JP4392581B2 - Language processing apparatus, language processing method, program, and recording medium - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a language processing device, a language processing method, a program, and a recording medium, and more particularly, to a language processing device, a language processing method, and a program that enable a registered word to be commonly recognized by a plurality of applications.
[0002]
[Prior art]
Speech recognition includes isolated word recognition for recognizing a single word and continuous word recognition for recognizing a word string composed of a plurality of words. In conventional continuous word recognition, having a “database of easy connection between words” called a language model prevents “sounds that are similar but messy words” from being generated as recognition results. It is out.
[0003]
However, since the language model describes only information about words that can be recognized from the beginning (hereinafter referred to as known words as appropriate), the words that are registered later (hereinafter referred to as registered words as appropriate) are correctly recognized. It was difficult to do. In other words, in isolated word recognition, if a word is registered in the recognition dictionary, the word is recognized thereafter. However, in continuous word recognition, it is not sufficient to register in the dictionary. Although it is necessary to reflect it, it was generally difficult to reflect it in the language model.
[0004]
Therefore, it has been proposed to classify registered words into categories such as “person name” and “place name” and prepare recognition grammars corresponding to the categories to recognize speech (for example, see Patent Document 1).
[0005]
Further, in a system in which there are a plurality of applications that use voice recognition and there are variable applications, when a word registered in one application is reflected in another application, a problem different from the case where there is one application occurs. . For example, if word registration is performed only for an application that has already been activated, it is difficult to reflect the registered word in an application that has been activated or installed after registration, unlike the case of a single application. There was a problem.
[0006]
Furthermore, when there are a plurality of applications, it is troublesome to delete the same registered word many times with a plurality of applications. Further, when there are a plurality of applications, it is easy to delete all registered words, but there is a problem that it is difficult to delete only a part of them or change pronunciation.
[0007]
That is, when there is one application, for example, a registered word to be deleted or changed can be specified by information such as “word registered at the nth time” or “nth entry in the recognition dictionary”, but there are a plurality of applications. In this case, it is difficult to specify each application because “word registered at the nth time” and “the number of dictionary entries added” differ.
[0008]
In addition, when there are a plurality of applications, a registered word can be specified by pronunciation. However, if a registered word is specified by pronunciation, there is a possibility that the homonym is deleted or changed.
[0009]
Therefore, instead of each application performing voice recognition individually, it is proposed that a module called “voice commander” performs voice recognition for all applications and transfers the recognition result to each application (for example, patents). Reference 1).
[0010]
[Patent Document 1]
JP 2001-216128 A
[0011]
[Problems to be solved by the invention]
However, in the invention described in Patent Document 1, the “voice commander” needs to have a recognition dictionary and a language model corresponding to each application. In other words, when developing “Speech Commander”, it is necessary to prepare a recognition dictionary and language model suitable for it, assuming which applications are used at the same time. On the other hand, there is a problem that it is difficult to reflect registered words.
[0012]
The present invention has been made in view of such a situation, and enables registered words to be used in common by a plurality of applications.
[0013]
[Means for Solving the Problems]
The language processing apparatus of the present invention includes a registered dictionary storage means for storing a registered dictionary in which words are registered, and a dedicated dictionary dedicated to the application in which words to be used for language processing used in the application are registered. For each application, Construction means to build based on the registered dictionary and , Including processing means for performing processing for adding, deleting, or changing words in the registered dictionary, and deleting means for deleting words in the dedicated dictionary, after all the words registered in the dedicated dictionary have been deleted The construction means rebuilds a dedicated dictionary based on a registered dictionary in which words are added, deleted, or changed It is characterized by that.
[0015]
The dedicated dictionary includes at least a fixed dictionary in which a predetermined word is registered in advance and a variable dictionary in which the registered word is variable, and the construction unit is configured to construct a variable dictionary of the dedicated dictionary. Can do.
[0016]
The dedicated dictionary further includes a category table in which word categories are registered, and the construction means registers the variable dictionary by registering the words in the category registered in the category table among the words in the registered dictionary in the variable dictionary. Can be built.
[0017]
Further provided is a language model storage unit that stores a language model that describes chain information indicating how words of a category are linked, and a recognition processing unit that performs speech recognition based on the dedicated dictionary and the language model. be able to.
[0018]
The language processing method of the present invention includes a registered dictionary storage step for storing a registered dictionary in which words are registered, and a dedicated dictionary dedicated to the application in which words to be used for language processing used in the application are registered. For each application, Build steps to build based on registered dictionaries and , A processing step for adding, deleting, or changing a word in the registered dictionary, a deleting step for deleting a word in the dedicated dictionary, and a word after all the words registered in the dedicated dictionary are deleted A rebuild step to rebuild a dedicated dictionary based on a registered dictionary that has been added, deleted or modified It is characterized by including.
[0019]
The program recorded in the recording medium of the present invention is a dedicated dictionary dedicated to the application in which words to be used for language processing used in the application are registered. For each application, Building steps to build based on a registered dictionary where words are registered; , A processing step for adding, deleting, or changing a word in the registered dictionary, a deleting step for deleting a word in the dedicated dictionary, and a word after all the words registered in the dedicated dictionary are deleted A rebuild step to rebuild a dedicated dictionary based on a registered dictionary that has been added, deleted or modified The Let computer run It is characterized by that.
[0020]
The program of the present invention is a dedicated dictionary dedicated to the application in which words to be processed in the language used in the application are registered. For each application, Building steps to build based on a registered dictionary where words are registered; , A processing step for adding, deleting, or changing a word in the registered dictionary, a deleting step for deleting a word in the dedicated dictionary, and a word after all the words registered in the dedicated dictionary are deleted A rebuild step to rebuild a dedicated dictionary based on a registered dictionary that has been added, deleted or modified Is executed by a computer.
[0021]
In the present invention , A dedicated dictionary dedicated to the application that registers the language processing target words used in the application. Words are registered for each application Built on the registered dictionary , Processing to add, delete, or change words in the registered dictionary, delete words in the dedicated dictionary, delete all words registered in the dedicated dictionary, add, delete, Or a dedicated dictionary is rebuilt based on the modified registration dictionary .
[0022]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows a configuration example of a robot control system 1 to which the present invention is applied.
[0023]
In this robot control system 1, the speech recognition engine unit 11 recognizes input speech data and generates a word string corresponding to the speech data as a recognition result. The voice recognition engine unit 11 uses the recognition result as the name registration application unit 21. ₁ Chat application part 21 ₂ Application unit 21 for voice commander _Three ..., other application part 21 _M And to the application management unit 31.
[0024]
Name registration application section 21 ₁ Chat application part 21 ₂ Application unit 21 for voice commander _Three ..., other application part 21 _M Performs various processes based on the recognition result supplied from the speech recognition engine unit 11.
[0025]
Name registration application section 21 ₁ Is based on the recognition result supplied from the speech recognition engine unit 11, and registers the robot name, user name, etc. by voice, and the other application units are the name registration application unit 21. ₁ Is used to control the operation of the robot in response to the utterance from the user.
[0026]
Therefore, the chat application unit 21 ₂ Application unit 21 for voice commander _Three ,..., And other application units 21 _M The voice recognition performed by the name registration application unit 21 ₁ It is necessary to correspond to the robot name, user name, etc. registered in (1).
[0027]
Chat application part 21 ₂ Makes the robot chat with the user by voice and uses the voice commander application unit 21. _Three Causes the robot to perform an action corresponding to the utterance from the user. For example, the voice commander application unit 21 _Three Responds to an utterance from the user such as “SD (robot name), move forward!” To move the robot forward.
[0028]
An arbitrary number of application units can be prepared. Hereinafter, the name registration application unit 21 ₁ Chat application part 21 ₂ Application unit 21 for voice commander _Three ..., and other application sections 21 _M When it is not necessary to distinguish each of the above, they are collectively referred to as an application unit 21 as appropriate.
[0029]
The application management unit 31 instructs the application unit 21 to start and end based on the recognition result supplied from the speech recognition engine unit 11. For example, when the recognition result “start voice commander” is supplied from the voice recognition engine unit 11, the application management unit 31 uses the voice commander application unit 21. _Three Start up. At this time, a plurality of application units may be activated simultaneously.
[0030]
Further, the application unit 21 and the application management unit 31 issue a task switching command to the speech recognition engine unit 11, and tasks corresponding to the respective commands (described later in FIG. 2) are stored in the speech recognition engine unit 11. To enable (active) or disable (deactive).
[0031]
FIG. 2 shows the configuration of the speech recognition engine unit 11. The user's utterance is input to the microphone 51, and the microphone 51 converts the utterance into an audio signal as an electric signal. The microphone 51 supplies this audio signal to an AD (Analog Digital) conversion unit 52. The AD converter 52 samples the audio signal that is an analog signal from the microphone 51, quantizes it, and converts it into audio data that is a digital signal. This audio data is supplied to the feature amount extraction unit 43.
[0032]
The feature amount extraction unit 53 extracts, for example, feature parameters such as spectrum, power linear prediction count, cepstrum count, line spectrum pair, and the like from the audio data from the AD conversion unit 52 for each appropriate frame, and supplies them to the matching unit 54. To do.
[0033]
Based on the feature parameters from the feature quantity extraction unit 53, the matching unit 54 uses a phoneme typewriter task 71. ₁ Application switching task 71 ₂ , Name registration task 71 _Three , Chat Task 71 _Four Voice commander task 71 _Five ... and other tasks 71 _N Among them, for each task enabled at that time, a word string closest to the voice (input voice) input to the microphone 51 is obtained as a recognition result while referring to a database inside the task as necessary. . The matching unit 54 supplies the recognition result to the application unit 21 and the application management unit 31 corresponding to each task.
[0034]
A task is a set of data necessary for voice recognition. That is, the speech recognition engine unit 11 is a program part for performing matching and the like, and a data part when classified into a data part such as an acoustic model, a language model, a recognition dictionary, and a program for accessing data. .
[0035]
Therefore, even when a plurality of applications perform speech recognition using different acoustic models, language models, and dictionaries, a single speech recognition engine unit can be provided by preparing a plurality of tasks. Details of the inside of the task will be described later with reference to FIG.
[0036]
Phonological typewriter task 71 ₁ Is a task that works as a phoneme typewriter, and is enabled by a command from the speech recognition engine unit 11. Using this phoneme typewriter, the matching unit 54 acquires a phoneme sequence for an input arbitrary voice, and also acquires pronunciation in kana notation. For example, “k / i / m / i / n / o / n / a / m / a / e / w / a / e / s / u / d” says “Your name is SDR” / i: / a: / r / u / d / a / y / o ”(where“ i: ”and“ a: ”are the long sounds of“ i ”and“ a ”, respectively) Acquired the Kana notation of “WSDDR Dayo”. The phoneme series and kana notation are used by the unknown word acquisition unit 56.
[0037]
Application switching task 71 ₂ Is a task corresponding to the application management unit 31 and is validated when a task switching command is supplied from the application management unit 31 after the application management unit 31 is activated. Application switching task 71 ₂ Thus, for example, the matching unit 54 recognizes the voice corresponding to the activation of the application unit such as “launch chat app”, “launch voice commander”, “launch name registration”, or an end command.
