JP2015099253A - Voice recognizing apparatus, voice recognizing method, voice recognizing program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognizing apparatus capable of correcting a piece of voice recognition even when a user suspends a pronunciation.SOLUTION: The voice recognizing apparatus includes: a recognition output section; a pronunciation relation determination section; and a recognition connecting section. The recognition output section outputs a piece of recognition of an input voice pronunciation. The pronunciation relation determination section determines whether a preceding input first voice pronunciation is continuous to a following input second voice pronunciation. When the second voice pronunciation is determined as the continuous pronunciation, the recognition connecting section connects the first recognition of first voice pronunciation and the second recognition of the second voice pronunciation to each other.

Description

  The embodiment described in this specification relates to a speech recognition technique for recognizing a character string input by speech.

  2. Description of the Related Art Conventionally, a speech recognition technology has been developed that converts speech information into character information by pattern matching between an input speech utterance and speech analysis information stored in advance. Since current speech recognition technology cannot completely eliminate the occurrence of misrecognition, various technologies for obtaining appropriate recognition results have been proposed and widely used.

  As voice recognition technology, for example, the most probable candidate from a plurality of recognition candidates is presented to the user as a recognition result, and when the presented recognition result is incorrect, the user re-enters the entire utterance, A technique for efficiently acquiring a correct recognition result by excluding the original recognition result is known. However, in this method, since the entire utterance is re-input, there is a problem that the operation burden on the user increases.

  On the other hand, in Patent Document 1, when the recognition result to be presented is incorrect, if the user re-speaks only the misrecognized part of the device, the misrecognized part is corrected and recognized based on the re-spoken part. Techniques for re-presenting results are disclosed. In Patent Document 1, this can reduce the burden of rephrasing the whole sentence and improve convenience.

  Further, in Patent Document 1, on the premise that information having a hierarchical structure such as an address and a telephone number is input, it is determined to which hierarchy the re-input utterance corresponds and the recognition result is corrected. At this time, in Patent Document 1, only a pattern under a layer that is erroneously recognized can be used as a target of collation, so that the efficiency and accuracy of the recognition process can be improved.

  In Patent Document 2, the user re-speaks only the misrecognized part, and deletes the recognition candidate selected at the time of the original utterance from the recognized part of the recognition candidate, thereby selecting and presenting the wrong candidate again. A technique for avoiding this problem has been proposed.

  However, in Patent Documents 1 and 2, as a general rule, an appropriate recognition result can be obtained only by either speaking and correcting the whole or correcting a part of it again. There is a problem that correction according to the content of recognition cannot be performed.

  Moreover, in patent document 2, although both a total correction and a partial correction can be performed, since the correction method is the same in each correction, the user's whether the user reuttered the whole or only the part was spoken. Cannot make corrections according to the utterance method.

  With respect to these problems, the apparatus of Patent Literature 3 first presents the most probable candidate from a plurality of recognition candidates to the user as a recognition result based on the user's utterance. When the recognition result presented by the device is incorrect, the user causes the device to present the recognition result again by speaking again. When the recognition result presented by the apparatus is correct, the user ends the voice recognition process of the apparatus by inputting an OK button or the like.

  When the device corrects the recognition result due to the user's recurrent utterance, the device compares the first utterance immediately before utterance with the current utterance after the immediately preceding utterance to correct, and similar and inconsistent portions. Is calculated. Then, the apparatus determines whether the current utterance is any one of “partially corrected utterance”, “overall corrected utterance”, and “new utterance”.

  “Partial correction utterance” is an utterance when the user re-utters only the misrecognized portion of the apparatus. The “whole correction utterance” is an utterance when the user wants to utter almost the entire content that the user wants to utter and the content differs only in the last utterance. The “new utterance” is an utterance in which almost the entire content that the user wants to utter is re-uttered and most of the content is different from the previous utterance.

  In this way, the device of Patent Document 3 determines whether the user's recurrent utterance is “partial corrected utterance”, “overall corrected utterance”, or “new utterance”, and corrects according to the type of the utterance The recognition result is corrected by the method, and the corrected recognition result is presented to the user.

