JP4475380B2 - Speech recognition apparatus and speech recognition program - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speech recognition method for recognizing human speech input via a microphone or the like.
[0002]
[Prior art]
In general, the speech recognition apparatus acoustically analyzes a speech input signal generated based on a user's utterance, and compares each acoustic likelihood (similarity) with a plurality of candidate word models prepared in advance. Degree) and the candidate having the highest acoustic likelihood (referred to as “first candidate”) is determined as the recognition result. If the recognition reliability of the first candidate is not sufficiently high, the speech recognition device determines that there is no correct recognition result, and talks back, such as “Please speak again” to prompt the user to speak again. The recognition process is performed again.
[0003]
The conventional speech recognition apparatus has low reliability of the recognition result, and performs the recognition process again using the same candidate as the previous time even when the user is requested to re-utter. Therefore, if the user repeats the utterance in the same way as the previous time, the recognition result will be the same as the previous time, so that the recognition rate for the recurrent utterance is not improved so much.
[0004]
[Problems to be solved by the invention]
One of speech recognition methods that improve such a point is described in Japanese Patent No. 3112037. In this speech recognition method, when a sufficiently reliable recognition result is not obtained in the recognition process for the first utterance of the user, narrowing down to a few candidates with high reliability and prompting the user to speak again . Further, with respect to a candidate having a higher degree of reliability obtained in the recognition process for the first utterance, those synonyms are added to the candidate to prompt the user to speak again, and the recognition process is performed again.
[0005]
However, in this method, if the correct answer is not included in the top candidates narrowed down by the first recognition result, the recognition cannot be performed. Further, even if a higher-ranked synonym is added to a candidate, if the user uses the same word as the previous one in a recurrent utterance, the meaning of adding the synonym to the candidate is lost.
[0006]
Another speech recognition method is described in Japanese Patent Application Laid-Open No. 11-119792. In the method described in this publication, a set of acoustically similar commands (referred to as “similar type commands”) and a corresponding paraphrase command set are defined and stored in advance. For example, when there are “raise window” and “lower window” as similar type commands, “open window” and “close window” are prepared as paraphrase commands. When the user utters a similar type command, the user requests to speak again using a paraphrase command for the utterance.
[0007]
However, in this method, it is necessary to preliminarily define the correspondence between the similar type command and the paraphrase command corresponding thereto and store it in a memory or the like. Therefore, when a large number of commands are used in the system, the storage capacity required for that purpose increases, leading to an increase in cost.
[0008]
The present invention has been made in view of the above points, and it is an object of the present invention to provide a speech recognition apparatus and a program that enable efficient and accurate recognition while minimizing the number of requests for recurrent speech to the user. .
[0009]
[Means for Solving the Problems]
According to one aspect of the present invention, the speech recognition apparatus responds to the speech input by a matching process between speech input means for receiving the speech input of the user and each standby word in a preset standby word group. Recognition processing means for performing recognition processing for determining a plurality of word candidates; determination means for determining whether or not a correct answer is included in the plurality of word candidates; and each of the plurality of word candidates and their consent word candidates And setting means comprising means for analyzing the phonemes constituting each word candidate, and when the determination means determines that the correct answer is not included in the plurality of word candidates, Each group set for each word candidate in the case where at least the consent word candidates corresponding to the word candidates are grouped from the plurality of word candidates and the consent word candidates. In between the-loop, (1) phoneme of overlapping small (2) Among the extracted combinations, Relative total sound Elementary One candidate is determined for each group so as to have a large number of combinations, and is set to the standby word group used in the next recognition process. When the determination unit determines that a correct answer is not included in the plurality of word candidates, the setting unit determines the word candidate and the consent word candidate from the plurality of word candidates and their consent word candidates. It is preferable to determine candidates between the groups as one group.
[0010]
The voice recognition apparatus receives voice input by the user such as a command, and determines word candidates corresponding to the voice input by the user by matching processing with a preset standby word. Then, it is determined whether or not a correct answer is included in the word candidate. When the determination unit determines that the correct answer is included, the word candidate is output as a recognition result. On the other hand, if the determination unit determines that the correct answer is not included, at least one of the word candidates and the consent word candidate having the same meaning as each word candidate, the consent word candidate corresponding to the word candidate is grouped into one group. (1) There is little overlap of phonemes between groups set for each word candidate (2) Among the extracted combinations, Relative total sound Elementary Candidates are determined between the groups so that there are many combinations, and these are used in the next recognition process. Therefore, since the next recognition process is performed using easy-to-identify word candidates from among the word candidates including the consent word, it is possible to improve the recognition rate of recurrent speech by the user.
[0011]
In one aspect of the speech recognition apparatus, the setting unit includes a unit that analyzes a phoneme constituting each word candidate for each of the plurality of word candidates and their consent word candidates, and a word having the least phoneme overlap. Means for setting a combination of candidates as the standby word.