[0038]
Name registration task 71 _Three Name registration application unit 21 ₁ In response to a command from the application management unit 31, the name registration application unit 21 ₁ Is started, the name registration application unit 21 ₁ When a task switching command is supplied from, it is validated. Name registration task 71 _Three Thus, the matching unit 54, for example, “Your name is <an unknown word representing a robot name>” and “My name is <an unknown word representing a person name>”. Recognize
[0039]
Chat task 71 _Four Voice commander task 71 _Five ... and other tasks 71 _N Respectively, the chat application part 21 ₂ Application unit 21 for voice commander _Three ..., other application part 21 _M When a task switching command is supplied from the corresponding application unit after the corresponding application unit is activated by a command from the application management unit 31, the task is validated.
[0040]
The matching unit 54 includes a chat task 71. _Four Thus, for example, it is possible to recognize an utterance as a chat from the user “SD (robot name), what time did it happen?”. The matching unit 54 also includes a voice commander task 71. _Five Thus, for example, it is possible to recognize an utterance as a command from the user “SD (robot name), move forward one step”.
[0041]
Moreover, the matching part 54 reflects the word registered in the common dictionary part 55 mentioned later on each task.
[0042]
In the following, the phonetic typewriter task 71 ₁ Application switching task 71 ₂ , Name registration task 71 _Three , Chat Task 71 _Four Voice commander task 71 _Five ... and other tasks 71 _N When it is not necessary to distinguish each of these individually, they are collectively referred to as task 71 as appropriate.
[0043]
The common dictionary unit 55 stores a common dictionary as a word dictionary commonly used in the task 71. In the common dictionary stored in the common dictionary section 55, pronunciation information and category information are described for all the words registered there. For example, when “SDR (robot name)” which is a proper noun is registered in the common dictionary, the pronunciation (phonological information) “SDR” and the category “_robot name_” are described in the common dictionary. Details will be described later with reference to FIG.
[0044]
The unknown word acquisition unit 56 performs a phonological typewriter task 71 for words (unknown words) such as names that are not registered in a recognition dictionary (a fixed word dictionary 131 described later with reference to FIG. 6). ₁ Is stored and the phoneme sequence and kana notation supplied from the matching unit 54 are stored, and thereafter, the speech of the word can be recognized (identified from other speech).
[0045]
That is, the unknown word acquisition unit 56 performs the phoneme typewriter task 71. ₁ The phoneme sequence and kana notation of the unknown word recognized by is classified into several clusters. Each cluster has an ID, a representative phoneme series, and a representative kana notation, and is managed by the ID.
[0046]
FIG. 3 shows the state of the cluster of the unknown word acquisition unit 56.
[0047]
When three voices of “Aka”, “Ao”, and “Midori” are input, the unknown word acquisition unit 56 converts the three input voices into “Aka” cluster 91 and “Ao” cluster corresponding to each of them. 92 and “Midori” cluster 93, and each cluster has a representative phoneme sequence (“a / k / a”, “a / o”, “m / i / d / o / r / i ”), typical kana notation (in the case of FIG. 3,“ red ”,“ ao ”,“ midori ”), and ID (in the case of FIG. 3,“ 1 ”,“ 2 ”,“ 3 ”).
[0048]
Here, when the voice “Aka” is input again, since the corresponding cluster already exists, the unknown word acquisition unit 56 classifies the input voice into the “Aka” cluster 91 and does not generate a new cluster. On the other hand, when the voice “Kuro” is input, the corresponding cluster does not exist, so the unknown word acquisition unit 56 newly generates the “Kuro” cluster 94 corresponding to “Kuro”. In addition, a representative phoneme sequence (“k / u / r / o” in the example of FIG. 3), a typical kana notation (“black” in the example of FIG. 3), and an ID (in FIG. 3) In the example, “4”) is added.
[0049]
When this method is used, the accuracy of the representative phoneme sequence and the representative kana pronunciation of each cluster can be improved by the user inputting the same voice many times. For example, when “Midori” is entered once, the phoneme typewriter misrecognizes and outputs the phonetic sequence “m / e / r / a / a” and the kana pronunciation “Mera”. . After that, the phonological sequence and kana pronunciation may converge to the correct values (“m / i / d / o / r / i” and “midori”) by repeating the utterance “Midori” many times. There is. Details of such word acquisition processing are disclosed in Japanese Patent Application No. 2001-097843 and Japanese Patent Application No. 2001-382579 previously proposed by the present applicant.
[0050]
Next, robot control processing in the robot control system 1 of FIG. 1 will be described with reference to FIGS. This process is started when the robot control system 1 is activated by the user.
[0051]
In step S1, the speech recognition engine unit 11 is activated, and the process proceeds to step S2. In step S2, the speech recognition engine unit 11 stores the contents (common dictionary) of the common dictionary unit 55 stored in a storage unit (not shown) at the end of the previous robot control system 1 (processing in step S17 described later). And the state of the cluster of the unknown word acquisition unit 56 is loaded. When the common dictionary and the cluster state are not stored in the storage unit, the common dictionary unit 55 and the unknown word acquisition unit 56 are left without any cluster entry. If the cluster state is stored in the storage unit but the common dictionary state is not stored, only the common dictionary is initialized (there is no entry). Conversely, if the common dictionary state is stored but the cluster state is not stored, the cluster-derived entry (the entry whose cluster ID is described in FIG. 24) is deleted from the common dictionary, and the kana pronunciation The entry of origin (the entry in which kana pronunciation is described in FIG. 24) remains.
[0052]
After the process of step S2, the process proceeds to step S3, where the speech recognition engine unit 11 performs the phoneme typewriter task 71. ₁ Enables phonetic typewriter task 71 ₁ Can be used for voice recognition, and the process proceeds to step S4. In step S4, the application management unit 31 is activated and proceeds to step S5.
[0053]
In step S5, the application management unit 31 performs a corresponding task 71 for application switching. ₂ Is enabled, and the process proceeds to step S6. In step S <b> 6, the speech recognition engine unit 11 recognizes an activation command for the application unit 21 input by voice to the microphone 51 and supplies the recognition result to the application management unit 31. Details of the voice recognition processing will be described later with reference to the flowchart of FIG.
[0054]
After the processing in step S6, the process proceeds to step S7 in FIG. ₁ Whether or not to start the name registration application unit 21 ₁ If it is determined to activate (for example, the recognition result is “activate name registration”), the process proceeds to step S8.
[0055]
In step S <b> 8, the application management unit 31 performs the name registration application unit 21. ₁ Start up. After the processing of step S8, the process proceeds to step S9, where the name registration application unit 21 ₁ Performs name registration processing. Details of the name registration processing will be described later with reference to the flowchart of FIG.
[0056]
In step S <b> 7, the application management unit 31 performs the name registration application unit 21. ₁ Is determined not to be activated, the process proceeds to step S10, and the chat application unit 21 is determined from the recognition result by the speech recognition engine unit 11. ₂ It is determined whether or not to start. In step S <b> 10, the application management unit 31 performs the chat application unit 21. ₂ (For example, when the recognition result is “activate chat”), the process proceeds to step S11 and the chat application unit 21 is activated. ₂ Start up.
[0057]
After the processing of step S11, the process proceeds to step S12, and the chat application unit 21 ₂ Performs chat processing. Details of the chat process will be described later with reference to the flowchart of FIG.
[0058]
In step S <b> 10, the application management unit 31 performs the chat application unit 21. ₂ Is determined not to be activated, the process proceeds to step S13, and the speech commander application unit 21 is determined from the recognition result by the speech recognition engine unit 11. _Three It is determined whether or not to start. In the process of step S13, the application management unit 31 performs the voice commander application unit 21. _Three (For example, when the recognition result is “voice commander activation”), the process proceeds to step S14, where the voice commander application unit 21 is activated. _Three Start up.
[0059]
After the process of step S14, the process proceeds to step S15, where the voice commander application unit 21 is processed. _Three Performs voice commander processing. Details of the voice commander process will be described later with reference to the flowchart of FIG.
[0060]
In step S <b> 13, the application management unit 31 performs the voice commander application unit 21. _Three 4, since the recognition result by the speech recognition engine unit 11 is incorrect (there may be an utterance other than application switching), the process returns to step S6 in FIG. A process for recognizing the input voice is performed.
[0061]
Thus, the application management unit 31 activates the application unit 21 according to the recognition result by the speech recognition engine unit 11.
[0062]
After the processes of steps S9, S12, and S15, the process proceeds to step S16, and the application management unit 31 determines whether or not to end the robot control process. For example, the application management unit 31 determines whether or not an end button (not illustrated) is pressed by the user, and determines that the robot control process is ended when the end button is pressed.
[0063]
If it is determined in step S16 that the robot control process is not to be terminated, the process returns to step S6 in FIG. 4 to repeat the process of recognizing the input voice. In step S16, if the application management unit 31 determines that the robot control process is to be ended (the end button has been pressed), the process proceeds to step S17, where the common dictionary of the common dictionary unit 55 and the cluster of the unknown word acquisition unit 56 are clustered. The state is stored in a storage unit (not shown).
[0064]
Then, when there is an active application unit 21, the application management unit 31 ends the application unit. At this time, the application unit 21 invalidates the corresponding task 71. The application management unit 31 also includes an application switching task 71. ₂ The speech recognition engine unit 11 disables the phoneme typewriter task 71. ₁ Is invalidated, and the application management unit 31 and the speech recognition engine unit 11 end the processing.
[0065]
In the above-described processing, the application unit performs the name registration application unit 21. ₁ Chat application part 21 ₂ Application unit 21 for voice commander _Three However, if there is another application part, if it is determined in step S13 that the voice commander application is not activated, the process does not return to step S6 but is similar to steps S7, S10, and S13. Whether or not to start another application is determined, and the other application is started according to the determination result.
[0066]
In the above-described processing, the end of the voice recognition is instructed by the user, but the robot control system 1 may automatically determine, for example, that the voice recognition ends when no voice is input for a predetermined time.
[0067]
According to the above-described processing, the application switching task 71 ₂ Is valid even while each application unit 21 is running, so that even if the utterance “Launch XX” is made while another application unit is running, it recognizes the utterance and responds accordingly. An application can be started. For example, the voice commander application unit 21 _Three Is activated, if the user utters “activate chat”, the chat application unit 21 ₂ Can be launched.
[0068]
In this case, start the new application part after closing the active application part, or start the new application part after the active application part is paused, Whether to restart the application unit or to start both in parallel is set in advance by a combination of the application units (may be dynamically determined from resource constraints such as memory).
[0069]
FIG. 6 shows the configuration of the task 71. The task 71 includes an acoustic model 111, a language model 112, a dictionary 113, a phoneme list 114, a kana phoneme conversion rule 115, and a search parameter 116.
[0070]
The acoustic model 111 stores a model representing acoustic features such as individual phonemes and syllables of speech to be recognized. As the acoustic model, for example, an HMM (Hidden Markov Model) can be used.
[0071]
The language model 112 describes information (hereinafter referred to as chain information as appropriate) indicating how words registered in the word dictionary of the dictionary 113 are linked (connected). The description method includes statistical word chain probability (n-gram), generation grammar, finite state automan, and the like.
[0072]
The language model 112 includes chain information about a category in which words are classified from a specific viewpoint, in addition to the chain information about words. For example, when “category consisting of a word representing a user name” is represented by a symbol “_user name_” and “category consisting of a word representing a robot name” is represented by a symbol “_robot name_”, the language model 112 is , “_User name_”, “_robot name_” chain information (chain of categories, chain of categories and words stored in the dictionary in advance) is also described.
[0073]
Accordingly, chain information can be acquired for words that are not included in the language model 112. For example, when the chain information of “SD” and “ha (particle)” is acquired, even if the chain information about “SD” is not described in the language model 112, “SD” is “_robot name”. If it is known that it belongs to the category represented by the symbol “_”, instead of acquiring the linkage information of “_Robot name_” and “ha”, the chain of “SD” and “ha” is obtained. Information can be acquired.
