eat" having a high similarity score is searched in the search of the predicate at the end of the sentence (step S1 in FIG. 9). Then, if the dependency relation rule of FIG. 8 is used thereafter, the phrase sequence “I want to eat pizza, another one, I want to eat” will be searched.

As described above, according to this embodiment, by providing a keyword estimation unit that estimates a keyword other than a predicate in the bunsetsu lattice and estimates the predicate from the keyword, the predicate at the end of the sentence is not clearly pronounced. It can be correctly recognized even when the phrase hypothesis with a high similarity score cannot be obtained during the phrase lattice.

(Second Embodiment) A second embodiment of the present invention will be described below with reference to the drawings. Since the configuration of the voice recognition apparatus of this embodiment is the same as that of the first embodiment, its detailed description is omitted and its operation will be described.

First, the voice signal is input to the voice analysis unit 1, the phrase lattice is created through the frame similarity calculation unit and the spotting unit, and the keyword estimation unit 8 estimates the keyword and the predicate that receives the keyword. This is the same as the first embodiment. The keyword estimation unit 8 checks whether the estimated predicate exists near the end of the speech section of the phrase lattice. When the estimated predicate exists in the phrase lattice, the input sentence estimation unit 7 searches the phrase string so as to connect the predicate and the preceding keyword as a clue. At this time, the dependency relation rule of the syntax information storage unit 6 is used. Here, there is a case where a clause hypothesis that satisfies the dependency relation rule cannot be found in the section between the keyword and the predicate. In this case, a process of searching for a clause hypothesis that satisfies the dependency relation rule in the voice section before the keyword is performed. In the worst case, the phrase sequence consisting of keywords and predicates is used as the recognition result.

FIG. 3 shows an example of the above operation. The example of FIG. 3 is a schematic diagram of the bunsetsu lattice obtained for the utterance "I want to drink another beer", and the bunsetsu hypothesis "another cup" that should be the correct answer does not exist in the bunsetsu lattice. It is assumed that the search is performed by a beam search with a beam width of 1. First, the phrase hypothesis "beer" having the highest similarity score in the phrase lattice is found as a keyword. Next, in the dependency relation rule of the syntax information storage unit 6, a predicate that can have “beer” as a dependency clause is searched for. Using the dependency relation rule of FIG. 8, "I want to drink" is estimated as a predicate. It is present if the hypothesis of this predicate exists near the end of the speech section of the bunsetsu lattice, so it exists, and while connecting the pre-adjacent bunsetsu hypothesis using the dependency relation rule of FIG. Search for a phrase sequence. However, the predicate "I want to drink"
Since there is no pre-adjacent bunsetsu hypothesis satisfying the dependency relation rule, the bunsetsu hypothesis is searched for a bunsetsu hypothesis that satisfies the dependency relation rule in the speech section before the keyword.
In the example of FIG. 3, since the keyword is at the beginning of the voice section, the search process ends here, and as a result, the phrase string “beer / want to drink” is obtained. Although this is not the correct phrase sequence, it can catch a part of the input utterance, and when the input is a sentence, a result close in meaning is often obtained.

On the other hand, when the search is performed by the beam search with the beam width of 1 in the conventional speech recognition apparatus, "I want to eat" having a high similarity score is searched in the search of the predicate at the end of the sentence (step S1 in FIG. 9). Then, if the dependency relation rule of FIG. 8 is used thereafter, the phrase sequence “I want to eat pizza, another one, I want to eat” will be searched. Even if the similarity score of the bunsetsu hypothesis “I want to drink” is higher than that of “I want to eat”, in the conventional speech recognition device that connects the preceding bunsetsu hypotheses that satisfy the dependency relation rule from the end of the sentence to the beginning of the sentence, If no pre-adjacent bunsetsu hypothesis satisfying the receiving relation rule is found, the search process does not continue halfway and no result is obtained.

As described above, the keyword estimation unit for estimating a keyword other than the predicate in the phrase lattice and estimating the predicate from the keyword and the estimation process of the input sentence are performed even when there is no proper pre-adjacent hypothesis in the lattice. By providing an input sentence estimation unit that can continue, the search process can proceed even if some bunsetsus in the input sentence are not caught as the bunsetsu lattice's bunsetsu hypothesis, and a bunsetsu string forming a part of the input sentence can be obtained. it can.

(Embodiment 3) A third embodiment of the present invention will be described below with reference to the drawings. Since the configuration of the voice recognition apparatus of this embodiment is the same as that of the first embodiment, its detailed description will be omitted and its operation will be described.

First, the voice signal is input to the voice analysis unit 1, the phrase lattice is created through the frame similarity calculation unit and the spotting unit, and the keyword estimation unit 8 estimates the keyword and the predicate that receives the keyword. This is the same as the first embodiment. The keyword estimation unit 8 checks whether the estimated predicate exists near the end of the speech section of the phrase lattice. If the estimated predicate does not exist near the end of the speech segment in the phrase lattice,
It is considered that the predicate is omitted, and the input sentence estimation unit 7 assumes that the predicate exists behind the voice section, and searches the phrase string so as to connect the predicate and the preceding keyword as a clue. At this time, the dependency relation rule of the syntax information storage unit 6 is used.

