JP2011221293A - Command processing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a command processing device which facilitates a screen operation of a correction work when a selected command is not a user's intended command.SOLUTION: A command processing device comprises: a command selection section 3 which selects an intermediate command which is likely to be a command candidate according to a likelihood worked-out by assuming a sound model and a language model from an input data while a recognition result of a vocabulary recognition is worked out by a voice recognition process section 2 according to the sound model and the language model with respect to the input data entered by a user; an operation cost database memory 7 which stores the operation cost for an operation between commands by defining an error of the operation made by the user as the operation cost; and a likelihood correction section 4 which selects a command from the intermediate commands for minimizing an worked-out evaluation value based on the operation cost.

Description

  The present invention relates to a command processing apparatus that executes processing in response to an operation command input by voice.

  Conventionally, a multimodal information device having a screen operation and a voice operation such as a navigation device has been developed. The screen operation is created along a hierarchical structure of commands in which the screen transition relationship is associated with a specific function of the device. For this reason, by proceeding with the operation, the user's desired command can be gradually approached and executed. However, it is necessary to perform one screen operation for each screen transition, which is inconvenient because several operations are required for execution depending on the command. The voice operation can be executed with one utterance regardless of the command, and the inconvenience of the screen operation can be compensated. However, a command different from the user's desired command may be executed due to erroneous recognition.

なお、式（１）において、Ｃはコマンド、Ｘは入力音声、Ｐ（Ｃ）はコマンドごとの事前に判明している出現確率、Ｐ（Ｘ｜Ｃ）は事前の学習された確率モデルに基づく入力音声の尤度である。ａｒｇ ｍａｘは要素の中で最大の値の要素を返す関数として与えられるものである。 A command processing apparatus that maps a vocabulary obtained by speech recognition processing to a command is used in the field of speech recognition, and is effective for operating devices with various vocabularies. For example, in the command processing device disclosed in Patent Document 1, a command is selected based on a vocabulary most likely to be acoustically and linguistically uttered by a user who has been recognized by speech recognition. Further, in the command processing device disclosed in Patent Document 2, based on the most probable vocabulary, using the conversion database in which the command and related phrase information are associated, the most probable as shown in the following formula (1). The selection of a new command is performed.

In Equation (1), C is a command, X is an input voice, P (C) is a known occurrence probability for each command, and P (X | C) is based on a prior learned probability model. This is the likelihood of the input speech. arg max is given as a function that returns the element with the maximum value among the elements.

Japanese Patent Laid-Open No. 9-50291 International Publication WO2007 / 114226

Since the conventional command processing apparatus is configured as described above, the command selection is performed based on the equation (1) and the command with the highest likelihood P (X | C) is executed, or In order to present multiple command candidates and allow the user to select them, when presenting multiple command candidates, to determine based on likelihood, when presenting multiple completely different command candidates as the content of the function, to confuse the user There was a problem that there was. This problem occurs especially when the function is complicated.
Furthermore, when a misrecognition occurs, it is not possible to select a command that is hierarchically similar so that the user can know how to make a transition, and this causes a great deal of effort for correction work. was there.

  The present invention has been made to solve the above-described problems. By selecting a command in consideration of the hierarchical relationship of commands, the correction performed when the selected command is not the user's desired command. The purpose is to reduce the labor associated with screen operations.

  The command processing device according to the present invention performs vocabulary recognition on input data input by a user, and outputs a probable vocabulary as a recognition result, and when the result appears according to a model of the recognition result On the other hand, a likelihood that is a likelihood of guessing the model when viewed from the input data is calculated, and a command selection unit that selects an intermediate command that is a probable command candidate based on the likelihood, and the user performs an operation error. Operation cost database memory that stores the operation cost associated with the operation between commands, and the likelihood of selecting the command that minimizes the evaluation value calculated based on the operation cost from the intermediate command A correction unit is provided.

  According to the present invention, when the recognition result calculated by performing vocabulary recognition on the input data input by the user by the recognition processing unit appears according to the model, the model is inferred from the input data. A command selection unit that calculates a likelihood that is a likelihood to be performed, selects an intermediate command that is a probable command candidate based on the likelihood, and defines an operation when the user makes an operation error as an operation cost, The operation cost database memory for storing the operation cost associated with the operation and the likelihood correction unit for selecting the command for minimizing the evaluation value calculated based on the operation cost from the intermediate command. Commands can be selected in consideration of hierarchical relationships, reducing the labor involved in screen operations, which is a modification of selected commands. Door can be.

It is explanatory drawing which shows an example of the hierarchical relationship in the command processing of this invention. It is a figure which shows the database showing the operation cost based on the screen operation in the command processing of this invention. It is explanatory drawing which shows an example of the average value of the operation cost in the command processing of this invention. It is explanatory drawing which shows an example of the hierarchical relationship in the command processing of this invention. It is explanatory drawing which shows calculation of the expected value of the operation cost in the command processing of this invention. It is a block diagram which shows the structure of the command processing apparatus by Embodiment 1 of this invention. It is a flowchart which shows operation | movement of the command processing apparatus by Embodiment 1 of this invention. It is explanatory drawing which shows an example of the conversion database of the command processing apparatus by Embodiment 1 of this invention. It is a block diagram which shows operation | movement of the command processing apparatus by Embodiment 2 of this invention. It is a flowchart which shows operation | movement of the command processing apparatus by Embodiment 2 of this invention.