[0074]
The category is not a classification based on semantic attributes (“_robot name_”, “_user name_”, “_place name_”, “_store name_”, etc.), but a classification based on part of speech (“_noun_ ”,“ _Verb_ ”,“ _particle_ ”, etc.). Hereinafter, the notation “_... _” Represents a category name.
[0075]
The dictionary 113 includes a fixed word dictionary 131, a variable word dictionary 132, and a category table 133.
[0076]
In the fixed word dictionary 131, words that are not subject to word registration and deletion, that is, pronunciations (phoneme series), phonemes, and words set in advance in the robot control system 1 (hereinafter referred to as “fixed words” as appropriate) are stored. Various types of information such as a model describing the syllable chain relationship are described.
[0077]
In the fixed word dictionary 131, for each task 71, information about a dedicated word used in the application unit 21 corresponding to the task 71 is described. The same applies to the acoustic model 111 and the language model 112 described above, and to the later-described category table 133, phoneme list 114, kana phoneme conversion rule 115, and search parameter 116.
[0078]
The variable word dictionary 132 describes various types of information such as a model describing a word to be registered and deleted, that is, a model describing the pronunciation, phoneme, and syllable chain relationship for the registered word. When a new registered word is registered, the registered word is reflected. This reflection process will be described later with reference to FIG. The deletion of words and the change of pronunciation can be performed only for the entries of the variable word dictionary 132. The variable word dictionary 132 does not have to be recorded in anything.
[0079]
The category table 133 stores a table indicating the correspondence between categories included in the language model 112 and word information included in the categories. Further, when the task 71 assigns a category-specific ID (category ID), the category table 133 also stores the correspondence between the category symbol and the ID. For example, when category ID “4” is assigned to the category “_robot name_”, category ID = 4 is also stored corresponding to “_robot name_”. The category table 133 stores nothing when the language model 112 does not include a category.
[0080]
The phoneme list 114 is a list of phoneme symbols used in the task 71. The kana phoneme conversion rule 115 is a rule for converting a kana character string into a phoneme sequence. Thus, by storing the kana phoneme conversion rule 115 for each task, the common dictionary unit 55 can hold a kana character string that is independent of the phoneme sequence as pronunciation information.
[0081]
The search parameter 116 holds a parameter used when the matching unit 54 performs matching (search). The parameters include a value depending on the acoustic model 111, a value depending on the number of vocabularies, a value depending on the type of the language model 112, and the like, and thus must be retained for each task. However, the task-independent parameters may be held in common in the recognition engine unit 11.
[0082]
In the above description, all data is stored for each task. However, the data used in common for a plurality of tasks can be shared between the tasks to reduce the memory usage. For example, if the phoneme list 114 is common to all tasks, only one phoneme list 114 may be prepared in the speech recognition engine unit 11, and each task may refer to it. In this case, it is sufficient to prepare only one kana phoneme conversion rule 115.
[0083]
There are two types of acoustic models 111 for quiet environments (acoustic models with high recognition rates in quiet environments) and noise models (acoustic models with reasonable recognition rates in noisy environments). One of them may be referred to every time.
[0084]
For example, name registration task 71 _Three And chat task 71 _Four Are assumed to be used in a quiet environment, so the audio commander task 71 is referred to by referring to the acoustic model 111 for a quiet environment. _Five Is assumed to be used in a noisy environment (an environment in which the robot's operation sound is high), so an acoustic model for a noisy environment can be referred to.
[0085]
FIG. 7 shows an example of the phoneme list 114 of FIG. In FIG. 7, one symbol represents one phoneme (corresponding to). In the phoneme list 114 of FIG. 7, a vowel + colon (for example, “a:”) represents a long sound, and “N” represents a repellent sound (“n”). "Sp", "silB", "silE", and "q" all represent silence, but "silence in utterance", "silence before utterance", "silence after utterance", "promotion sound"("")"".
[0086]
FIG. 8 shows an example of the kana phoneme conversion rule 115 of FIG. According to the kana phoneme conversion rule 115 in FIG. 8, for example, a kana character string “SD” is converted into a phoneme sequence “e / s / u / d / i: / a: / r / u”. .
[0087]
Next, examples of the language model 112 and the dictionary 113 (FIG. 6) for each task are shown.
[0088]
FIG. 9 shows a phonetic typewriter task 71. ₁ An example of the language model 112 (FIG. 6) is shown. In FIG. 9, the variable “$ SYLLABLE” on the first line means any one of the kana notations because all kana notations are connected by “|” meaning “or”. .
[0089]
That is, here, the phoneme typewriter task 71 ₁ 9 is a task for speech recognition in units of syllables, the language model 112 in FIG. 9 has a chain rule that any syllable can be arbitrarily connected in the BNF (Backus-Naur-Form) format. This is expressed in grammar. The language model 112 may be a statistical language model described later.
[0090]
FIG. 10 shows a phonological typewriter task 71. ₁ An example of the fixed word dictionary 131 (FIG. 6) is shown. “Symbol” is a character string for identifying a word, and for example, kana notation can be used. Entries with the same symbol are considered to be entries with the same word. The language model 112 is represented using this symbol. In addition, "<First> and `` “<Termination>” is a special symbol, and represents “silence before utterance” and “silence after utterance”, respectively (the same applies to FIG. 11 described later).
[0091]
“Transcription” represents a notation of a word, and a character string output as a recognition result is this transcription. The “phoneme sequence” is a phoneme sequence representing the pronunciation of a word.
[0092]
Phonological typewriter task 71 ₁ In the variable word dictionary 132, a phoneme typewriter task 71 is stored. ₁ Since it is not supposed to add a word to, nothing is remembered. Also, the phoneme typewriter task 71 ₁ As shown in FIG. 9, the language model 112 does not include a category, so nothing is stored in the category table 133.
[0093]
FIG. 11 shows an application switching task 71. ₂ An example of the language model 112 (FIG. 6) is shown. The language model 112 in FIG. 11 is described in a BNF format grammar. The variable “$ APPLICATIONS” on the first line is connected to all application names (“chat”, “voice commander”, “name registration”, etc.) with “|” meaning “or”. Means one of the names.
[0094]
In addition, the variable “$ UTTERANCE” on the second line has “[]” meaning “can be omitted” added to “_robot name_” and “wo” respectively. ) Start application name () "means. Here, “robot name” indicates a word registered in the category “_robot name_”.
[0095]
For example, if “SDR” is registered in “_Robot name_”, utterances such as “Start SDR, voice commander ()”, “Start voice commander ()”, etc. 11 language models 112 are recognized.
[0096]
Thus, by describing the language model 112 using the category name, even if the word is newly registered, the word is included in the category described in the language model 112. Can recognize an utterance including the newly registered word by using the language model 112.
[0097]
FIG. 12 shows an application switching task 71 ₂ An example of the fixed word dictionary 131 (FIG. 6) is shown. In the fixed word dictionary 131 in FIG. 12, transcription and phoneme series are described for symbols (such as “chat” and “speech commander” in FIG. 11) described in the grammar of the language model 112 in FIG. Yes.
[0098]
FIG. 13 shows an application switching task 71. ₂ An example of the category table 133 (FIG. 6) is shown. The category table 133 stores information on the types of categories used in the language model 112 and the words belonging to the categories. When the language model 112 is as shown in FIG. ₂ Since the language model 112 uses the category “_robot name_”, “_robot name_” is entered in the category table 133 as shown in FIG. In FIG. 13, the set of words belonging to the category “_Robot name_” is an empty set, indicating that there are no words belonging to “_Robot name_” yet.
[0099]
As shown in FIG. 13, even when a category is entered in the category table 133, if there is no word belonging to the entry (in the case of an empty set), the variable word dictionary 132 contains the category. Information on the belonging word is not stored.
[0100]
FIG. 14 shows a name registration task 71. _Three An example of the language model 112 (FIG. 6) is shown. The language model 112 in FIG. 14 is described in a BNF format grammar. The variable “$ UTTERANCE” is “I [name] is <OOV> [is] [named]” and “You [name] is <OOV> [name]]” or “or” “|”, Which means “,” and “[]”, which means “can be omitted”, are added to “name”, “is”, “say”, and “to say”. ing.
[0101]
Therefore, using the language model 112 in FIG. 14, “I (name) is <OOV> (is)” (or “you” is <OOV>). Be recognized. Note that <OOV> is a symbol meaning “Out Of Vocabulary”, and means an arbitrary pronunciation phrase (word not described in the fixed word dictionary 131).
[0102]
By using the symbol <OOV>, for example, “My name is Taro” and “Your name is SDR” (“Taro” and “SDR” are described in the fixed word dictionary 131. 14), “<First> My name is <OOV><Terminal>” and “<First> Your name is <OOV>” in the language model 112 of FIG. 14 respectively. By doing this, you can get voice recognition results such as “My name is Taro” and “Your name is SDR”.
[0103]
FIG. 15 shows a name registration task 71. _Three An example of the fixed word dictionary 131 (FIG. 6) is shown. In the fixed word dictionary 131, transcriptions and phoneme sequences are described for symbols described in the grammar of the language model 112 as shown in FIG.
[0104]
Name registration task 71 _Three Here, the variable word dictionary 132 includes a name registration task 71. _Three Since it is not supposed to add a word to, nothing is remembered. Also, a name registration task 71 _Three As shown in FIG. 14, the language model 112 does not include a category, so nothing is stored in the category table 133.
[0105]
FIG. 16 shows a chat task 71. _Four An example of the language model 112 (FIG. 6) is shown. Since the chat has many vocabulary and utterance variations, a statistical language model is used as the language model 112. The statistical language model is a model in which word chain information is described with conditional probabilities. In the language model 112 in FIG. 16, a tri-type representing the sequence of three words 1, 2, 3; Gram is used.
[0106]
In FIG. 16, “P (word 3 | word 1 word 2)” is a probability that “word 3” appears next when “word 1” and “word 2” are arranged in the word string. Represents. For example, when there is a sequence of “<start>“ _robot name_ ””, the probability of “ha” appearing next is “0.012”. This probability is obtained in advance by analyzing a text describing a large amount of chat. In addition to the tri-gram, a bi-gram (probability of two chains), a uni-gram (word appearance probability), and the like can be used as the language model 112 as necessary.
[0107]
In the language model 112 of FIG. 16 as well, the grammar is described using categories in addition to words, as in the case of FIG. That is, in FIG. 16, “_robot name_” and “_place name_” mean the categories “_robot name_” and “_place name_”, and the tri-gram is described using these categories. Thus, when a word representing a robot name or a place name is registered in the variable word dictionary 132, the word is added to the chat task 71. _Four Can be recognized.
[0108]
FIG. 17 shows a chat task 71. _Four An example of the fixed word dictionary 131 is shown. In the fixed word dictionary 131, transcriptions and phoneme sequences are described for symbols described in the grammar of the language model 112 as shown in FIG.
[0109]
FIG. 18 shows a chat task 71. _Four An example of the category table 133 is shown. The category table 133 stores information on the types of categories used in the language model 112 and the words belonging to the categories. When the language model 112 is as shown in FIG. _Four In the language model 112, two categories of “_robot name_” and “_place name_” are used. Therefore, the category table 133 includes “_robot name_” as shown in FIG. And two categories "_place name_" are entered. In FIG. 18, the words belonging to the categories “_robot name_” and “_place name_” indicate that there is nothing yet.