  Thereby, in Patent Document 3, when the erroneous recognition of the device is slight in the first utterance, the user utters the whole again by speaking only the erroneously recognized part or correcting the erroneously recognized part. Thus, the erroneously recognized portion can be corrected by the correction method corresponding to “partial correction utterance” or “total correction utterance”. Also, when misrecognition occurs in a concentrated manner such as a burst error in the first utterance, the user can cause the apparatus to recognize the voice as “new utterance” by repeating the entire utterance again.

  As described above, since the apparatus disclosed in Patent Document 3 can perform correction according to the user's utterance method, it is highly convenient and can output a highly accurate voice recognition result.

JP 2002-287792 A JP 2003-316386 A JP 2007-93789 A

  By the way, the reason why erroneous recognition occurs in the apparatus is caused not only by noise in the surrounding environment and incomplete recognition method, but also when the utterance itself is incomplete. For example, if the user bites, the user may not be able to convey the intended content and may interrupt the utterance. In such a case, the device cannot recognize the entire user's utterance and cannot recognize the error. Recognition occurs.

  When the user interrupts the utterance, in order to correct the misrecognition of the device, the user needs to perform the re-utterance and cause the device to correct the recognition result for the previous utterance by the re-utterance (current utterance). is there. However, the conventional apparatus is inconvenient because it is necessary for the user to speak the entirety including the interrupted portion and to cause the apparatus to perform a correction corresponding to the “totally modified utterance” or “new utterance”.

  An object of the present invention is to provide a technique that is convenient for correcting speech recognition when a user interrupts speech.

  In general, according to the embodiment, the speech recognition apparatus includes a recognition output unit, an utterance relationship determination unit, and a recognition connection unit. The recognition output unit outputs recognition of the input voice utterance. The utterance relationship determination unit determines whether or not the second voice utterance input later is a continuous utterance continuing from the first voice utterance input earlier. When it is determined that the second voice utterance is a continuous utterance, the recognition connection unit connects the first recognition, which is the recognition of the first voice utterance, and the second recognition, which is the recognition of the second voice utterance.

  In general, according to the embodiment, the speech recognition method recognizes an input speech utterance, and the second speech utterance input later is an utterance from the middle of the first speech utterance input earlier. If it is determined whether the second voice utterance is a continuous utterance, the recognition of the first voice utterance and the recognition of the second voice utterance are coupled.

  In general, according to the embodiment, an input voice utterance is recognized, and it is determined whether or not a second voice utterance input later is a continuous utterance that is an utterance from the middle of the first voice utterance input first. When determining that the second voice utterance is a continuous utterance, the voice recognition apparatus is caused to cause the recognition result of the first voice utterance to be connected to the recognition result of the second voice utterance. A speech recognition program to be executed by a computer.

It is a block diagram which shows the structure of a speech recognition apparatus. It is a flowchart of a voice recognition process. It is a figure which shows the data structure of the recognition candidate which a recognition candidate production | generation part outputs. It is a figure which shows an example of the data recognized by voice recognition processing. It is a flowchart of the utterance relation determination method of the last utterance and the present utterance. It is a figure which shows an example of the data recognized by voice recognition processing. It is a figure which shows the example of a user's utterance. It is a figure which shows the phoneme of the last speech and the present speech. It is a figure which shows the example of a user's utterance.

Hereinafter, embodiments will be described with reference to the drawings.
FIG. 1 is a block diagram showing the configuration of the speech recognition apparatus.
The speech recognition apparatus 100 includes a microphone 111, a display 112, a speaker 113, and an utterance information storage unit 121 as a hardware configuration.

  The speech recognition apparatus 100 includes, as a software configuration, an input reception unit 101, a recognition candidate generation unit 102 (recognition output unit), an utterance relationship determination unit 103, a partial correction unit 104a, an overall correction unit 104b, and a recognition candidate selection. Unit 105, output control unit 106, and recognition coupling unit 107.

Hereinafter, each element will be described with reference to the flowchart of FIG.
The microphone 111 receives a user's utterance.
The input receiving unit 101 performs A / D (Analogue to Digital) conversion on the user's utterance received by the microphone 111 (Act 1).

The recognition candidate generation unit 102 performs speech recognition processing on the user's utterance to generate a recognition candidate (Act 2).
The recognition candidate generation unit 102 performs frequency analysis by FFT (Fast Fourier Transform) and the like, for each predetermined section of input speech, time-series characteristic information (for example, spectrum) necessary for speech recognition for each section. Output to. The predetermined section refers to a unit such as a character string such as a word, a phoneme string such as an accent phrase, a phoneme, or a syllable.