FIG. 4 shows an example of the above operation. In the example of Fig. 4, "one more beer" with the predicate omitted
FIG. 3 is a schematic diagram of a bunsetsu lattice obtained for the utterance of FIG.
First, the phrase hypothesis "beer" having the highest similarity score in the phrase lattice is found as a keyword. Next, in the dependency relation rule of the syntax information storage unit 6, a predicate that can have “beer” as a dependency clause is searched for. Using the dependency relation rule of FIG. 8, "I want to drink" is estimated as a predicate. Since it does not exist when checking whether the hypothesis of this predicate exists near the end of the speech section of the phrase lattice, it is assumed that the predicate "I want to drink" is omitted,
Using the predicate and the preceding keyword as a clue, the bunsetsu string is searched while connecting the pre-adjacent bunsetsu hypotheses using the dependency relation rules of FIG. As a result, "one more beer"
Can be correctly searched and is output as a recognition result.

As described above, the keyword estimation unit that estimates the keywords other than the predicate in the voice section and determines whether the predicate is omitted, and the input sentence estimation unit that can estimate the input sentence in which the predicate is omitted. By providing
Can be recognized correctly.

(Embodiment 4) A fourth embodiment of the present invention will be described below with reference to the drawings. Since the configuration of the voice recognition apparatus of this embodiment is the same as that of the first embodiment, its detailed description will be omitted and its operation will be described.

First, the voice signal is input to the voice analysis unit 1, the phrase lattice is created through the frame similarity calculation unit and the spotting unit, and the keyword estimation unit 8 estimates a keyword and a predicate to receive the keyword. This is the same as the first embodiment. The keyword estimation unit 8 checks whether the estimated predicate exists near the end of the speech section of the phrase lattice. If the estimated predicate does not exist near the end of the speech segment in the phrase lattice,
It is considered that the predicate is omitted, and the input sentence estimation unit 7 assumes that the predicate exists behind the voice section, and searches the phrase string so as to connect the predicate and the preceding keyword as a clue. At this time, the dependency relation rule of the syntax information storage unit 6 is used. If a hypothesis of an estimated predicate is found as a pre-adjacent clause hypothesis during the search, it is interpreted that the input is an inverted expression, and the predicate hypothesis is connected to continue the search.

FIG. 5 shows an example of the above operation. The example of FIG. 5 is a schematic diagram of the phrase lattice obtained for the utterance of "another drink of beer" which is an inverted expression. First, the phrase hypothesis "beer" having the highest similarity score in the phrase lattice is found as a keyword. Next, in the dependency relation rule of the syntax information storage unit 6, a predicate that can have “beer” as a dependency clause is searched for. Using the dependency relation rule of FIG. 8, "I want to drink" is estimated as a predicate. If the hypothesis of this predicate does not exist when checking whether it exists near the end of the speech section of the bunsetsu lattice, it does not exist, so it is considered that the predicate "I want to drink" is omitted, and the predicate and the preceding keyword are used as clues to the relationship in FIG. The bunsetsu sequence is searched while connecting the preceding bunsetsu hypotheses using the receiving relation rule. First, the bunsetsu hypothesis "One more cup"
Was found, and "I want to drink" as the bunsetsu hypothesis in front of it
Can be found. Since this is an estimated predicate, the input is interpreted as an inverted expression, and the predicate hypothesis is connected to continue the search. In FIG. 5, the bunsetsu hypothesis "I want to drink" connects the bunsetsu hypothesis "another cup" and "beer", and the search process ends. As a result, the phrase sequence "beer, want to drink, another cup" can be correctly searched and output as a recognition result.

As described above, the keyword estimation unit for estimating the keywords other than the predicate in the voice section and determining whether the predicate is omitted, and the input sentence estimation unit for estimating the input sentence of the inverted expression are provided. Therefore, it can be correctly recognized.

In addition, although the example using the phrase lattice is described in the above embodiment, the present invention is naturally applied to the case using the word lattice.

[0031]

As described above, according to the present invention, the characteristics of the "basic unit of recognition" such as a phoneme and a syllable are analyzed by analyzing the input speech for each frame which is a unit of analysis and obtaining a characteristic parameter. A standard pattern storage unit that stores a standard pattern to represent, a frame similarity calculation unit that calculates the similarity between the characteristic parameter and the standard pattern for each frame, and information about pronunciation in "language units" such as words and phrases. To create a lattice by calculating the degree of similarity in an arbitrary section of the input voice with respect to all pronunciation units stored in the pronunciation dictionary and the pronunciation information storage unit in which the above is represented by a phonetic symbol representing the basic unit of recognition. The spotting section, the syntax information storage section that stores the connection information between linguistic units such as words and clauses in the sentence to be recognized, and the latte created by the spotting section. An input sentence estimation unit that estimates a sentence that is an input voice from a keyword and a keyword estimation unit that estimates a keyword to be used as a clue when estimating the input sentence by the input estimation unit from the lattice created by the spotting unit are provided. As a result, it is possible to recognize sentence speech that is not required in the clause lattice as a clause hypothesis with a high similarity score for the predicate at the end of the sentence, sentence speech in which the predicate has been omitted, and sentence speech with an inverted expression. The voice recognition device can be realized.