[0110]
FIG. 19 shows a voice commander task 71. _Five An example of the language model 112 (FIG. 6) is shown. The language model 112 in FIG. 19 is described in a BNF format grammar. The variable “$ NUMBER” in the first line is connected to the number (“1”, “2”, “3”, etc.) by “|”, which means “or”. Means one.
[0111]
The variable “$ DIRECTION” on the second line is connected to the direction (“front”, “back”, “right”, “left”, etc.) with “|” meaning “or”. Means one of the following. The variable “UTTERANCE” on the third line is “_Robot name_”, “$ DIRECTION”, and “$ NUMBER steps” with “advance” added, and the variable “$ UTTERANCE” “[Robot name_]”, “to $ DIRECTION”, and “$ NUMBER steps” are appended with “[]” meaning “can be omitted”.
[0112]
Accordingly, in the language model 112 of FIG. 19, for example, a voice “3 steps ahead of (robot name)” is recognized.
[0113]
FIG. 20 shows a voice commander task 71. _Five An example of the fixed word dictionary 131 is shown. In the fixed word dictionary 131, transcriptions and phoneme sequences are described for symbols described in the grammar of the language model 112 as shown in FIG.
[0114]
Note that the symbols for “1” and “Walk” are duplicated, but this means that “1” and “Walk” have two pronunciations (“Ichi” and “It”, “Ho” and “Po”, respectively). )). As a result, for example, utterances with different pronunciations “Ichiho” and “Ippo” can be recognized as the same “1 step”.
[0115]
When the language model 112 is as shown in FIG. _Five Since only the category “_robot name_” is used in the language model 112 of FIG. _Five In the category table 133, the application switching task 71 shown in FIG. ₂ This is the same as the category table 133. When the word belonging to “_robot name_” has not been spoken yet, the voice commander task 71 _Five Nothing is stored in the variable word dictionary 132.
[0116]
Next, the name registration application unit 21 ₁ However, the name registration process performed in step S9 of FIG. 5 will be described in detail with reference to the flowchart of FIG. This process is performed by the name registration application unit 21 according to the user's speech. ₁ Started when is started. Before this process is started, the user can enter a name by, for example, a voice input mode for inputting a name by voice or a kana input by a keyboard or the like as a name registration mode for registering a name by a mode switching button (not shown). Select either Kana input mode to input.
[0117]
In step S41, the name registration application unit 21 ₁ Is the name registration task 71 of the speech recognition engine unit 11. _Three And this name registration task 71 _Three To be able to recognize voice.
[0118]
After the process of step S41, the process proceeds to step S42, and the name registration application unit 21 ₁ Determines whether the name registration mode is the voice input mode. If it is determined that the name registration mode is the voice input mode, the process proceeds to step S43, the matching unit 54 performs name recognition processing, and the process proceeds to step S44. (Alternatively, if the user speaks in step S42, it is determined that “name is input by voice” and the process proceeds to step S43. If a not-illustrated kana input button is pressed, “name is input in kana characters” The process proceeds to step S46.) Details of this name recognition process will be described later with reference to FIG.
[0119]
In step S44, the name registration application unit 21 ₁ Determines whether the name speech recognition result (recognized name) obtained by performing the name recognition processing in step S43 in the matching unit 54 is correct. This determination is made, for example, by speaking the recognition result toward the user and determining whether or not an OK button (not shown) has been operated by the user.
[0120]
If it is determined in step S44 that the name speech recognition result is not correct, the user is prompted to speak again, and the process returns to step S43 to perform the name recognition process again. If it is determined in step S44 that the recognition result is correct, the process proceeds to step S47.
[0121]
On the other hand, in step S42, the name registration application unit 21 ₁ If it is determined that the name registration mode is not the voice input mode, the process proceeds to step S45 to determine whether or not the name registration mode is the kana input mode.
[0122]
If it is determined in step S45 that the name registration mode is not the kana input mode, the name registration mode is not selected by the user. Therefore, the process waits until the name registration mode is selected, and the user selects the name input mode. Wait and return to step S42.
[0123]
If it is determined in step S45 that the name registration mode is the kana input mode, the process proceeds to step S46, and the name registration application unit 21. ₁ Obtains a kana string of names input by the user and a category of the name.
[0124]
As a method of inputting a kana string, for example, a method in which a user temporarily connects a keyboard to input kana characters, a method of inputting using various switches of the robot, a paper on which characters are written, etc. are shown to the robot. Character recognition (for example, see Japanese Patent Application No. 2001-135423), wireless LAN (Local Area Network) etc., connecting a robot and a personal computer, transferring from the personal computer to the robot, via the Internet, etc. There are methods for downloading to a robot. Also, in the method of recognizing characters by showing a paper or the like on which a character is written to a robot, a character string mixed with kana-kanji characters is input instead of inputting kana characters, and the name registration application unit 21 ₁ However, it may be converted into a kana string (see Japanese Patent Application 2001-135423).
[0125]
Furthermore, the user does not input the name string, but an entry with the name kana character added to the common dictionary of the common dictionary section 55 in advance, and the name registration application section 21. ₁ May acquire a kana string of names by referring to the common dictionary unit 55.
[0126]
After the process of step S44 or S46, the process proceeds to step S47, and the name registration application unit 21 ₁ Determines the category of the name to be registered. When the name registration mode is the kana input mode, the name registration application unit 21 ₁ Determines the category (input by the user) acquired in step S46 as the category of the name to be registered.
[0127]
In other words, in the Kana input mode, in step S46, the user inputs not only the name but also the category of the name, and the category of the name input by the user is determined as the name category to be registered. On the other hand, when the name registration mode is the voice input mode, the name registration application unit 21. ₁ Determines the category of the name obtained by the name recognition process in step S43.
[0128]
For example, when the recognition result supplied from the speech recognition engine unit 11 starts with “Kimi”, the category to which the registered name belongs is assumed to be “_robot name_”, and when it starts with “I”, It is assumed that the category to which the registered name belongs is “_user name_”. Various category estimation methods disclosed in Japanese Patent Application No. 2001-382579 previously proposed by the present applicant can also be used.
[0129]
After the process of step S47, the process proceeds to step S48, and the name registration application unit 21 ₁ Controls the matching unit 54 to enter the pronunciation information and category of the name to be registered in the common dictionary of the common dictionary unit 55, and proceeds to step S49. In step S49, the name registration application unit 21 ₁ Controls the matching unit 54 to change the contents of the common dictionary into the chat task 71. _Four Voice commander task 71 _Five ... and other tasks 71 _N To reflect. Details of this reflection will be described later with reference to the flowchart of FIG.
[0130]
In this way, by reflecting the name registered in the common dictionary on other tasks, the registered name can be recognized also by other tasks.
[0131]
After the processing of step S49, the process proceeds to step S50, and the name registration application unit 21. ₁ Determines whether to end the name registration process. This determination is made, for example, depending on whether or not an OK button (not shown) has been operated (pressed) by speaking to the user a question of whether to end or not. If it is determined in step S49 that the name registration application is not terminated (for example, the OK button has not been pressed), the process returns to step S42 to perform processing for registering another name.
[0132]
If it is determined in step S50 that the name registration process is to be terminated (for example, an OK button has been pressed), the process proceeds to step S51, and the name registration application unit 21. ₁ Is a name registration task 71 _Three Is disabled, and the process proceeds to step S52. In step S52, the name registration application unit 21 ₁ Ends the process.
[0133]
FIG. 22 is a flowchart illustrating the name recognition process performed by the matching unit 54 in FIG. 2 in step S43 in FIG.
[0134]
In step S61, the matching unit 54 determines whether or not sound is input to the microphone 51. If it is determined that no sound is input, the matching unit 54 waits until the sound is input. If it is determined in step S61 that sound has been input, the process proceeds to step S62. The voice input here may be a normal conversation such as “My name is Taro” or “Your name is SDR”, and the user is conscious of name registration, “Taro” “ It is not necessary to enter the name “SD” alone.
[0135]
In step S62, the matching unit 54 recognizes the voice and extracts the name. For example, when an utterance “Your name is SDR” is made, a name registration task 71 having a language model 112 as shown in FIG. 14 and a fixed word dictionary 131 as shown in FIG. _Three Referring to FIG. 5, the matching unit 54 generates, for example, a recognition result “<head> Your name is <OOV><end>”. In addition, the matching unit 54 obtains information on which section of the utterance <OOV> is (from what second to how many seconds of the first utterance).
[0136]
Further, the matching unit 54 performs a phonological typewriter task 71 having a language model 112 as shown in FIG. 9 and a fixed word dictionary 131 as shown in FIG. 10 for the same utterance. ₁ For example, “k / i / m / i / n / o / n / a / m / a / e / w / a / e / s / u / d / i: / a: / r / Obtain the phoneme sequence “u / d / a / y / o” and the kana sequence “Kiminonamaewa SDL Ayo”.
[0137]
Then, based on the information on which section of the utterance <OOV> is the matching unit 54, from the obtained phoneme sequence and kana sequence, the section corresponding to <OOV>, that is, the phoneme sequence of the name section A kana string is cut out to obtain a phoneme sequence “e / s / u / d / i: / a: / r / u” and a kana string “SDR”. The matching unit 54 also obtains audio data for the same section. Details of the process for extracting the name are disclosed in Japanese Patent Application No. 2001-382579 previously proposed by the present applicant.
[0138]
After the process in step S62, the process proceeds to step S63, and the matching unit 54 supplies the phoneme sequence, kana string, and speech data of the name extracted in the process of step S62 to the unknown word acquisition unit 56 to perform clustering. Details of clustering are disclosed in Japanese Patent Application No. 2001-097843 previously proposed by the present applicant. As a result of this clustering, each cluster of the unknown word acquisition unit 56 has a representative phoneme sequence and a kana string.
[0139]
After the process of step S63, the process proceeds to step S64, and the recognition result of the voice recognized in step S62 (for example, the kana string “Kinominama ESD Ryo”) is used as the name registration application unit 21. ₁ To supply.
[0140]
FIG. 23 shows an example of the feature space clustered by the unknown word acquisition unit 56 in the process of step S63 of FIG. In FIG. 23, a feature space defined by two feature amounts (feature parameters) 1 and 2 is shown in order to avoid making the diagram complicated (the same applies to FIG. 3 described above). In FIG. 23, four names “Arara”, “Saga”, “Tokyo”, and “Taro” are clustered in the feature space.
[0141]
That is, in FIG. 23, four clusters of “arara” cluster 151, “sato” cluster 152, “tokyo” cluster 153, and “taro” cluster 154 are configured in the feature space. Each cluster has a representative phoneme sequence (“a / r / a / r / a”, “s / a / n / i:”, “t / o: / ky / o in the example of FIG. 23). : ”,“ T / a / r / o / u ”), typical Kana notation (in the example of FIG. 23,“ Arara ”,“ Sunny ”,“ Tokyo ”,“ Taro ”)) and ID (Fig. In the case of 23 examples, “1”, “2”, “3”, “5”) are added.
[0142]
FIG. 24 shows an example of a common dictionary of the common dictionary unit 55 in which word information is entered in step S48 of FIG. In FIG. 24, the entry in the first row indicates that the pronunciation is input in a kana string, the pronunciation is a character string “SDR”, and the category “_robot name_” is input. .