[0012]
According to this aspect, word candidates including consent word candidates are analyzed from the aspect of phonemes that are their constituent elements, and a combination of word candidates having the smallest phoneme overlap is used as a standby word. Therefore, the recognition processing can be performed in a state where it is easy to distinguish from each other in the speech recognition processing.
[0013]
In another aspect of the speech recognition apparatus described above, the setting unit includes: a unit that analyzes a phoneme constituting each word candidate for each of the plurality of word candidates and their consent word candidates; And a means for setting a combination of word candidates having a small total number of phonemes as the standby word.
[0014]
According to this aspect, word candidates including consent word candidates are analyzed from the aspect of phonemes that are their constituent elements, and a combination of word candidates with the smallest phoneme overlap and the largest total phoneme number is set as a standby word. Use as Therefore, the recognition process can be performed in a state in which the voice recognition process is easier to distinguish.
[0015]
In another aspect of the above speech recognition apparatus, the setting means determines a standby error word that means that the user's voice input corresponds to a word candidate other than the word candidate included in the standby word as the standby word. Can be included in a group. Therefore, when the current standby word does not include the correct answer, the user utters the standby error word, so it can be determined whether or not the current standby word includes the correct answer.
[0016]
In still another aspect of the speech recognition apparatus, the setting unit includes a storage unit that stores a standby word group used in the past, and when the determination unit determines that the standby error word is correct, The setting means can set the previous standby word group stored in the storage means to the standby word group used in the next recognition process. As a result, when the correct answer is not included in the current standby word group, the range of the standby words can be expanded to search for the correct answer.
[0017]
In still another aspect of the speech recognition apparatus, the standby error word may be “other” or a synonym thereof.
[0018]
In still another aspect of the speech recognition apparatus, when the user's speech input is the standby error word, the standby error word corresponds to the word candidate in the standby word group at that time. It is possible to exclude word candidates other than the word candidates to be included from the word candidates to be included in the next standby word group. Since the standby error word indicates that there is no correct answer in the word candidates in the current standby word group, it does not make sense to include them in the next standby word group. Therefore, by excluding those word candidates that are known not to be correct from the next word candidates, it is possible to narrow down the word candidates and obtain a correct answer efficiently.
[0019]
Still another aspect of the speech recognition apparatus described above is, when the determination unit determines that a correct answer is not included in the plurality of word candidates, standby words belonging to the standby word group set by the setting unit, Notification means for notifying the user by at least one of synthesized speech output or character display can be provided. Thereby, since the standby word is notified to the user by the synthesized speech, the user can easily know the word to be re-spoken.
[0020]
In still another aspect of the speech recognition apparatus, the determination unit can relax a criterion for determining the word candidate as correct each time the recognition process is repeated. As a result, the correct answer can be easily obtained as the recognition process is repeated, and the efficiency of the recognition process can be increased. In one preferable example, the determination unit determines that the word candidate is correct when the reliability of the word candidate is equal to or higher than a predetermined threshold, and each time the recognition process is repeated. The threshold value can be lowered.
[0021]
In still another aspect of the above speech recognition apparatus, the setting unit determines a combination of candidates that are most easily identified from the plurality of word candidates, their consent word candidates, and the standby error word, and performs next recognition. The standby word group used in the processing can be set. Therefore, the recognition process can be performed in a state in which the voice recognition process is easier to distinguish.
[0022]
In another aspect of the present invention, a speech recognition program executed by a computer uses the computer to perform a matching process with each standby word in a preset standby word group by using a voice input unit that receives a user's voice input. Recognition processing means for performing recognition processing for determining a plurality of word candidates corresponding to the voice input, means for determining whether or not a correct answer is included in the plurality of word candidates, the plurality of word candidates and their agreed words For each of the candidates, the determination means is configured to function as a setting unit including a unit that analyzes a phoneme constituting each word candidate, and the determination unit determines that a correct answer is not included in the plurality of word candidates. When the means determines, at least one consent word candidate corresponding to the word candidate is selected from the plurality of word candidates and their consent word candidates. Among the groups in the case of the groups (1) phonemes overlap small (2) Among the extracted combinations, Relative total sound Elementary One candidate is determined in each group so as to have a large number of combinations, and functions are set to the standby word group used in the next recognition process.
[0023]
DETAILED DESCRIPTION OF THE INVENTION
Preferred embodiments of the present invention will be described below with reference to the drawings.
[0024]
[Configuration of voice recognition device]
FIG. 1 shows a functional configuration of a speech recognition apparatus according to an embodiment of the present invention. In FIG. 1, a speech recognition device 10 includes a subword acoustic model storage unit 1, a dictionary 2, a word model generation unit 3, an acoustic analysis unit 4, a recognition processing unit 5, an additional information collection unit 6, and a recognition trust. A degree calculation unit 7, a recurrent speech control unit 8, a synthesized speech generation unit 9, a speaker 11, a microphone 12, and a switch SW1.