  The recognition candidate generation unit 102 compares the feature information with a speech analysis pattern stored in a speech recognition dictionary (not shown) to calculate a score representing similarity, and outputs a recognition candidate in a lattice format.

FIG. 3 is a diagram illustrating a data structure of recognition candidates output from the recognition candidate generation unit 102.
When “WA-TA-SHI-WA-KO-UEN-NI-I-KI-TA-I”, which means “I want to go to the park”, is input as a new utterance, The lattice shown in FIG. 3 is output as a recognition candidate. In the upper part of FIG. 3, each phoneme included in the input speech is associated with a numerical value indicating the input position of each phoneme.

  Lattice is generated by connecting nodes with recognition results arranged in chronological order from left to right with arcs with scores that indicate the degree of similarity when collated with speech analysis patterns stored in a speech recognition dictionary (not shown). Is done. A score of the node is attached to an arc input to the node (hereinafter referred to as an input arc). The score indicates the likelihood (likelihood) of the recognition candidate.

  A node is generated for each predetermined section of input speech such as a character string such as a word, a phoneme string such as an accent phrase, a phoneme, and a syllable. For example, the node 201 in FIG. 3 represents a Japanese word “I” composed of three phonemes “WA”, “TA”, and “SHI” uttered in a section from 0 to 2 in the input position. The symbol “/” represents the part of speech of the word. “/ N” in the node 201 indicates that the word is a noun.

  The recognition candidate generation unit 102 generates a plurality of corresponding nodes when there are a plurality of recognition candidates. In the example of FIG. 3, the recognition candidate generation unit 102 generates three recognition candidates, each having a node with an input position interval of 4 to 7 and a node with an input position interval of 9 to 10. The score of each node is 5, 3, 2 and 6, 2, 2.

  The recognition candidates generated by the recognition candidate generation unit 102 are stored in the utterance information storage unit 121. The utterance information storage unit 121 also stores a recognition result generated by a recognition candidate selection unit 105 described later and a recognition candidate generated by a recognition connection unit 107 described later.

  When the user's utterance “I want to go to the park” is a new utterance, the utterance relation determination unit 103 determines that the user's utterance is a new utterance (Act 3 in FIG. 2: Act 3). New utterance).

In this case, the recognition candidate selection unit 105 selects the nodes “performance” and “living” with the highest score in the recognition candidates shown in FIG. "I want to live in a performance" is output as a recognition result (Act 4).
The output control unit 106 displays a recognition result such as “I want to live in a performance” output from the recognition candidate selection unit 105 on the display 112 (Act 5).

The apparatus 100 displays the recognition result of the new utterance on the display 112 in this way. If the recognition result includes an error, the user corrects the error by the utterance.
As a method for correcting the recognition result of the apparatus 100, conventionally, there are a correction method using a partially corrected utterance, an overall corrected utterance, and a new utterance.

The correction method based on the partial correction utterance is a method in which the apparatus 100 corrects only the erroneous portion by re-speaking only the erroneous portion of the immediately preceding utterance.
The correction method based on the overall correction utterance is a method in which only a part is changed and the entire utterance is again uttered, and the apparatus 100 is made to integrate the lattice of the immediately preceding utterance and the current utterance and output the recognition result again.

  The correction method based on a new utterance is a method in which correction is performed by causing the apparatus 100 to recognize a voice separately from the immediately preceding utterance by performing an utterance with many inconsistent portions with the immediately preceding utterance. For example, when the user bites in the middle and interrupts the utterance, the entire sentence is uttered again from the beginning, thereby causing the apparatus 100 to recognize the voice as a new utterance.

  In the present embodiment, one of the features is that in addition to the conventional three utterance correction methods, there is a correction method using continuous utterances.

  Briefly speaking, the correction method based on continuous utterance means that if the user bites in the middle and interrupts the utterance, the utterance was interrupted, for example, "We are good corporate citizens and have high ethical standards." In this case, it is a method of causing the apparatus 100 to recognize the voice by connecting the current utterance and the immediately preceding utterance by uttering “the ethics and the spirit of the law, each country and...

  Since the correction method by partial correction utterance and the correction method by whole correction utterance are disclosed in Patent Document 3 (Japanese Patent Laid-Open No. 2007-93789), they will be briefly described below. After describing these conventional correction methods, the correction method based on continuous speech will be specifically described.