[0143]
The entry in the second line is a representative kana notation in which pronunciation is input by voice, the kana notation of the pronunciation and the phoneme sequence are added to the cluster whose ID of the unknown word acquisition unit 56 is “5” (FIG. 23). In this example, “Taro”) and a phoneme sequence (“t / a / r / o:” in the example of FIG. 23) are indicated. The category of the entry in the second row is determined by the name registration application unit 21 in step S47 of FIG. 21, and is “_user name_”. For example, when the user utters “My name is Taro”, an entry such as the second line is configured in the common dictionary unit 55.
[0144]
Similarly, in the entries on the third and fourth lines, the pronunciation is input in a kana string, the pronunciations are character strings “Sunny Taro” and “Kitashinagawa”, respectively, and the category is “_user name_”. This indicates that “_place name_” has been entered. The entry on the fifth line is a representative kana notation (the pronunciation is input by voice, the kana notation of the pronunciation and the phoneme sequence are added to the cluster whose ID of the unknown word acquisition unit 56 is “3” ( In the example of FIG. 23, “Tokyo”) and a phoneme sequence (“t / o: / ky / o:” in the example of FIG. 23) are shown. Further, the category of the entry in the fifth row is determined to be “_place name_” by the name registration application unit 21.
[0145]
In the common dictionary, for a word whose pronunciation is input in a kana string, a pair of a kana string and a category representing the pronunciation of the word is registered in one entry, and for a word whose pronunciation is spoken, A set of ID and category representing the cluster of words is registered in one entry.
[0146]
FIG. 25 is a flowchart for explaining the process in which the matching unit 54 reflects the contents of the common dictionary unit 55 in the task in the process of step S49 of FIG. This process is performed for each activated task.
[0147]
In step S81, the matching unit 54 initializes the variable word dictionary 132 and the category table 133 in the task 71 (FIG. 6). That is, the variable word dictionary 132 is in a state where there is no entry, and the category table 133 is in a state where no word belongs to each category.
[0148]
After the process of step S81, the process proceeds to step S82, and the matching unit 54 reflects the contents of the common dictionary unit 55 in the variable word dictionary 132 and the category table 133.
[0149]
That is, the matching unit 54 selects, from the common dictionary of the common dictionary unit 55, a category that is common (identical) with the category entered in the category table 133, and the category and the cluster ID corresponding to the category. Or acquire kana pronunciation (kana string). Furthermore, the matching unit 54 acquires a kana string corresponding to the cluster ID from the unknown word acquisition unit 56 when acquiring the cluster ID from the common dictionary.
[0150]
When the matching unit 54 acquires a kana string of words belonging to the category selected from the common dictionary of the common dictionary unit 55 as described above, the matching unit 54 enters the kana string in the variable word dictionary 132. In addition, the matching unit 54 enters the word information represented by the kana string acquired from the common dictionary into the corresponding category of the category table 133.
[0151]
According to the above processing, in each task, after the variable word dictionary 132 is initialized, the contents of the common dictionary are reflected. That is, the variable word dictionary 132 is constructed or reconstructed based on the contents of the common dictionary. Therefore, it is possible to easily maintain consistency in each task as compared with a method of deleting or changing a specific entry in the dictionary.
[0152]
Further, according to the above-described processing, each time a word registered by voice is reflected in each task, the latest pronunciation information at that time is acquired from the unknown word acquisition unit 56, so that it is registered in the variable word dictionary 132. Thereafter, the pronunciation information is updated only by supplying voice data to the unknown word acquisition unit 56, and the matching unit 54 can recognize the voice by referring to the latest pronunciation information at that time.
[0153]
FIG. 26 is a block diagram illustrating the reflection process of FIG. When a kana string is described corresponding to a category in the common dictionary of the common dictionary unit 55, the kana string is registered in the variable word dictionary 132, and the same category as the category of the common dictionary in the category table 133 is set. Information of words represented by the kana string in the common dictionary is registered.
[0154]
On the other hand, when the cluster ID is described corresponding to the category in the common dictionary, the unknown word acquisition unit 56 is referred to, and the representative kana string and the representative phoneme sequence corresponding to the cluster ID are registered in the variable word dictionary 132. Thus, the word information represented by the cluster ID of the common dictionary is registered in the same category as the category of the common dictionary in the category table 133. Note that in the speech recognition process described later, both the fixed word dictionary 131 and the variable word dictionary 132 are used.
[0155]
FIG. 27 shows an application switching task 71 that reflects the contents of the common dictionary unit 55 shown in FIG. ₂ This is an example of the variable word dictionary 132. Application switching task 71 ₂ When the category table 133 of FIG. 13 is as shown in FIG. 13, since the category common to the common dictionary of FIG. 24 is “_robot name_”, the matching unit 54 searches the common dictionary of FIG. Acquires the pronunciation of “SD” corresponding to the name “_”.
[0156]
Then, the matching unit 54 enters the kana pronunciation “SD” acquired from the common dictionary of FIG. 24 in the transcription of the variable word dictionary 132, as shown in FIG. Further, the matching unit 54 converts the phoneme sequence corresponding to the transcription “SD” to “e / s / u /” corresponding to the kana pronunciation “SD” based on the kana phoneme conversion rule 115 (FIG. 8). d / i: / a: / r / u ”is described.
[0157]
Further, the matching unit 54 registers “OOV00001” as a symbol of a word represented by the transcription “SDR”. Here, the symbol is “OOV00001” which means “OOV” + serial number ”, but the symbol may be a character string that can uniquely identify the word. That is, as a symbol, for example, “_robot name _ :: OOV00001” or the like can be used by adding a category name to the head.
[0158]
FIG. 28 shows an application switching task 71 reflecting the contents of the common dictionary of FIG. ₂ An example of the category table 133 is shown. As shown in FIG. 27, when the contents of the common dictionary of FIG. 24 are reflected in the variable word dictionary 132, the contents of the category table 133 are words in the category “_robot name_” shown in FIG. Is registered, the symbol “OOV00001” of the word belonging to the category “_robot name_” registered in the variable word dictionary 132 in FIG. 27 is entered.
[0159]
Next, a process in which the matching unit 54 deletes or changes the word registered in the common dictionary of the common dictionary unit 55 in step S48 of FIG. 21 will be described with reference to the flowchart of FIG. The process of deleting or changing a word in the common dictionary starts, for example, when there is a command from the name registration application unit 21 or when it is necessary to delete an unnecessary registered word due to memory restrictions Is done.
[0160]
In addition, the process of deleting or changing a word in the common dictionary includes, for example, deleting the cluster in the unknown word acquisition unit 56, or changing the ID assigned to the cluster by dividing and merging the cluster, When it is necessary to match the ID assigned to the cluster of the unknown word acquisition unit 56 and the ID described in the common dictionary (the cluster ID described in FIG. 24), the ID described in the common dictionary is This is done to rewrite.
[0161]
Further, the process of deleting or changing the common dictionary is performed by deleting information on the category from the common dictionary when all the tasks described in the language model 112 are no longer used, and reducing the common dictionary. This is done to make it easier.
[0162]
When the unknown word acquisition unit 56 changes the representative phoneme sequence and the kana string of the cluster, the change contents are reflected in the common dictionary by the reflection process in FIG. There is no need to perform a process (hereinafter referred to as a process of changing and deleting as appropriate).
[0163]
In step S <b> 101, the matching unit 54 determines a word to be subjected to the change deletion process from the common dictionary, and proceeds to step S <b> 102. The target word may be determined by the user using a button (not shown), or may be estimated and determined by the matching unit 54.
[0164]
In step S102, the matching unit 54 determines whether or not to delete the word that is the target of the change deletion process. If it is determined to delete the word, the process proceeds to step S103. In step S103, the matching unit 54 deletes the entry of the word that is the target of the change deletion process from the common dictionary. Deletion means deleting an entry specified by a category and pronunciation information, deleting entries of a specific category at once, or deleting entries having specific pronunciation information (kana string or cluster ID) at once. It means to do.
[0165]
On the other hand, in step S102, when the matching unit 54 determines not to delete the word that is the target of the change deletion process, the process proceeds to step S104, determines whether or not to change the word, and determines not to change the word. If YES in step S102, the process returns to step S102 and waits until it is determined to be either change or deletion.
[0166]
In step S104, if the matching unit 54 determines to change the word to be changed and deleted, the matching unit 54 proceeds to step S105 and changes the entry of the word to be changed and deleted in the common dictionary.
[0167]
For example, when the cluster of the unknown word acquisition unit 56 is divided or merged and the ID number of the cluster changes, the matching unit 54 matches the cluster ID of the common dictionary so that the unknown word acquisition unit 56 is matched. To change. For example, when the user wants to correct the kana string input at the time of registration later, the matching unit 54 uses the name registration application unit 21. ₁ , The kana pronunciation of the word that is the target of the common dictionary (the word entered in the common dictionary in step S48 in FIG. 21) is converted into the kana string that the user has entered after determining the target word of the common dictionary. change.
[0168]
After the process of step S103 or the process of step S105, the process proceeds to step S106, and the matching unit 54 performs the reflection process of FIG. 25 to reflect the contents of the common dictionary in each task.
[0169]
As described above, when the words in the common dictionary are deleted or changed, the contents after the change are reflected in each task, so that the consistency of the registered words in each application unit can be maintained.
[0170]
30 and 31 show an example in which the matching unit 54 changes the entry of a word in the common dictionary in the process of step S105 of FIG. For example, when the cluster of the unknown word acquisition unit 56 whose ID is “5” is divided into a cluster whose ID is “8” and a cluster whose ID is “9”, the matching unit 54 shows the common dictionary in FIG. 30A. The state is changed to the state shown in FIG. 30B.
[0171]
That is, the matching unit 54 deletes the entry (the entry on the first line in FIG. 30A) whose cluster ID is “5” in the common dictionary unit 55 and registers “_user name_” registered in the deleted entry. Two entries of the category are registered. Further, the matching unit 54 describes the cluster ID numbers “8” and “9” in the two new entries, respectively (entries on the first and second lines in FIG. 30B).
[0172]
For example, when the cluster of the unknown word acquisition unit 56 with the cluster ID “5” and the cluster with the ID “3” are merged, and the cluster with the ID “10” is newly generated, the matching unit 54 The common dictionary is changed from the state shown in FIG. 31A to the state shown in FIG. 31B.
[0173]
That is, the matching unit 54 changes the cluster IDs of the entries having the cluster IDs “5” and “3” in the common dictionary (all entries in FIG. 31A) to “10”. Two entries of the category “_” and the corresponding cluster ID number “10” are made one (for example, one of them is deleted) (FIG. 31B).
[0174]
Next, the chat process in step S12 of FIG. 5 will be described in detail with reference to the flowchart of FIG.
[0175]
In step S121, the chat application unit 21 ₂ The chat task 71 _Four Is enabled, and the process proceeds to step S122. In step S122, the chat application unit 21 ₂ Controls the matching unit 54 to perform a reflection process as shown in FIG. 25, and changes the contents of the common dictionary of the common dictionary unit 55 to the chat task 71. _Four (Variable word dictionary 132 and category table 133) are reflected. Therefore, chat task 71 _Four Can obtain words that are registered, changed, and deleted in the common dictionary while they are invalid.
[0176]
After the processing of step S122, the process proceeds to step S123, and the chat application unit 21 ₂ Controls the speech recognition engine unit 11 to perform speech recognition processing, and proceeds to step S124. Details of the voice recognition processing will be described later with reference to FIG.