[0025]
The subword acoustic model storage unit 1 stores an acoustic model in units of subwords such as previously learned phonemes. Here, the “phoneme” is a minimum unit in which sound used in a certain language is analyzed and defined from the viewpoint of the discrimination function, and is classified into consonants and vowels. A “subword” is a unit constituting each word, and one word is constituted by a set of subwords. The subword acoustic model storage unit 1 stores subword acoustic models corresponding to phonemes such as vowels and consonants. For example, in the case of the word “aka”, the subwords constituting this are “a”, “k”, and “a”.
[0026]
The dictionary 2 stores word information related to words that are subject to speech recognition. Specifically, for a plurality of words, combinations of subwords constituting the word are stored. For example, for the word “red”, it is stored as information that the subwords constituting it are “a”, “k”, and “a”.
[0027]
The word model generation unit 3 generates a word model that is an acoustic model of each word. Specifically, for a certain word, a word model of the word is generated using word information stored in the dictionary 2 and a subword acoustic model stored in the subword acoustic model storage unit 1. For example, in the case of the word “Aka”, the dictionary 2 stores that the word “Aka” is composed of subwords “a”, “k”, and “a” as word information. Also, subword acoustic models corresponding to the subwords “a”, “k”, and “a” are stored in the subword acoustic model storage unit 1. Therefore, the word model generation unit 3 refers to the dictionary 2 to check the subwords constituting the word “red”, obtains the subword acoustic model corresponding to them from the subword acoustic model storage unit 1 and combines them to obtain the word “red” (Aka) "is generated.
[0028]
The acoustic analysis unit 4 acoustically analyzes the speech signal input to the speech recognition device 10 via the microphone 12 and converts it into a feature vector series. The recognition processing unit 5 collates the feature vector of the utterance voice obtained by the acoustic analysis unit 4 with a plurality of word models generated by the word model generation unit 3 (matching process), and each word model for the user's utterance voice. Calculate the acoustic likelihood of. The word model collated at this time is also called a “word candidate”. The recognition processing unit 5 matches a plurality of predetermined word candidates with a feature vector sequence corresponding to the user's uttered speech, and calculates an acoustic likelihood for each word candidate.
[0029]
In practice, when a user utters a word, several words that the user is expected to utter in that situation are determined as word candidates (this is also referred to as a “standby word”). . When a feature vector sequence corresponding to the user's utterance is obtained, it is matched with a predetermined word candidate (standby word), and the acoustic likelihood for each word candidate is calculated individually.
[0030]
The additional information collection unit 6 collects additional information such as the user's past utterance history. When the voice input device of the present invention is used for a command input unit of a car navigation device, the additional information includes position information of a vehicle on which the car navigation device is mounted. The recognition reliability calculation unit 7 calculates the recognition reliability of each word candidate based on the acoustic likelihood of each word candidate for the user utterance calculated by the recognition processing unit 5. The recognition reliability is an index indicating to what degree the probability that the word candidate matches the word actually spoken by the user. The higher the recognition reliability, the higher the probability that the word candidate matches the word spoken by the user, that is, the probability of being correct. The lower the recognition reliability, the lower the probability that the word candidate is correct.
[0031]
Specifically, the recognition reliability calculation unit 7 weights the acoustic likelihood of each word candidate calculated by the recognition processing unit 5 using the additional information obtained by the additional information collection unit 6. The recognition reliability of each word candidate with respect to the user's speech is calculated. For example, when the additional information collected by the additional information collecting unit 6 has a history that the user frequently utters a word in the past, the recognition reliability of the same word candidate is set high. . In addition, when a word related to the current position of the vehicle is uttered, the reliability of the word can be set high. Note that the above-described method of calculating the recognition reliability is an example, and various other methods of calculating the recognition reliability can be used in the present invention.
[0032]
The recurrent speech control unit 8 is an element that plays a central role in the present invention, and controls word candidates in the recurrent speech. FIG. 2 shows the internal configuration of the recurrent speech control unit 8. As shown in FIG. 2, the recurrent speech control unit 8 includes a reliability analysis unit 81, a candidate selection unit 82, a standby word selection unit 83, a first candidate information extraction unit 84, and a synthesized speech information generation unit 85. And a switch SW2.
[0033]
The recurrence control unit 8 receives the reliability information 20 from the recognition reliability calculation unit 7. The reliability information 20 includes word candidate information indicating a plurality of word candidates for the user's utterance, and recognition reliability information calculated by the recognition reliability calculation unit 7 for each word candidate. That is, the reliability information 20 indicates which word candidate has what degree of reliability.