<About correction method using partial correction utterance>
First, a method for correcting misrecognition of the apparatus 100 by partial correction utterance will be described with reference to the flowcharts of FIGS. 4 and 5.

  For the new utterance “WA-TA-SHI-WA-KO-UEN-NI-I-KI-TA-I” meaning “I want to go to the park”, as described above, the device 100 is shown in FIG. Consider the case where “I want to live in a performance” is output as a recognition result, as shown in a).

  In this case, the user can correct the erroneous recognition of the device 100 by re-inputting only the utterance “KO-UEN” meaning “park” in order to cause the device 100 to correct the recognition result error. Yes (Act1 in FIG. 2).

  As shown in FIG. 4B, the recognition candidate generation unit 102 recognizes four recognition candidates “performance”, “park”, and “sponsor” as recognition candidates for the current utterance “KO-UEN” made after the previous utterance. , “Good performance” is output (Act 2). Assume that the score of each recognition candidate is 7, 7, 5, 1.

  The utterance relationship determination unit 103 compares the immediately preceding utterance and the current utterance, and calculates a similar part and a mismatched part (Act 31 in FIG. 5). For example, the method disclosed in Patent Document 2 can be used to calculate the similar portion. That is, the utterance relation determination unit 103 determines the ratio of phoneme strings or character strings that are common to both of a plurality of phoneme strings or character strings that are recognition candidates generated for each predetermined section of the input speech of the immediately preceding utterance and the current utterance. Is equal to or greater than a predetermined threshold, the predetermined section is determined as a similar portion. A section that is not determined to be a similar part is a mismatched part.

  The utterance relationship determination unit 103 determines whether a part of the immediately preceding utterance and all parts of the current utterance are similar (Act 32).

In this example, all parts of the current utterance "KO-UEN" are part of the previous utterance "WA-TA-SHI-WA-KO-UEN-NI-I-KI-TA-I""KO-UEN" (Act 32: YES), the utterance relationship determination unit 103 determines that the current utterance is a partially modified utterance.
The partial correction unit 104a performs partial correction processing (Act 6 in FIG. 2). That is, as shown in FIG. 4B, the partial correction unit 104a sets the score of the word “performance” corresponding to the current utterance in the last utterance to 1 to 1, which is the lowest score among the four recognition candidates. Set to 0 which is the subtracted value.

As a result, the recognition candidate selection unit 105 selects a node “park” having a score of 7 and having the maximum score, and outputs “I want to live in a park” that connects the selected nodes as a recognition candidate (FIG. Act 4 of 2).
The output control unit 106 displays the recognition result “I want to live in the park” on the display 112 (Act 5).

  Then, the user next speaks only “I-KI-TA-I” which means “I want to go” in the same way as above, and the recognition result “I live in the park” of the device 100. The misrecognized part “I want to live” in “I want” can be corrected to “I want to go”.

  When the device 100 displays the recognition result “I want to go to the park” and the user determines that the recognition result is correct and clicks the OK button, for example, the device 100 (the output control unit 106) causes the speaker 113 to At “I want to go to the park” and output the voice, and the voice recognition process is terminated.

<About the correction method by the whole correction utterance>
The correction method based on the overall correction utterance causes the apparatus 100 to correct erroneous recognition by changing only a part and then speaking the whole again.

FIG. 6 is a diagram for explaining the overall correction process. In FIG. 6, nodes in the section where the input position is from 8 to 15 are omitted.
For example, "EI-GA-WO-MI-TA-I-NO-DE-SU-GA-DO-KO-EI-" means that the user wants to see a movie, but where should I go? When “KE-BA-YO-I-DE-SU-KA” is spoken, the device 100 generates the lattice shown in FIG. 6 (a) and connects the nodes with the highest score. The user can use the correction method based on the entire correction utterance when outputting “Where should I go?” As a recognition result.

  In this case, the user changes the ending "DE-SU-KA" to "I-KA" and the entire modified utterance "EI-GA-WO-MI-TA-I-NO-DE-SU-GA-DO-KO" -EI-KE-BA-YO-II-KA "can be entered again (Act 1).