[0177]
In step S124, the chat application unit 21 ₂ Acquires a recognition result from the speech recognition engine unit 11 and generates a response to the recognition result. That is, the robot responds to the utterance from the user. For example, when the utterance from the user is “What time did SDR (robot name) occur?”, The chat application unit 21 ₂ Generates a response for the time when the robot woke up (activated) (eg, “7 o'clock”) and causes the robot to speak.
[0178]
After the processing of step S124, the process proceeds to step S125, and the chat application unit 21 ₂ Determines whether to end the process. This determination is performed by, for example, the chat application unit 21. ₂ However, it is made depending on whether the user has operated (pressed) (pressed) an OK button (not shown) or not.
[0179]
If it is determined in step S125 that the process is not terminated, the process returns to step S123, and the same process is repeated thereafter. That is, the robot continues chatting with the user.
[0180]
If it is determined in step S125 that the process is to end, the process proceeds to step S126, and the chat application unit 21 is started. ₂ The chat task 71 _Four Is disabled and the process proceeds to step S127. In step S127, the chat application unit 21 ₂ Ends the process.
[0181]
In the above-described processing, the chat application unit 21 every time the user speaks once. ₂ Generated a response, but the robot may urge the user to speak by voluntarily speaking.
[0182]
In the process of FIG. 32, the chat application unit 21 ₂ The chat command processing for the voice commander 21 has been described. _Three Voice commander processing ... Other application section 21 _M This process is also performed in the same manner. However, in step S124, processing based on the speech recognition result by the speech recognition engine unit 11 is performed in accordance with the application unit 21.
[0183]
FIG. 33 shows the content of the common dictionary in the common dictionary unit 55 shown in FIG. _Four The state reflected in the variable word dictionary 132 is shown.
[0184]
Chat task 71 _Four When the category table 133 of FIG. 18 is as shown in FIG. 18, the categories common to the common dictionary of FIG. 24 are “_robot name_” and “_place name_”. The first entry in FIG. 24 is acquired as a common dictionary entry corresponding to “name_”, and the fourth and fifth entries in FIG. 24 are acquired as entries corresponding to “_place name_”. Further, the Kana pronunciation “SD” is acquired from the first entry, the Kana pronunciation “Kitashinagawa” is acquired from the fourth entry, and the cluster ID number “3” is acquired from the fifth entry.
[0185]
Then, as shown in FIG. 33, the matching unit 54 enters “SD” and “Kitashinagawa” in the transcription of the variable word dictionary 132. Further, the matching unit 54 applies “e / s / u / d / i:” to the phoneme sequence of the variable word dictionary 132 in accordance with the transcription “SD” based on the Kana phoneme conversion rule 115 (FIG. 8). “/ a: / r / u” and “k / i / t / a / sh / i / n / a / g / a / w / a” are described corresponding to the transcription “Kitashinagawa”.
[0186]
Further, the matching unit 54 extracts the cluster having the cluster ID “3” from the unknown word acquisition unit 56, and acquires a representative phoneme sequence and a kana string. For example, when the unknown word acquisition unit 56 is in the state shown in FIG. 23, the matching unit 54 selects the phoneme sequence “t / o: / ky / o:” from the cluster 153 with the cluster ID “3” and “ Get Kana line "Tokyo". Then, as shown in FIG. 33, the matching unit enters the acquired phoneme sequence “t / o: / ky / o:” and the kana sequence “Tokyo” into the phoneme sequence and transcription of the variable word dictionary 132, respectively. To do.
[0187]
Further, the matching unit 54 transcribes “OOV00001” as a word symbol represented by the transcription “SD” and “OOV00002” as a word symbol represented by the transcription “Kitashinagawa”. “OOV00003” is registered as a symbol of the word represented by the option “Tokyo”.
[0188]
In this case, the phoneme typewriter task 71 is used. ₁ And chat task 71 _Four The phonetic typewriter task 71 assumes that the kana phonological sequence rules 115 of the ₁ A representative phoneme sequence of the cluster obtained by using the chat task 71 as it is. _Four The kana phonological sequence rule 115 is registered in the variable word dictionary 132 of the phonological typewriter task 71. ₁ And chat task 71 _Four When the kana phonological sequence rules 115 of the different are different, the matching unit 54 acquires a representative kana string of the cluster from the unknown word acquisition unit 56, and the chat task 71. _Four The phoneme sequence of the variable word dictionary 132 is described based on the kana phoneme sequence rule 115.
[0189]
FIG. 34 shows that the content of the common dictionary in FIG. 24 is the chat task 71 in FIG. _Four The state reflected in the category table 133 is shown. In the category table 133, for the category of “_robot name_”, words belonging to the category “_robot name_” (words whose transcription is “SD” (FIG. 33)) are stored in the variable word dictionary 132. The symbol “OOV00001” is entered. Further, with respect to the category “_place name_” in the category table 133, the words (transcriptions “Kitashinagawa” and “Tokyo” (FIG. 33)) belonging to the category “_place name_” are stored in the variable word dictionary 132. The registered symbols “OOV00002” and “OOV00003” are entered.
[0190]
Next, the speech recognition process performed by the speech recognition engine unit 11 of FIG. 2 in the process of step S123 of FIG. 32 will be described in detail with reference to the flowchart of FIG. This process is started when a voice is input to the microphone 51 from the user, and the application switching task 71 is started. ₂ , Chat Task 71 _Four Voice commander task 71 _Five ... and other tasks 71 _N Of these, it is performed for each active task.
[0191]
The audio signal generated by the microphone 51 is converted into audio data that is a digital signal by the AD conversion unit 52 and supplied to the feature amount extraction unit 53 in step S141. After the process of step S141, the process proceeds to step S142, and the feature quantity extraction unit 53 extracts a feature quantity such as a mel cepstrum from the supplied audio signal, and the process proceeds to step S143.
[0192]
In step S143, the matching unit 54 connects some of the words represented by the symbols of the fixed word dictionary 131 and the variable word dictionary 132, generates a word string, and calculates an acoustic score. The acoustic score represents how close (acoustically) a word string that is a candidate for the speech recognition result and the input speech are as sounds.
[0193]
After the processing of step S143, the process proceeds to step S144, and the matching unit 54 selects a predetermined number of word strings having a high acoustic score based on the acoustic score calculated in step S143, and proceeds to step S145.
[0194]
In step S145, the matching unit 54 calculates the language score of each word string selected in step S144 using the language model 112, and proceeds to step S146. For example, when grammar or finite state automan is used as the language model 112, when the word string can be accepted by the language model 112, the language score is “1” and cannot be accepted The language score is “0”.
[0195]
The matching unit 54 may leave the word string selected in step S144 when it can be accepted, and delete the word string selected in step S144 when it cannot be accepted.
[0196]
When a statistical language model is used as the language model 112, the generation probability of the word string is used as a language score. Details of the method for obtaining the language score are disclosed in Japanese Patent Application No. 2001-382579 previously proposed by the present applicant.
[0197]
For example, the voice commander application unit 21 _Three When the speech recognition process is performed in the voice commander process, when the matching unit 54 selects the word string “<start> OOV00001 forward <end>” in the process of step S144, the language score is the word string “< Since “start> OOV00001 forward <end>” can be accepted by the language model 112 of the grammar shown in FIG. 19, it becomes “1”.
[0198]
That is, the matching unit 54 recognizes that the category of the symbol “OOV00001” is “_robot name_” with reference to the category table 133 (FIG. 28), and the word string “<beginning” obtained in step S144. > OOV00001 move forward <end> can be converted to the word string “<start> _robot name_advance <end>” using the category name and accepted by the language model 112 shown in FIG. Judge that it is possible.
[0199]
On the other hand, for example, when the word string “<Proceed to <start> OOV00001” <end> is selected in step S144, the matching unit 54 refers to the category table 133 (FIG. 28) and refers to the category “OOV00001” Is “_robot name_” and the word string “<start> Go to OOV00001 before <end>” obtained in step S144 is replaced with the word string “<start> _robot name” using the category name. It is determined that it cannot be accepted by the language model 112 shown in FIG. 19, and the language score of this word string is set to “0”.
[0200]
In step S146, the matching unit 54 integrates the acoustic score calculated in step S143 and the language score calculated in step S145, sorts each word string, for example, the word string having the largest integrated score. Is determined as a recognition result.
[0201]
As a result, the most appropriate word string acoustically and linguistically is determined as the recognition result.
[0202]
After the processing of step S146, the process proceeds to step S147, and the matching unit 54 determines whether the recognition result includes a word registered in speech (words clustered in the unknown word acquisition unit 56).
[0203]
If it is determined in step S147 that the word registered by speech is included in the recognition result, the process proceeds to step S148, where the matching unit 54 supplies the word to the unknown word acquisition unit 56, and the unknown word acquisition unit 56 performs re-clustering. Then, the process proceeds to step S149.
[0204]
For example, in step S144, if the word string “<tip> I went to Tokyo today <end>” including the place name (unknown word) “Tokyo”, the matching unit 54 is an unknown word. “Tokyo” voice data, phonetic typewriter task 71 ₁ The phoneme sequence (for example, “t / o: / ky / o:”) and the kana string (for example, “Tokyo”) recognized with reference to FIG. Then, the unknown word acquisition unit 56 performs reclustering.
[0205]
As a result, the amount of speech data supplied to the unknown word acquisition unit 56 increases, and the representative phoneme sequence and the representative kana sequence of each cluster may be updated to correct values. However, as a side effect, even after a correct phoneme kana sequence / kana sequence is acquired, there is a possibility that it will be changed to an incorrect value by re-clustering. In order to prevent such a side effect, when there is an instruction from the user, the pronunciation can be fixed by describing the kana string at that time in the entry of the common dictionary. For example, in FIG. 23, when the kana string of the cluster with ID = 3 is pronounced “Tokyo”, the location described as “cluster ID = 3” in the common dictionary in FIG. 24 is “Kana pronunciation: Tokyo”. (The fifth entry is the target of the rewriting). By doing this, even if the kana string of the cluster with ID = 3 changes to something other than “Tokyo”, the pronunciation of the fifth entry in the common dictionary is fixed to “Tokyo”.
[0206]
On the other hand, in step S147, when the matching unit 54 determines that the speech recognition result does not include a word registered by speech, the matching unit 54 skips step S148 and proceeds to step S149.
[0207]
In step S149, the matching unit 54 supplies the recognition result determined in step S146 to the application unit 21 corresponding to the task.
[0208]
Here, the chat application unit 21 ₂ In the chat process of FIG. 36, for example, FIG. 36 shows an expression for obtaining the language score when the matching unit 54 selects the word string “<Start> OOV00001 occurred at <End>” in step S144 of FIG. Show.
[0209]
The language score “Score (<start> OOV00001 when it occurred <end>)” generates the word string “<start> OOV00001 occurred when <end>” as shown in equation (1). It is a probability.
[0210]
The value of the language score “Score (<start> OOV00001 happened at <end>)” is exactly “P (<start>) P (OOV00001 | <start) as shown in Equation (2) >) P (ha | <top> OOV00001) P (what | | <top> OOV00001 is) P (at | <top> OOV00001 is at what time) P (was | | at the beginning> OOV00001 is at what time) P (of | <First> OOV00001 occurred at what time) P (<Termination> | <Start> OOV00001 occurred at what time), but as shown in FIG. 16, the language model 112 uses a tri-gram. So, the condition parts “<Start> OOV00001 is”, “<Start> OOV00001 is what time”, “<Start> OOV00001 is what time”, “<Start> OOV00001 is what time”, and “<Start> OOV00001 is “What happened” is the maximum two words “OOV00001”, “What time”, “When”, “Wake up”, and “Wake up” Approximate with conditional probabilities limited to “Tano” (Equation (3)).