[0034]
The reliability analysis unit 81 may determine the word candidate having the highest reliability (hereinafter referred to as “first word candidate”) among the plurality of word candidates included in the reliability information 20 as the recognition result. Or not, that is, whether or not the first word candidate is correct. This determination can be made using, for example, the reliability of the first word candidate and the reliability of the second word candidate. That is, the reliability of the first word candidate is sufficiently high and is greater than or equal to a predetermined threshold α (condition 1), and the difference between the reliability of the first word candidate and the reliability of the second word candidate is When the two conditions of sufficiently large and above a predetermined threshold β (condition 2) are satisfied, the first word candidate is determined to be correct. On the other hand, if neither of the conditions 1 and 2 is satisfied, the first word candidate is not regarded as a correct answer. A method other than the above can also be adopted as a method of determining the first word candidate as the correct answer. For example, it is possible to determine whether or not the first word candidate is correct by using the reliability of a predetermined number n of word candidates.
[0035]
When it is determined that the first word candidate is correct, the recognition level analysis unit 81 supplies the switches SW1 and SW2 with control signals that connect the switch SW1 shown in FIG. 1 and the switch SW2 shown in FIG. 2 to the terminal T1. To do. On the other hand, when it is determined that the first word candidate is not correct, the recognition degree analysis unit 81 supplies a control signal for connecting the switches SW1 and SW2 to the terminal T2 to the switches SW1 and SW2.
[0036]
The first candidate information extraction unit 84 receives the reliability information 20 from the recognition reliability calculation unit 7 via the switch SW2 when the reliability analysis unit 81 determines that the first word candidate is correct. Then, information indicating that the first word candidate is correct, information indicating what is the first candidate word determined to be correct, pronunciation information corresponding to the first word candidate, and the like, a synthesized speech information generation unit 85. The first candidate information extraction unit 84 outputs information indicating what is the first candidate word to the outside as a recognition result.
[0037]
When the first word candidate is correct, the synthesized voice information generation unit 85 generates synthesized voice information for notifying the user of the recognition result based on the information from the first candidate information extraction unit 84. Output to the synthesized speech generation unit 9.
[0038]
The synthesized speech generation unit 9 shown in FIG. 1 generates a synthesized speech including a word determined to be correct based on the synthesized speech information input from the synthesized speech information generation unit 85, and recognizes it by outputting from the speaker 11. Notify the user of the result. Notifying the user of the recognition result means that, for example, when the word candidate determined to be correct is “red”, a synthesized speech such as “red” is output. Thereby, the user confirms the recognition result. In this embodiment, a method of notifying the user of the recognition result by audio output from the speaker 11 is employed, but instead of or in addition to this, the user is visually notified of the recognition result by a display or the like. It can also be configured as follows.
[0039]
On the other hand, when the recognition level analysis unit 81 determines that the first candidate is not correct, the speech recognition apparatus 10 requests the user to re-utter. In that case, the switch SW2 is connected to the terminal T2, and the reliability information 20 is supplied to the candidate selection unit 82. The switch SW1 is also connected to the terminal T2, and the standby word selection unit 83 is connected to the word model generation unit 3. The candidate selection unit 82 narrows down some word candidates with high reliability (hereinafter referred to as “correct word candidates”) from among all word candidates for which the reliability is calculated. For example, a word candidate whose reliability difference from the first word candidate is a predetermined threshold value γ or less is set as a correct word candidate. Then, the identification information of the determined correct word candidate is supplied to the standby word selection unit 83.
[0040]
The standby word selection unit 83 determines a standby word group for the user's recurrent utterance (that is, a combination of words used as word candidates in the recognition process for the user's recurrent utterance). This most typical method is a method of setting the correct word candidate selected by the candidate selection unit 82 as a standby word. As a result, in the previous speech recognition process, candidates with high recognition reliability are set as standby words. However, in this case, when the user's previous utterance and the recurrent utterance are exactly the same (for example, when the user repeatedly utters “red”), the recognition result may not be determined to be correct as in the previous utterance. There is sex. Therefore, in the present invention, the word used as the standby word in the recurrent utterance is a different word that is a synonym of the correct word candidate and can be easily identified by the recognition process, thereby increasing the recognition rate in the recurrent utterance. I have to. That is, the standby word selection unit 83 sets a combination of words that are synonyms and are easily identified based on the correct word candidates supplied from the candidate selection unit 82 as standby words for recurrence. Here, one suitable example of “a combination of easy-to-identify words” is a synonym of a correct word candidate, and there is little overlap of phonemes (condition A) and a large number of total phonemes (condition B). It is a combination of words. This is because, when words are compared acoustically from the viewpoint of speech recognition, generally, the fewer phonemes overlap and the greater the number of phonemes in a word, the easier it is to identify the words.