The recognition candidate generator 102 generates the lattice shown in FIG. 6B (Act 2).
The utterance relationship determination unit 103 compares the immediately preceding utterance with the current utterance that is the overall corrected utterance, and calculates a similar portion (Act 31). The lattice of FIG. 6 (a) and the lattice of FIG. 6 (b) differ only in the expression of the endings, so the interval from 0 to 18 in the input position is a similar part.

  The proportion of phoneme mismatches between the immediately preceding utterance and the current utterance is about 14% (= (22-19) × 100/22). If the threshold is set to 15%, the proportion of mismatched parts is smaller than the threshold. As a result, the utterance relation determination unit 103 does not mean that all parts of the current utterance are similar to a part of the immediately preceding utterance (Act 32: NO in FIG. 5), and the ratio of inconsistent parts is smaller than the threshold (Act 33). : NO), it is determined that the re-input current utterance is an overall corrected utterance.

Thereby, the overall correction unit 104b executes the overall correction process (Act 7 in FIG. 2).
In the overall correction process, the lattice of FIG. 6A for the previous utterance and the lattice of FIG. 6B for the current utterance are integrated, and the lattice of FIG. 6C is output.

  For example, since the same node 1202 exists in FIG. 6B with respect to the node 1201 in the section where the input position is 4 to 6 in FIG. 6A, the score 10 of the node 1201 and the score 10 of the node 1202 are added. The calculated value 20 is calculated as the score of the node 1203 in FIG.

  Further, for example, since the node 1204 corresponding to the section whose input position is 0 to 3 in FIG. 6A does not exist in FIG. 6B, the node is added to the lattice in FIG. The node 1205 of 6 (c) is output.

  In the lattice of FIG. 6C integrated by such an overall correction process, the recognition candidate selection unit 105 concatenates the nodes having the maximum score, and the recognition result “I want to watch a movie. Is generated (Act 4), and the output control unit 106 displays the recognition result on the display 112 (Act 5). As described above, the apparatus 100 generates a correct recognition result as a result of selecting a node by integrating the lattice of the immediately preceding utterance and the current utterance.

<About correction method by continuous utterance>
FIG. 7 is a diagram illustrating an example of a user's utterance.
When a user wants to say, “As a good corporate citizen, we have high ethical standards and a law-abiding spirit and fulfill our responsibility to each country and local community.” Eventually, we may interrupt the speech “We are good corporate citizens and have high ethical standards”. In this case, the apparatus 100 outputs the recognition result as “we are good corporate citizens and high ethical standards”.

  In such a case, the user makes the apparatus 100 a continuous utterance (second voice utterance) that continues to the immediately preceding utterance “has an ethical and legal spirit and fulfills responsibility to each country and community”. By inputting, the apparatus 100 can be made to recognize the full text that the user wants to speak (Act 1, 2).

  In the utterance relation determination process and the recognition linking process for linking to the recognition result of the immediately preceding utterance, the current utterance “has a spirit of ethics and legal compliance and fulfills responsibility to each country and community” is a selection candidate. The lattice contains multiple nodes, and the selection of multiple generated nodes has not been completed. However, in the following, for the sake of easy understanding, the correct utterance is selected from the current utterance recognized in the lattice state. A description such as a recognition result after completion, that is, as shown in the current utterance in FIG.

FIG. 8 is a diagram showing phonemes of the immediately preceding utterance and the current utterance.
The utterance relation determination unit 103 agrees with the previous utterance “We are a good corporate citizen, high ethical standards and societies” and the current utterance “has a spirit of ethics and legal compliance and fulfills responsibility to each country and community”. The part and the inconsistent part are calculated (Act 31). The utterance relation determining unit 103 determines, for example, from the phonemes of the immediately preceding utterance and the current utterance, the “ethics” portion of the section where the input position of the immediately preceding utterance is 18 to 23, and the section where the input position of the current utterance is 1 to 6 The “ethics” part is similar, and the other parts are calculated as inconsistent.

Then, the utterance relation determination unit 103 does not mean that part of the immediately preceding utterance and all of the current utterance are similar (Act 32: NO), and the ratio of the mismatched part is high between the immediately preceding utterance and the current utterance, and the ratio of the mismatching part is the threshold Although it exceeds 15% (Act 33: YES), a part of the previous utterance “with ethics” matches a part of the current utterance “with ethics”, so it is determined that the current utterance is a continuous utterance that continues to the previous utterance. (Act 34: YES).
Therefore, the recognition linking unit 107 performs a recognition linking process for linking the recognition result of the immediately preceding utterance and the recognition candidate of the current utterance.