[0211]
This conditional probability is obtained by referring to the language model 112 (FIG. 16), but since the language model 112 does not include the symbol “OOV00001”, the matching unit 54 refers to the category table 133 of FIG. Then, it recognizes that the category of the word represented by the symbol “OOV00001” is “_robot name_”, and converts “OOV00001” to “_robot name_”.
[0212]
That is, as shown in Expression (4), “P (OOV00001 | <first>)” is changed to “P (_robot name_ | <first>) P (OOV00001 | _robot name)”, and “P (_Robot name_ | <head>) "/ N". N represents the number of words belonging to the category “_robot name_” in the category table 133.
[0213]
That is, when the probability is described in the form of P (X | Y), if the word X is a word belonging to the category C, P (C | Y) is obtained from the language model 112, and the value P (X | C) Multiply (probability of generating word X from category C). Assuming that all the words belonging to category C are generated with equal probability, if there are N words belonging to category C, P (X | C) can be approximated to 1 / N.
[0214]
In FIG. 34, only the word represented by the symbol “OOV00001” belongs to the category “_robot name_”, so “N” is “1”. Therefore, as shown in Expression (5), “P (ha | <start> OOV00001)” becomes “P (ha | <head> _robot name_)”. Also, “P (what time | OOV00001 is)” becomes “P (what time | _robot name_ is), as shown in Expression (6).
[0215]
Thereby, the language score can be calculated even for a word string including a variable word, and the variable word can appear in the recognition result.
[0216]
In the above example, the activation of the application unit 21 and the validity of the task 71 are linked with the termination of the application unit 21 and the invalidation of the task 71. However, this is performed at another timing, for example, the activation of the application unit 21 It is also possible to switch between valid / invalid of a task many times and control a plurality of tasks with one application.
[0217]
In this case, a task that frequently switches between enabling and disabling frequently is not efficient if memory is allocated or released repeatedly, so a flag (a flag indicating that the task is invalid) is set even after it is disabled. You can just leave the memory reserved.
[0218]
Further, in the above example, it is assumed that nothing is stored in the common dictionary of the common dictionary unit 55 when the robot system is activated. However, even if several words are stored in the common dictionary in advance. Good. For example, since the product name of a robot is often registered in the name of the robot, the product name of the robot may be registered in advance in the category “_robot name_” of the common dictionary.
[0219]
FIG. 37 shows an example of a common dictionary when the robot product name “SD” is entered in the category “_robot name_” when the robot system is activated. In FIG. 37, when the robot system is activated, the kana pronunciation “SD” is entered in the category “_robot name_”, so that the user can pronounce the kana pronunciation “SD” without registering the name. The word can be used to control the robot.
[0220]
In the above example, it is assumed that no unknown word acquisition unit cluster is generated in the initial stage (at the time of shipment). However, if a cluster is prepared for the main name from the beginning, when the name is input by voice in the name registration process of FIG. 21, the name for which the cluster is prepared is easily recognized. For example, if a cluster as shown in FIG. 3 is prepared at the time of shipment, the pronunciation of voices “red”, “blue”, “midori”, and “black” can be recognized (acquired) with a correct phoneme sequence. Furthermore, it is not desirable that the pronunciation of a name for which a cluster is prepared changes after it is registered in the common dictionary. Therefore, when registering pronunciation information in the common dictionary, instead of describing the cluster ID, describe it with kana pronunciation (representative kana string of clusters) such as “red”, “ao”, “midori”, “black”. .
[0221]
Furthermore, in the above-described example, the matching unit 54 reflects the contents of the common dictionary on all tasks. However, the matching unit 54 may reflect only on the tasks to be reflected. For example, a number (task ID) is added in advance to the task, and the common dictionary in FIG. 24 is expanded to provide a column indicating “list of tasks in which this entry is valid (or invalid)”. In the processing, the matching unit 54 only needs to reflect the contents of the common dictionary only to the task to which the task ID described in the column indicating “list of tasks for which this entry is valid” is added.
[0222]
FIG. 38 shows an example in which the ID of a task to be reflected is described in a column representing “valid task” in the common dictionary. In FIG. 38, the words whose kana pronunciation belonging to the category “_robot name_” is represented by “SD” have the task IDs “1”, “2”, and “4”. Only the variable word dictionary 132 and the category table 133 of the tasks “1”, “2”, and “4” reflect the contents of the common dictionary of words represented by kana pronunciation “SDR”.
[0223]
In the above example, the word stored in the fixed word dictionary 131 is a word described in the language model 112, and the word stored in the variable word dictionary 132 is a word belonging to a category. However, a part of the words belonging to the category may be stored in the fixed word dictionary 131.
[0224]
FIG. 39 shows an application switching task 71. ₂ FIG. 40 shows an example of the category table 133 at the time of activation. That is, in the category table 133 of FIG. 40, the category “_robot name_” and the word symbol “OOV00001” belonging to the category “_robot name_” are registered in advance. 40 includes the symbol “OOV00001”, the transcription “SD” of the word represented by the symbol “OOV00001”, and the phoneme sequence “e / s / u / d / i”. : / a: / r / u ”is registered in advance.
[0225]
In this case, the speech recognition process is performed assuming that the word “SDR” belongs to the category “_robot name_”. That is, the word “SDR” is treated as a robot name from the beginning. However, since the word “SD” is stored in the fixed word dictionary 131, it cannot be deleted or changed.
[0226]
In this way, for example, by storing in the fixed word dictionary 131 in advance a word that is assumed to be set as a name, such as a product name of the robot, the user can control the robot without registering the name. Can do.
[0227]
In the above example, the category symbols are common to all tasks, but may not be common. In this case, a conversion table as shown in FIGS. 41 to 44 may be prepared in the task.
[0228]
That is, for example, when a category “_ROBOT_NAME_” and a category “_USER_NAME_” are described in a task T, according to the conversion table of FIG. The content of the common dictionary of words belonging to is reflected in the category “_ROBOT_NAME_”. In the task T, the contents of the common dictionary of words belonging to the category “_user name_” are reflected in the category “_USER_NAME_”.
[0229]
Further, for example, when a category “_proper noun_” is described in a certain task T, according to the conversion table of FIG. 42, in the task T, the common dictionary unit 55 sets the category “_robot name_”. Both the contents of the common dictionary of words belonging to and the contents of the common dictionary of words belonging to the category “_user name_” are reflected in the category “_proper noun_”.
[0230]
Further, for example, when a category “_Last Name_” and a category “_First Name_” are described in a certain task T, according to the conversion table of FIG. The contents of the common dictionary of words belonging to “_user name_” are converted and copied into the category “_last name_” and the category “_name_”. In the step of reflecting the contents of the common dictionary to this task (FIG. 25), for example, the second entry in FIG. 24 is changed from “_user name_cluster ID = 5” to “_surname_cluster ID =” according to the conversion table. 5 ”and“ _name_cluster ID = 5 ”are converted and duplicated, and then reflected in the fixed word dictionary and the category table of this task.
[0231]
Further, for example, when a category is not described in a certain task T, according to the conversion table of FIG. 44, the category “_robot name_” and category “_user name_” in the common dictionary unit 55 are included in the task T. , A word belonging to the category “_place name_” is represented by the symbol “UNK”. “UNK” means “Unknown word”.
[0232]
As a result, even in a task for which no category is described, the matching unit 54 can simply select the category “_robot name_” and the category “_user name_” simply by describing the symbol “UNK” in the language model 112. , Words belonging to the category “_place name_” can be recognized.
[0233]
FIG. 45 shows an example of the external configuration of a biped robot equipped with the robot control system 1 to which the present invention is applied. In the robot 201, a head unit 211 is disposed on the upper portion of the body unit 213, and arm units 212A and 212B having the same configuration are disposed on the upper left and right of the body unit 213, respectively. Leg units 214A and 214B having the same configuration are attached to predetermined positions on the lower left and right sides of 213, respectively.
[0234]
The head unit 211 includes CCD (Charge Coupled Device) cameras 221A and 221B that function as “eyes” of the robot 201, microphones 222A and 222B that function as “ears”, and a speaker 223 that functions as a “mouth”. Are arranged at predetermined positions.
[0235]
FIG. 46 shows an example of the electrical configuration of the robot. The unit control system 231 and the dialogue control system 232 control the operation of the robot 201 in accordance with commands from the robot control system 1. That is, the unit control system 231 controls the head unit 211, the arm units 212A and 212B, and the leg units 214A and 214B of the robot 201 as necessary, and causes the robot 201 to perform a predetermined operation. Further, the dialogue control system 232 controls the utterance of the robot 201 and causes the speaker 223 to make a predetermined utterance as necessary.
[0236]
In the above description, a word is a unit that should be handled as one unit in the process of recognizing speech, and does not necessarily match a linguistic word. For example, “Taro-kun” may be treated as a single word, or may be treated as two words “Taro” and “you”. In addition, may be dealing with more is a major unit of "Hello Taro" or the like as one word.
[0237]
A phoneme is one that is more convenient in terms of processing if it is treated acoustically as one unit, and does not necessarily match phonemic phonemes or phonemes. For example, the “To” part of “Tokyo” can be represented by three phonetic symbols “t / o / u”, or a symbol “o:” that is a long sound of “o” is prepared. May be. Further, it may be expressed as “t / o / o”. In addition, symbols that indicate silence can be prepared, or further divided into "silence before speech", "short silence between speeches", and "silence of the part". Good.
[0238]
Although the robot apparatus has been described above, the present invention can be applied to an apparatus having an application using speech recognition, speech synthesis, translation, and other language processing.
[0239]
Furthermore, the present invention can be applied to, for example, an apparatus for extracting a predetermined term from a dictionary registered in Kojien and creating the term dictionary.
[0240]
In the above description, the case where there are a plurality of application units has been described. However, the number of application units may be one.
[0241]
The series of processes described above can be executed by hardware or can be executed by software. In this case, the processing described above is executed by a personal computer 600 as shown in FIG.
[0242]
47, a CPU (Central Processing Unit) 601 performs various processes according to a program stored in a ROM (Read Only Memory) 602 or a program loaded from a storage unit 608 to a RAM (Random Access Memory) 603. Execute. The RAM 603 also appropriately stores data necessary for the CPU 601 to execute various processes.
[0243]
The CPU 601, ROM 602, and RAM 603 are connected to each other via an internal bus 604. An input / output interface 605 is also connected to the internal bus 604.
[0244]
The input / output interface 605 includes an input unit 606 including a keyboard and a mouse, a display including a CRT and an LCD (Liquid Crystal Display), an output unit 607 including a speaker, a storage unit 608 including a hard disk, a modem, and the like. A communication unit 609 including a terminal adapter is connected. The communication unit 609 performs communication processing via various networks including a telephone line and CATV.
[0245]
A drive 610 is connected to the input / output interface 605 as necessary, and a removable medium 621 composed of a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is appropriately attached, and a computer program read therefrom is It is installed in the storage unit 608 as necessary.
[0246]
When a series of processing is executed by software, a program constituting the software executes various functions by installing a computer incorporated in dedicated hardware or various programs. For example, it is installed in a general-purpose personal computer from a network or a recording medium.
[0247]
As shown in FIG. 47, this recording medium is not only composed of a package medium composed of a removable medium 621 on which a program is recorded, which is distributed to provide a program to a user, separately from the computer. A ROM 602 storing a program and a hard disk including a storage unit 608 provided to the user in a state of being pre-installed in the apparatus main body.