[0041]
This will be described in more detail. In the dictionary 2, synonyms (consent words) having the same meaning and different pronunciation are prepared for one word. Now, it is assumed that there are two correct candidate words selected by the candidate selection unit 82: “aka” and “ao”. Also, “redo (reqdo)” is stored in the dictionary 2 as a consensus word for “red”, and “buruu” is stored in the dictionary 2 as a consensus word for “red”. Assume. In this case, the phoneme “a” is duplicated between “red” and “blue”, and the phoneme “o” is duplicated between “red” and “blue”. The combination is “red” and “blue”, or “red” and “blue”. Further, considering condition B, among these combinations, the combination of “Rado” and “Buru” has a larger total number of phonemes. "Is set as a standby word. As another example, for example, when “mizuiro” is further stored in the dictionary 2 as a consensus word for “Ao”, the total number of phonemes is the highest among the combinations having the smallest number of identical phonemes. Many combinations of “Aka” and “Mizuiro” are set as standby words. Thus, in the present invention, among the correct word candidates and their consent words, the most easily identifiable combination is set as a standby word for the next recurrent utterance. Thereby, the recognition accuracy in the recognizing process for recurrent speech can be improved.
[0042]
Further, in the present invention, in the case of recurrent utterance, words such as “other”, “other”, and “different” indicating that the word included in the talkback prompting recurrent utterance is other than the correct word are recurred. It is characterized by including it in the talkback that encourages the talk. Thereby, when the correct answer is not included in the word asked to the user by the talkback prompting the recurrent speech, the speech recognition apparatus 10 can know it. For example, as a result of the first utterance, the correct answer candidate words are narrowed down to “red” and “blue”, and “red” and “Mizuiro” are finally determined as standby words as described above. In that case, in the talkback prompting the recurrence, is the voice recognition device 10 for example “Aka”, “Mizuiro” or “Other”? " On the other hand, if the user re-speaks “other”, it is understood that the word spoken by the user is neither “red” nor “Mizuiro”. Then, the voice recognition device 10 recognizes that the previous narrowing-down was an error, and can search for word candidates other than “Aka” and “Mizuiro”.
[0043]
In this way, the standby word selection unit 83 supplies information including the number, pronunciation, meaning (reading of the original word), and the like of standby words for recurrent speech to the word model generation unit 3 via the switch SW1 as standby word information 83a. At the same time, it is supplied to the synthesized speech information generation unit 85. In this case, the word model generation unit 3 generates a word model of the standby word included in the standby word information 83a and uses it for the matching process by the recognition processing unit 5 in the recognizing speech recognition process. In other words, in the above-described example, the word models “Aka”, “Mizuiro”, and “Other” are subjected to the matching process in the recognition process of the re-spoken word. In addition, the synthesized voice information generation unit 85 is “Aka”, “Mizuiro”, or “Other” as a talkback for prompting a recurrence based on the standby word information 83a? ”Is generated. The synthesized voice information is output as synthesized voice from the speaker 11 by the synthesized voice generation unit 9.
[0044]
In this way, the speech recognition apparatus 10 includes words of combinations that are easy to identify among correct answer candidate words during talkback, and further includes words such as “others” indicating other words. Encourage users to talk again. Thereby, the recognition accuracy at the time of recurrent speech can be improved.
[0045]
If the first word candidate cannot be determined to be correct in the recognition process after re-utterance, the same re-utterance process can be repeated. In the re-utterance process, the threshold value used when the reliability analysis unit 81 determines the first word candidate as the correct answer may be gradually relaxed so that the correct answer is easily determined. it can.
[0046]
Further, when it is determined that the word candidate corresponding to the word “others” is correct in the recurrent utterance (including the case of multiple times), that is, in the current standby candidate word designated by the user during the talkback If it is determined that there is no correct answer, the standby word selection unit 83 returns the standby word to the state at the time of the previous utterance. The reason is as follows. For example, when it is determined in the recognition process for the m-th utterance that the first word candidate is not correct, the candidate selection unit 82 narrows down the standby words for the (m + 1) -th utterance to only the top candidates. However, the fact that the user uttered “other” in the (m + 1) th utterance means that there is no correct word among the standby words set at that time, and narrowing down was incorrect ( It means "waiting error". Therefore, the standby words are narrowed down and returned to the previous state (the state at the m-th utterance) to widen the range of word candidates, and if necessary, recurrent speech is urged.
[0047]
In this case, the reliability analysis unit 81 connects the switches SW1 and SW2 to the terminal T2. The standby word selection unit 83 stores the previous standby word group when determining a standby word group for the next utterance. That is, the standby word selection unit 83 stores all the past standby word groups, and in the case of a standby error, the standby word group of the previous time is used in the next recurrence and recognition process.
[0048]
Thus, if necessary, the recurrent utterance is repeated, and when the reliability analysis unit 81 finally determines that a certain first word candidate is correct, the first word candidate is recognized as a recognition result from the speech recognition device 10 to the external device. Sent to. The external device is a device that uses a recognition result by the voice recognition device 10 as a command or the like. For example, when the speech recognition device 10 is used as the input unit of the car navigation device as described above, the recognition result is supplied to the controller of the car navigation device, and processing corresponding to the content (command) is executed.
[0049]
[Voice recognition processing]
Next, the speech recognition process executed by the speech recognition apparatus 10 will be described with reference to FIG. FIG. 3 is a flowchart of the voice recognition process.