  First, the recognition linking unit 107 determines whether or not there is one matching portion between the immediately preceding utterance (recognition result) and the current utterance (recognition candidate) (Act 8 in FIG. 2). When there is one matching portion as in this example (Act 8: YES), the recognition connecting unit 107 determines that the matching portion “ethical view” is a connecting portion between the immediately preceding utterance and the current utterance (Act 9).

  Then, the recognition linking unit 107 associates the current utterance with the immediately preceding utterance and overwrites the immediately preceding utterance. In other words, the recognition linking unit 107 says that “we are high as a good corporate citizen” in front of the connection part “ethics” of the previous utterance, and “compliance of legal compliance” after the connection part “ethics” of the current utterance. “I have a spirit and fulfill my responsibility to each country and local community” with the connection part “ethics” in between (Act 10).

  The recognition candidates connected in this way are lattices including a plurality of nodes that are selection candidates as described above. For example, “SE-I-SI-N” in the current utterance includes “mind”, “cleanness”, “political tribunal”, and the like as nodes.

  Therefore, the recognition candidate selection unit 105 concatenates the nodes with the highest scores, and as a recognition result, for example, “As a good corporate citizen, we have high ethical standards and a spirit of compliance, and are responsible for each country and community. Is output (Act 4), and the output control unit 106 displays the recognition result on the display 112 (Act 5). When the recognition result includes misrecognition, the user may cause the apparatus 100 to correct the misrecognized portion by partial correction utterance or overall correction utterance.

Hereinafter, with reference to FIG. 9, a recognition and connection process in the case where there are a plurality of matching portions between the immediately preceding utterance and the current utterance will be described.
The user said, “As a good corporate citizen, we have a high ethical sense and a law-abiding spirit, and to each country and local community.” The device 100 recognizes the recognition result as “We are a good corporate citizen. , Have high ethical standards and a law-abiding spirit, and output to each country and community.

  In this case, the user can input the continuous utterance “Contribute to the development of the global society by fulfilling the responsibility to society and respecting its culture and history” (Act 1, 2).

  The utterance relationship determination unit 103 calculates a matching part and a mismatching part between the immediately preceding utterance and the current utterance (Act 31). As described above, the utterance relation determination unit 103 has a ratio of the mismatched portion between the immediately preceding utterance and the current utterance being equal to or greater than the threshold (Act 32, 33: NO), Since “Society” matches, the current utterance is determined as a continuous utterance (Act 34: YES).

  Subsequently, the recognition linking unit 107 determines whether or not there is one matching portion between the immediately preceding utterance and the current utterance (Act 8). In the case of this example, in the current utterance “I will fulfill my responsibility to society and respect its culture and history and contribute to the development of the global society.” Since the “society” of a certain “global society” matches “society” in “to each country and local community” of the previous utterance, the recognition connecting unit 107 determines that there are a plurality of matching parts in the current utterance (Act 8: NO).

  As described above, when a plurality of matching parts are included in at least one of the immediately preceding utterance and the current utterance in the immediately preceding utterance and the current utterance (in this example, two matching parts “society” are included in the current utterance). The connecting unit 107 determines that the matching part behind the immediately preceding utterance and the matching part ahead of the current utterance are the connecting parts of the immediately preceding utterance and the current utterance (Act 11).

  The recognizing connecting unit 107 is connected to the front of the current utterance as “We are good corporate citizens, with high ethical standards and law-abiding spirit, each country and region” in front of the connected part “society” in the previous utterance. “Society” will be linked to “the rear side” and the culture and history will be respected to contribute to the development of the global society. ”(Act 12).

  The recognition candidates connected in this way are lattices originally including a plurality of nodes as described above.

  Therefore, the recognition candidate selection unit 105 concatenates the nodes with the highest scores, and as a recognition result, for example, “As a good corporate citizen, we have high ethical standards and a spirit of compliance, and are responsible for each country and community. The output control unit 106 displays the recognition result on the display 112 (Act 5). ”(Act 4). The output control unit 106 displays the recognition result on the display 112 (Act 5). When the recognition result includes misrecognition, the user may cause the apparatus 100 to correct the misrecognized portion by partial correction utterance or overall correction utterance.