[0248]
In the present specification, the step of describing a computer program includes not only processing performed in time series according to the described order but also processing executed in parallel or individually even if not necessarily processed in time series. Is also included.
[0249]
Further, in this specification, the system represents the entire apparatus composed of a plurality of apparatuses.
[0250]
By the way, in Japanese Patent Application No. 2001-382579 previously proposed by the present applicant, a series of processes until the word acquired by the unknown word acquisition mechanism is reflected in the language model and the word reflected in the language model are An invention about a process for making it appear in a recognition result is disclosed.
[0251]
However, the invention of Japanese Patent Application No. 2001-382579 is composed of an application for registering one word and an application that uses one registered word, and it is assumed that there are multiple applications that perform speech recognition. Therefore, in a system in which there are a plurality of applications and a variable number of applications, the above-described problem when reflecting registered words in the application, and the above-described problem when deleting or changing registered words for a plurality of applications are described. It was difficult to solve the problem.
[0252]
In addition, in Japanese Patent Application No. 2002-072718 previously proposed by the applicant, the word “Taro”, which is an unknown word, is extracted from the utterance “My name is Taro (unknown word).” An invention of obtaining as a name is disclosed.
[0253]
However, the invention of Japanese Patent Application No. 2002-072718 is based on the above-described problem of reflecting unknown words in the language model, the above-mentioned problem of reflecting registered words in the application, and the words registered for a plurality of applications. It has been difficult to solve the above-described problems when deleting or changing the.
[0254]
Therefore, in the invention of Japanese Patent Application No. 2001-382579 and the invention of Japanese Patent Application No. 2002-072718, the above-mentioned problem of reflecting an unknown word in a language model, the above-mentioned problem of reflecting a registered word in its application, In addition, it has been difficult to solve all the problems described above when deleting or changing words registered for a plurality of applications.
[0255]
However, in the robot control system 1 of FIG. 1, since the category is described in the language model 112, the above-described problem of reflecting the unknown word in the language model by causing the unknown word to belong to the category is solved. Can do.
[0256]
Further, in the robot control system 1 of FIG. 1, the variable word dictionary 132 in which words to be used for speech recognition used in the application are registered or reconstructed based on the common dictionary. It is possible to solve the above-described problem when reflecting a word that has been applied to the application and the above-described problem when deleting or changing a word registered for a plurality of applications.
[0257]
Furthermore, in the case of a system that does not have a keyboard (for example, a robot), there is a problem that it is difficult to input pronunciation information at the time of word registration. As a means for solving the problem, for example, a phonological typewriter is used. A method for inputting pronunciation information by voice has been proposed.
[0258]
However, the phonological typewriter sometimes misrecognizes, and if the phonological typewriter is used as it is, there is a problem that a word may be registered with a wrong pronunciation. For example, if the phonetic typewriter recognizes the pronunciation of “SDR” and the phoneme typewriter misrecognizes and outputs the result of “Ilnyal”, if “Ilnyal” is adopted as pronunciation information, in order to word is registered, and later, for example, "es Dr, Hello" that although the speech is difficult to be recognized, "Isuniyaru, Hello" speech that the context of recognized easily occurs.
[0259]
Therefore, in the robot control system 1 of FIG. 1, each time a word registered by voice is reflected in each task, the latest pronunciation information at that time is acquired from the unknown word acquisition unit 56, so that the phonological typewriter has an error. Even after the recognized and misrecognized words are registered in the variable word dictionary 132, the pronunciation information is updated only by supplying voice data to the unknown word acquisition unit 56, and the latest pronunciation information at that time can be obtained. There is a possibility of obtaining a normal recognition result.
[0260]
That is, in the robot control system 1 of FIG. 1, the above-described problem of reflecting an unknown word in a language model, the above-described problem of reflecting a registered word in the application, and the words registered for a plurality of applications The above-described problems when deleting or changing, and the above-described problems when inputting pronunciation information of a word to be registered in a system without a keyboard, that is, all the problems described above can be solved.
[0261]
【The invention's effect】
As described above, according to the present invention, words can be registered. In particular, even when registering words corresponding to a plurality of applications, the registered words can be commonly used in each application. In addition, words registered before the application is started can also be used in the application. Furthermore, even when the registered word is changed, consistency can be maintained in each application.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration example of a robot control system to which the present invention is applied.
2 is a block diagram illustrating a configuration example of a speech recognition engine unit in FIG. 1. FIG.
FIG. 3 is a diagram illustrating an example of a cluster of an unknown word acquisition unit in FIG. 2;
4 is a flowchart for explaining robot control processing in the robot control system of FIG. 1; FIG.
FIG. 5 is a flowchart for explaining robot control processing in the robot control system of FIG. 1;
6 is a diagram illustrating a configuration example of the task in FIG. 2. FIG.
7 is a diagram showing an example of a phoneme list in FIG. 6. FIG.
8 is a diagram illustrating an example of a kana phoneme conversion rule in FIG. 6; FIG.
FIG. 9 is a diagram illustrating an example of a language model of the phoneme typewriter task of FIG. 2;
10 is a diagram showing an example of a fixed word dictionary of the phoneme typewriter task of FIG. 2; FIG.
FIG. 11 is a diagram illustrating an example of a language model of the application switching task in FIG. 2;
12 is a diagram showing an example of a fixed word dictionary of the application switching task of FIG. 2;
13 is a diagram showing an example of a category table for the application switching task in FIG. 2; FIG.
14 is a diagram illustrating an example of a language model of the name registration task in FIG. 2;
15 is a diagram showing an example of a fixed word dictionary of the name registration task of FIG. 2;
16 is a diagram illustrating an example of a language model of the chat task in FIG. 2;
FIG. 17 is a diagram showing an example of a fixed word dictionary of the chat task in FIG. 2;
18 is a diagram showing an example of a category table for the chat task in FIG. 2. FIG.
FIG. 19 is a diagram illustrating an example of a language model of the voice commander task in FIG. 2;
20 is a diagram showing an example of a fixed word dictionary of the voice commander task of FIG. 2;
FIG. 21 is a flowchart for describing name registration processing in step S9 of FIG. 5;
FIG. 22 is a flowchart illustrating name recognition processing in step S43 of FIG.
FIG. 23 is a diagram illustrating an example of a cluster of the unknown word acquisition unit in FIG. 2;
FIG. 24 is a diagram illustrating an example of the common dictionary unit in FIG. 2;
FIG. 25 is a flowchart illustrating the reflection process in step S49 of FIG.
FIG. 26 is a block diagram illustrating the reflection process in FIG. 25;
FIG. 27 is a diagram showing an example of a variable word dictionary of the name registration task of FIG. 2;
FIG. 28 is a diagram showing an example of a category table for the name registration task in FIG. 2;
FIG. 29 is a flowchart illustrating word deletion or change processing in the matching unit of FIG. 2;
30 is a diagram illustrating an example of changing the common dictionary unit in FIG. 2;
FIG. 31 is a diagram showing an example of changing the common dictionary part in FIG. 2;
32 is a flowchart for explaining the chat process in step S12 of FIG.
FIG. 33 is an example of a variable word dictionary of the chat task in FIG. 2;
FIG. 34 is an example of a chat task category table of FIG. 2;
FIG. 35 is a flowchart for describing the speech recognition processing in step S123 of FIG.
FIG. 36 shows an example of a language score calculation formula.
FIG. 37 shows a modification of the common dictionary part of FIG.
38 is a diagram showing a modification of the common dictionary unit in FIG. 2;
FIG. 39 is a diagram showing a modification of the fixed word dictionary of FIG.
40 is a diagram showing an example of the category table in FIG. 6. FIG.
FIG. 41 is a diagram illustrating an example of a category conversion table.
FIG. 42 is a diagram illustrating an example of a category conversion table.
FIG. 43 is a diagram illustrating an example of a category conversion table.
FIG. 44 is a diagram illustrating an example of a category conversion table.
45 is a perspective view showing an external configuration of a robot. FIG.
FIG. 46 is a block diagram showing an electrical configuration of the robot.
FIG. 47 is a diagram illustrating an example of a personal computer.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 11 Speech recognition engine part, 21 Application part, 31 Application management part, 51 Microphone, 52 AD conversion part, 53 Feature-value extraction part, 54 Matching part, 55 Common dictionary part, 56 Unknown word acquisition part, 71 Task, 111 Acoustic model , 112 language model, 113 dictionary, 114 phoneme list, 115 kana phoneme conversion rule, 116 search parameter, 131 fixed word dictionary, 132 variable word dictionary, 133 category table

Claims (7)

A language processing apparatus having a plurality of applications using language processing,
A registered dictionary storage means for storing a registered dictionary in which words are registered;
A construction means for constructing a dedicated dictionary dedicated to the application, in which words to be used for language processing used in the application are registered, based on the registered dictionary for each application ;
Processing means for performing processing for adding, deleting, or changing words to the registered dictionary;
Deleting means for deleting words in the dedicated dictionary;
With
After all the words registered in the dedicated dictionary are deleted,
The language processing apparatus , wherein the construction unit reconstructs the dedicated dictionary based on the registered dictionary in which a word is added, deleted, or changed . The dedicated dictionary includes at least a fixed dictionary in which predetermined words are registered in advance, and a variable dictionary in which registered words are variable,
The language processing apparatus according to claim 1, wherein the construction unit constructs the variable dictionary of the dedicated dictionary. The dedicated dictionary further includes a category table in which word categories are registered,
The construction unit, among the words of the registration dictionary, the word of the registered in the category table category, by registering the variable dictionary, according to claim 2, characterized in that constructing the variable dictionary Language processor. Language model storage means for storing a language model that describes chain information indicating how words of the category are chained;
Recognition processing means for performing speech recognition based on the dedicated dictionary and the language model;
Further comprising
The language processing apparatus according to claim 3. A language processing method of a language processing apparatus having a plurality of applications using language processing,
A registered dictionary storage step for storing a registered dictionary in which words are registered;
A construction step in which a word dedicated to language processing used in the application is registered, and a dedicated dictionary dedicated to the application is built for each application based on the registration dictionary ;
A processing step of performing processing for adding, deleting, or changing a word with respect to the registered dictionary;
A deletion step of deleting words in the dedicated dictionary;
And reconstructing the dedicated dictionary based on the registered dictionary in which words are added, deleted, or changed after all the words registered in the dedicated dictionary are deleted. Language processing method. A program that performs language processing of multiple applications,
A construction step for constructing a dedicated dictionary dedicated to the application in which words to be used for language processing used in the application are registered based on the registered dictionary in which words are registered for each application ;
A processing step of performing processing for adding, deleting, or changing a word with respect to the registered dictionary;
A deletion step of deleting words in the dedicated dictionary;
Causing a computer to execute a restructuring step of reconstructing the dedicated dictionary based on the registered dictionary in which words are added, deleted, or changed after all the words registered in the dedicated dictionary are deleted. A recording medium on which a program is recorded . A program that performs language processing of multiple applications,
A construction step for constructing a dedicated dictionary dedicated to the application in which words to be used for language processing used in the application are registered based on the registered dictionary in which words are registered for each application ;
A processing step of performing processing for adding, deleting, or changing a word with respect to the registered dictionary;
A deletion step of deleting words in the dedicated dictionary;
Causing a computer to execute a rebuilding step of reconstructing the dedicated dictionary based on the registered dictionary in which words are added, deleted, or changed after all the words registered in the dedicated dictionary are deleted. A program characterized by

Images