[Brief description of drawings]

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

FIG. 2 is a conceptual diagram of a lattice search operation of the voice recognition device in the first embodiment.

FIG. 3 is a conceptual diagram of a lattice search operation of the voice recognition device in the second embodiment.

FIG. 4 is a conceptual diagram of a lattice search operation of the voice recognition device in the third embodiment.

FIG. 5 is a conceptual diagram of a lattice search operation of the voice recognition device in the fourth embodiment.

FIG. 6 is a configuration diagram of a voice recognition device according to a conventional technique.

FIG. 7 is a diagram showing a part of a phrase lattice obtained as a result of spotting.

FIG. 8 is a diagram showing description contents of syntax information.

FIG. 9 is a PAD diagram explaining a conventional phrase sequence search process.

[Explanation of symbols]

 1 voice analysis unit 2 standard pattern storage unit 3 frame similarity calculation unit 4 pronunciation information storage unit 5 spotting unit 6 syntax information storage unit 7 input sentence estimation unit 8 keyword estimation unit

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Mitsuru Endo 3-10-1 Higashisanda, Tama-ku, Kawasaki-shi, Kanagawa Matsushita Giken Co., Ltd.

Claims (9)

[Claims] 1. A voice analysis unit for analyzing input speech for each frame as a unit of analysis to obtain a characteristic parameter, and a standard pattern for storing a standard pattern representing a feature of a phoneme or a syllable as a "basic unit of recognition". A storage unit, a frame similarity calculation unit that calculates the similarity between the characteristic parameter and the standard pattern for each frame, and information regarding the pronunciation of a word or phrase that is a "linguistic unit" represents the basic unit of recognition. A pronunciation information storage unit represented by phonetic symbols, a spotting unit that creates a lattice by calculating the similarity in an arbitrary section of the input speech with respect to all the language units stored in the pronunciation information storage unit, and a recognition target From the lattice information created by the spotting section and the syntax information storage section that stores the connection information between the language units in the sentence An input sentence estimation unit that estimates a sentence represented by a voice as a recognition result, and a keyword estimation unit that estimates a keyword to be a clue when estimating the input sentence in the input estimation unit from the lattice created in the spotting unit. A voice recognition device provided. 2. The voice recognition device according to claim 1, wherein the keyword estimation unit estimates a keyword other than the predicate in the voice section and estimates the predicate from the keyword. 3. A keyword estimation unit for estimating a keyword other than a predicate in a voice section to determine whether or not the predicate is omitted, and an input sentence estimation unit for estimating an input sentence in which the predicate is omitted. The voice recognition device according to claim 2. 4. A keyword estimating unit for estimating a keyword other than a predicate in a voice section to determine whether or not the predicate is omitted, and an input sentence estimating unit for estimating an input sentence of an inverted expression. The voice recognition device according to claim 2. 5. The speech recognition according to claim 2, further comprising an input sentence estimation unit that continues the input sentence estimation process even when there is no suitable preceding neighbor hypothesis in the lattice. apparatus. 6. A keyword estimation unit estimates a keyword other than a predicate in a voice section and determines whether or not the predicate is omitted, and an input sentence estimation unit determines an input sentence in which the predicate is omitted and an input sentence of an inverted expression. The estimation is performed according to claim 2.
The voice recognition device described. 7. A keyword estimation unit for estimating a keyword other than a predicate in a speech section to determine whether or not the predicate is omitted, and an input sentence in which the predicate is omitted is estimated, and an appropriate pre-adjacent word is added in the lattice. The speech recognition apparatus according to claim 2, further comprising an input sentence estimation unit that continues the input sentence estimation process even when no hypothesis exists. 8. A keyword estimation unit for estimating a keyword other than a predicate in a voice section to determine whether or not the predicate is omitted, and an input sentence of an inverted expression are estimated, so that an appropriate pre-adjacent hypothesis is generated in the lattice. The speech recognition apparatus according to claim 2, further comprising an input sentence estimation unit that continues the estimation processing of the input sentence even when it does not exist. 9. A keyword estimation unit for estimating a keyword other than a predicate in a voice section and determining whether or not the predicate is omitted, and an input sentence in which the predicate is omitted and an input sentence of an inverted expression are suitable in the estimation lattice. The speech recognition apparatus according to claim 2, further comprising an input sentence estimation unit that continues the estimation process of the input sentence even when there is no such pre-adjacent hypothesis.