  Note that the matching part between the immediately preceding utterance and the current utterance is limited to cases where phonemes are continuous beyond the threshold (for example, three or more phonemes are continuous), or nodes or feature information continues beyond the threshold depending on the setting. It can be limited to cases.

<New utterance judgment method>
The utterance relation determination unit 103 does not mean that part of the immediately preceding utterance and all of the current utterance are similar (Act 32: NO), and the ratio of the mismatched portion between the immediately preceding utterance and the current utterance is high and exceeds the threshold of 15% (Act 33: YES), if a part of the immediately preceding utterance and a part of the current utterance do not match, the current utterance is determined as a new utterance (Act 34: NO).

  According to this embodiment, it is not only determined whether the voice re-input to the user is an utterance for correcting the whole or an utterance for correcting a part of the voice, but an utterance intended for continuous input Whether the correction method can be changed according to the determination result. For this reason, the present embodiment does not require the user to instruct the correction method, and the optimum correction method can be automatically applied according to the input method, so that it is highly convenient and can output a highly accurate speech recognition result.

(Modification)
In the embodiment, the recognition linking unit 107 connects the recognition result (recognition) of the immediately preceding utterance and the recognition candidate (recognition) that is a lattice of the current utterance for which the node has not been selected. In the embodiment, the recognition output unit that outputs the recognition of the input speech utterance includes the recognition candidate generation unit 102 and the recognition candidate selection unit 105 in the immediately preceding utterance, and the recognition candidate generation unit 102 in the current utterance. It is comprised including.

  However, when the utterance relationship determining unit 103 determines that the current utterance is a continuous utterance, the recognition linking unit 107 selects a node for a recognition candidate that is a lattice and generates a recognition result. The recognition result (recognition) of the immediately preceding utterance may be connected to the recognition result (recognition) of the current utterance. In this case, the recognition output unit includes a recognition candidate generation unit and a recognition candidate selection unit.

The order of the processes in the embodiment may be different from the order exemplified in the embodiment.
The present invention can be implemented in various other forms without departing from the spirit or main features thereof. Therefore, the above-described embodiment is merely an example in all respects and should not be interpreted in a limited manner. The scope of the present invention is indicated by the scope of claims, and is not restricted by the text of the specification. Further, all modifications, various improvements, alternatives and modifications belonging to the equivalent scope of the claims are all within the scope of the present invention.

  DESCRIPTION OF SYMBOLS 100 ... Voice recognition apparatus, 102 ... Recognition candidate output part (recognition output part), 103 ... Speech relation determination part, 105 ... Recognition candidate connection part (recognition output part), 107 ... Recognition connection part.

Claims (5)

A recognition output unit that outputs recognition of an input speech utterance;
An utterance relation determination unit that determines whether or not the second voice utterance input later is a continuous utterance continued from the first voice utterance input first;
A recognition linking unit that connects the first recognition, which is the recognition of the first voice utterance, and the second recognition, which is the recognition of the second voice utterance, when it is determined that the second voice utterance is a continuous utterance; ,
A speech recognition apparatus comprising: The apparatus of claim 1.
The utterance relation determination unit is configured such that the entire second recognition is not included in the first recognition, the ratio of the mismatched portions of the first and second recognitions is equal to or greater than a threshold, and a part of the first recognition And a part of the second recognition, a speech recognition device that determines that the second speech utterance is a continuous utterance. The apparatus of claim 2.
In a case where a plurality of matching parts of the first and second recognitions are included in at least one of the first and second recognitions, the utterance relationship determination unit includes a matching part located behind the first recognition, A matching part in front of the two recognitions is determined as the connection part of the first and second recognitions,
The recognition connection unit is a speech recognition device that connects the front side of the rear connection part in the first recognition and the front connection part to the rear side of the second recognition with the connection part interposed therebetween. Recognize incoming speech utterances
It is determined whether the second voice utterance input later is a continuous utterance that is an utterance from the middle of the first voice utterance input first,
A speech recognition method for connecting recognition of the first speech utterance and recognition of the second speech utterance when it is determined that the second speech utterance is a continuous utterance. Recognize incoming speech utterances
It is determined whether the second voice utterance input later is a continuous utterance that is an utterance from the middle of the first voice utterance input first,
When determining that the second voice utterance is a continuous utterance, a voice that causes a computer to cause the voice recognition device to connect the recognition of the first voice utterance and the recognition of the second voice utterance. Recognition program.

Images