[0050]
First, in step S1, initial setting for recognizing the first utterance by the user is performed. Specifically, the recurrent utterance control unit 8 connects the switch SW1 to the terminal T1 side, and sets all words in the dictionary 2 storing word candidate information to be recognized as standby words for the first utterance. Then, the utterance counter c is set to 1. The utterance counter indicates a standby word group for an utterance to be recognized. That is, the utterance counter c = 1 corresponds to the standby word group for the first utterance (in the above example, all words stored in the dictionary 2), and the utterance counter c = 2 is narrowed down once after the first utterance. Corresponds to the standby word group.
[0051]
Next, in step S <b> 2, based on the standby word group, the word model generation unit 3 uses the subword acoustic model stored in the subword acoustic model storage unit 1 to generate a word model. Thereby, all the word models corresponding to the standby word group for the first utterance are prepared.
[0052]
Next, voice recognition processing is performed in step S3. That is, the user speaks, and the corresponding speech signal is input to the acoustic analysis unit 4 via the microphone 12. The acoustic analysis unit 4 performs acoustic analysis of the speech signal and obtains a feature vector series. And the recognition process part 5 performs the matching process with the feature vector of an utterance audio | voice signal, and each word model prepared in step S2, and calculates the acoustic likelihood between both for every word model.
[0053]
Next, in step S4, the recognition reliability calculation unit 7 weights the acoustic likelihood for each word candidate calculated in step S3 by using the additional information collected by the additional information collection unit 6, The recognition reliability of each word candidate is calculated. The additional information is, for example, the user's past utterance history or the position information of the vehicle on which the navigation device is mounted.
[0054]
Next, in step S5, the reliability analysis unit 81 analyzes whether the first word candidate with the highest recognition reliability is a correct answer based on the recognition reliability of each word candidate. As described above, this analysis can be performed using the reliability of the first word candidate and the reliability of the second word candidate, for example.
[0055]
Next, in step S6, the reliability analysis unit 81 determines whether or not the first word candidate is correct based on the result of the analysis in step S5. If it is determined that the first word candidate is correct, the process proceeds to step S7. On the other hand, if it is determined that the first word candidate is not correct, the process proceeds to step S14.
[0056]
When it is determined in step S6 that the first word candidate is correct, in step S7, the reliability analysis unit 81 determines whether or not the first word candidate is a word corresponding to “other”. As described above, the word candidates corresponding to “others” are used to correct the standby word group when the correct words are not included in the standby words as a result of narrowing down the standby words. If the first word candidate corresponds to “others”, the process proceeds to step S10. If the first word candidate does not correspond to “others”, the process proceeds to step S8.
[0057]
When the process proceeds to step S8, it means that the first word candidate is correct and is not an “other” word candidate. That is, the first word candidate is good as a recognition result. Therefore, the first candidate information extraction unit 84 extracts the first word candidate from the reliability information 20, information indicating that the first word candidate is correct, and what is the first candidate word determined to be correct. The information indicating the first candidate word and the pronunciation information corresponding to the first word candidate are supplied to the synthesized speech information generation unit 85, and the information indicating the first candidate word determined to be correct is output as a recognition result to the outside. To do.
[0058]
In step S9, based on the pronunciation information corresponding to the first word candidate, the synthesized speech information generating unit 85 generates synthesized speech information and supplies it to the synthesized speech generating unit 9. The synthesized speech generating unit 9 The reading of the candidate word is output as synthesized speech. For example, when the first word candidate is “red”, the speaker notifies the user of the recognition result as “red?”.
[0059]
On the other hand, if it is determined in step S6 that the first word candidate is not correct, the candidate selection unit 82 selects the correct word candidate in step S14. Specifically, the candidate selection unit 82 selects a correct word candidate using the recognition reliability of the first word candidate. By this processing, word candidates used in the recognition processing at the next utterance are narrowed down.
[0060]
Next, in step S15, based on the correct word candidate selected by the candidate selection unit 82, the standby word selection unit 83 generates a combination of words that are easy to identify and have different pronunciations. Specifically, the standby word selection unit 83 determines a word candidate having the smallest number of identical phonemes and a large number of total phonemes as a standby word among combinations of consent words corresponding to correct answer candidate words. Then, a standby word group including those standby words is set. In this standby word group, in addition to the above words, words corresponding to “others” are included. And the standby word selection part 83 acquires the word information corresponding to those standby words from the dictionary 2, sends it to the word model production | generation part 3, and produces a corresponding word model. Thus, the standby word group is updated.
[0061]
The standby word selection unit 83 stores the standby word group before update. This is because when the user speaks “other” in the next utterance, it is necessary to use the standby word group from the previous one again. The standby word selection unit 83 also supplies the selected standby word group to the synthesized speech information generation unit 85.
[0062]
In step S16, the synthesized speech information generation unit 85 and the synthesized speech generation unit 9 output the synthesized speech of the standby word determined in step S15 as a talkback that encourages re-speech. For example, if the standby word is determined to be “Aka”, “Ao”, or “Other” in step S15, a synthesized voice such as “Aka, Ao? Is output.
[0063]
Next, in step S17, the utterance counter c is incremented by one. As a result, the increased utterance counter c indicates that the standby word group has shifted to the state after one update with respect to the previous standby word group. Then, the process returns to step S2, a word model of words included in the standby word group determined in step S15 is generated, and a recognition process for the recurrent utterance is performed.
[0064]
If it is determined in step S7 that the first word candidate corresponds to “others”, it indicates that the word in the standby word group at that time does not include the correct word, that is, a standby error. Yes. Therefore, the process proceeds to step S10, and it is determined whether or not the value of the utterance counter c is 1. When the utterance counter c = 1, the current recognition process is performed for the first utterance, and the combinations of standby words at that time are set for all word candidates included in the dictionary. Therefore, the word spoken by the user is not originally included in the dictionary 2. In this case, the recognition process is terminated with no candidate.
[0065]
On the other hand, if the utterance counter c is not 1, the process proceeds to step S11. In step S11, the standby word selection unit 83 subtracts 1 from the value of the utterance counter c, and sets the previous standby word group stored in advance. When the user utters “others”, the correct word is not included in the current standby word group, so the recognition process is performed again after returning to the standby word group used in the previous recognition process. Is done. Note that when the standby word is updated in step S14, the standby word selection unit 83 stores the standby word group in the state before the update, so that it may be read and set. At that time, the standby word selection unit 83 causes the word corresponding to “others” (also referred to as “standby error word”) to be included in the standby word group.
[0066]
Next, in step S <b> 12, the standby word selection unit 83 supplies the standby word group thus determined to the word model generation unit 3 and the synthesized speech information generation unit 85. The word model generation unit 3 generates a word model corresponding to these standby words so that it can be used in the next recognition process. In addition, the synthesized speech information generation unit 85 and the synthesized speech generation unit 9 use the supplied standby word information to output a synthesized speech of the corresponding word.
[0067]
As described above, the user's utterance is determined until the first word candidate is determined to be the correct answer and is output as a recognition result (step S9) or the recognition process is terminated as no candidate (step S10: Yes). Recognition processing is performed while the standby word group is updated according to the content. If the reliability of the first word candidate is not high enough to determine that it is correct, the standby words are narrowed down based on the reliability, etc., and further, the consensus word of the narrowed word is acoustically identified The standby word group is updated by setting the easy-to-use combination words as standby words for the next utterance. Therefore, the recognition rate due to recurrent speech can be improved, and the user's speech can be recognized quickly and efficiently.
[0068]
[Modification]
In the recurrent speech control unit 8 illustrated in FIG. 2, the reliability analysis unit 81 determines whether or not the first word candidate is correct using the first word candidate and the second word candidate. Instead, the reliability analysis unit 81 can be configured to determine whether or not the first word candidate is correct using the top n candidate words with the highest recognition reliability. In that case, n word candidates having higher recognition reliability are determined during the process of determining whether or not the first word candidate is correct. Therefore, when n word candidates having higher recognition reliability are determined, they can be used as correct word candidates after narrowing down. In this way, the reliability analysis unit 81 can perform the process of the candidate selection unit 82, and the candidate selection unit 82 can be omitted. In this case, information on correct word candidates is input from the reliability analysis unit 81 to the standby word selection unit 83.
[0069]
In the speech recognition processing shown in FIG. 3, when it is determined in step S7 that the first word candidate corresponds to “others” and it is determined in step S8 that the utterance counter c is not 1, the utterance counter is decremented by 1. The previous standby word group is set to be used in the next utterance. However, if the determination in step S7 is “Yes”, it means that there is no correct word in the previous standby word group, so that these words are included in the next standby word group. There is no meaning. For example, in the utterance using the standby word group “Aka”, “Ao”, “Other”, the user uttered “Other” means that the user uttered the word “Aka” or “Ao”. not. Therefore, the standby word selection unit 83 can set a standby word group by excluding “red”, “blue”, and their synonyms from the previous standby word group acquired in step S11. Thereby, it is possible to further improve the efficiency of the recognition process by removing words that are clearly not correct from the standby word group.
[0070]
In addition, it is possible to implement | achieve said speech recognition apparatus by comprising each element which comprises said speech recognition apparatus 10 as a computer program, and performing in the apparatus provided with a computer. For example, in a car navigation device or an AV device equipped with a computer, the voice input function can be realized by using the above computer program.
[0071]
In the above embodiment, among the correct word candidates and their consent words, the most easily identified combination is set as the standby word for the next recurrent utterance. May be determined.
[0072]
In addition to the correct word candidates and their agreed words, a standby error word indicating that the word included in the talkback for prompting recurrent speech is other than the correct word may be added to determine the most easily identifiable combination. .
[0073]
【The invention's effect】
As described above, according to the present invention, when there is a high possibility that the recognition result is erroneous, the possibility of misrecognition can be reduced by prompting the user to speak again. Also, if the recognition result for a certain utterance cannot be determined to be correct, the standby word used at that time and the consent word, which are acoustically easy to identify, are set as the standby word for the next utterance. Thus, the same recognition result is not repeated, and the recognition rate by the next utterance is improved. In addition, by including words such as “others” indicating words other than the current standby word in the confirmation talkback prompting recurrent speech, it is possible to remove words that are not correct, thereby efficiently and quickly. It is possible to reach a correct answer.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a schematic configuration of a speech recognition apparatus according to an embodiment of the present invention.
2 is a block diagram showing an internal configuration of a recurrent speech control unit shown in FIG. 1. FIG.
FIG. 3 is a flowchart showing voice recognition processing by the voice recognition apparatus shown in FIG. 1;
[Explanation of symbols]
1 Subword acoustic model storage
2 Dictionary
3 Word model generator
4 Acoustic analysis section
5 recognition processing section
6 Additional information collection department
7 Recognition reliability calculator
8 Recurrence control unit
9 Synthetic speech generator
10 Voice recognition device
11 Speaker
12 Microphone
81 Reliability Analysis Department
82 Candidate selection part
83 Standby word selector
84 1st candidate information extraction part
85 Synthetic voice information generator

Claims (10)

Voice input means for receiving the user's voice input;
Recognition processing means for performing recognition processing for determining a plurality of word candidates corresponding to the voice input by matching processing with each standby word in a preset standby word group;
Determining means for determining whether or not a correct answer is included in the plurality of word candidates;
Setting means comprising means for analyzing the phonemes constituting each word candidate for each of the plurality of word candidates and their consent word candidates;
The setting means, when the determination means determines that a correct answer is not included in the plurality of word candidates, the consent word candidate corresponding to at least the word candidate from the plurality of word candidates and their consent word candidates the between each group the set for each word candidate in the case of the one group, (1) extracts the combination phonemes overlap has little, (2) among combinations the extracted relatively total said to be a combination sound prime number is often determines one candidate in each group, the speech recognition apparatus characterized by setting the standby word group is used in the next recognition processing.   When the determination unit determines that a correct answer is not included in the plurality of word candidates, the setting unit determines the word candidate and the consent word candidate from the plurality of word candidates and the consent word candidates. The speech recognition apparatus according to claim 1, wherein candidates are determined between the groups as one group.   2. The standby word group includes a standby error word meaning that the voice input of the user corresponds to a word candidate other than the word candidate included in the standby word, or the standby word group. The speech recognition apparatus according to 2. The setting means includes storage means for storing standby word groups used in the past,
When the determination unit determines that the standby error word is correct, the setting unit sets the previous standby word group stored in the storage unit as a standby word group to be used in the next recognition process. The speech recognition apparatus according to claim 3.   The speech recognition apparatus according to claim 3, wherein the standby error word is “other” or a synonym thereof.   If the user's voice input is the standby error word, the word candidates other than the word candidate corresponding to the standby error word among the word candidates in the standby word group at that time are determined as the next standby word. The speech recognition apparatus according to claim 3, wherein the speech recognition apparatus is excluded from word candidates to be included in a group.   When the determination unit determines that a correct answer is not included in the plurality of word candidates, the standby word belonging to the standby word group set by the setting unit is sent to the user by at least one of synthesized speech output or character display. The voice recognition apparatus according to claim 1, further comprising a notification unit that notifies the user.   The speech recognition apparatus according to claim 1, wherein the determination unit relaxes a criterion for determining the word candidate as a correct answer every time the recognition process is repeated.   The determination unit determines that the word candidate is correct when the reliability of the word candidate is equal to or higher than a predetermined threshold, and decreases the threshold every time the recognition process is repeated. The speech recognition apparatus according to claim 8, wherein: A speech recognition program executed by a computer comprising:
Voice input means for receiving the voice input of the user,
A recognition processing means for performing a recognition process for determining a plurality of word candidates corresponding to the voice input by a matching process with each standby word in a preset standby word group;
Means for determining whether or not a correct answer is included in the plurality of word candidates;
For each of the plurality of word candidates and their consent word candidates, function as setting means comprising means for analyzing the phonemes constituting each word candidate,
And when the determination means determines that the correct answer is not included in the plurality of word candidates, the setting means corresponds to at least the word candidates from the plurality of word candidates and their consent word candidates agree word candidates among the groups in the case of a single group (1) extracts the combination overlap has little of phonemes, (2) among combinations the extracted, often relatively Sooto prime number A speech recognition program that functions to determine one candidate in each group so as to be combined and to set the candidate word group to be used in the next recognition process.

Images