  In the present invention, in the command processing device, an operation cost based on the screen operation is defined in consideration of correction in the case of erroneous recognition, and a command is selected so as to minimize the average value (evaluation value) of the operation cost. It is characterized by that. In the description of the present invention, first, a hierarchy diagram of commands based on screen operations and operation costs will be described, and then a voice command selection procedure for minimizing the average value of operation costs based on operation costs will be described.

FIG. 1 is an explanatory diagram showing an example of a hierarchical relationship between commands of an operation cost database according to the present invention.
A command based on a screen operation presents an operation screen, and a screen transition is performed by repeatedly performing an operation selected by the user from the screen, and finally a desired command is executed. An intermediate node having lower nodes represents a selection screen representing command classification, and a terminal node represents a specific command. The user executes specific commands associated with the end nodes F11 to F13 and F21 to F23 through the intermediate nodes F10 and F20 by screen operation from the initial node F0. On the other hand, in the voice operation, the end nodes F11 to F13 and F21 to F23, which are the end commands, can be executed from the initial node F0 by directly speaking the command.

  Next, the operation cost database will be described. When performing an operation based on a screen operation, the user confirms the content presented on the screen and repeats the menu selection. If an operation other than the desired operation is selected by mistake, it is necessary to return to the upper hierarchy and select another menu. The operation cost is associated with the screen transition, and specifically is associated with the average operation time. The cost generated for each screen operation will be described with reference to FIG. FIG. 2 is a diagram showing the operation cost based on the screen operation as a database. Specifically, the operation cost generated in accordance with the movement between commands shown in FIG. 1 is shown in the database.

  Hierarchy No is “0” for the initial node, “1” for the intermediate node, “2” for the end node, and the value increases as the level moves down. Corresponding command IDs and command names are also shown. The operation cost for moving between commands is also described. For example, the operation cost for moving from command F11 to F12 is the operation cost “1” for returning from F11 to F10, and the operation cost “1” for proceeding from F10 to F12. Therefore, the expected value of the operation cost is calculated as “12” with a likelihood 6 of the number of operations 2 × F12 in total.

In FIG. 2, for example, when the commands output from the command selection unit 3 described later have the same or approximate likelihoods in “TOLL_FIRST_ROUTE”, “RECOMMENDED_ROUTE”, and “SHORTEST_ROUTE”, the following formula is used under the following conditions: The optimum command is obtained from (2).
<Conditions>
・ The simplest operation (one movement between commands) is calculated as the operation cost “1”.
F10 has a score of “3”, F11 and F12 have a score of “6”, and F13 has a score of “5”.
Expression (2) is an expression for calculating an average value of operation costs and outputting a command having the smallest value as the final command C ^ and is expressed as follows.

  In Equation (2), C and C ′ are individual voice commands, and X is an input voice. Also, l () is a loss function that is a function that returns an operation cost that occurs in accordance with the dictionary transition in parentheses, and P (C ′ | X) is a likelihood that the command C ′ is obtained when the speech X is observed. is there. arg min is a function that returns the element having the smallest value among the elements. Equation (2) corresponds to the case where the loss function l in Bayesian decision theory is considered in Equation (1), and can be transformed into Equation (4) under the condition of Equation (3), and is therefore equal to Equation (1). The Bayesian decision theory is described in Reference 1.

Reference 1.
Taizo Iijima: Pattern Recognition Theory, Basic Information Engineering Series 6, Morikita Publishing (1989)

FIG. 3 is a diagram illustrating an average value of operation costs when each command is a command desired by the user, that is, a correct command, calculated using Expression (2). Referring to FIG. 3, F10 that minimizes the average operation cost is finally output as a command.
In this example, a command that is not desired by the user is not selected, it is possible to present options to the user at the intermediate node, select the desired command in the next operation, and suppress operation costs. In addition, when the user's utterance is ambiguous, the operation cost of the user is reduced by presenting the command when there is a relatively close command, as compared with the conventional technique that always moves to the intermediate node.

  In consideration of the case where F11 and F21 are unlikely to be ambiguous, the average operation cost is not calculated for all commands, but a specific group (in the example of the hierarchical relationship between commands shown in FIG. 4). It is also possible to obtain the average value of the operation cost by limiting to the group of F10 to F13 shown. A specific configuration will be described in Embodiment 2.

Further, in Expression (2), as shown in Expression (5), the expected value of the operation cost and the weight of likelihood can be changed.

In Equation (5), C and C ′ are individual voice commands, and X is an input voice. Further, l () is a loss function that is a function that returns an expected value of the operation cost that occurs in association with the transition of the event in parentheses, and P (C ′ | X) is a command C ′ when the voice X is observed. Likelihood. arg min is a function that returns an element having the smallest value among the elements, λ ₁ is given as a weight of the operation cost, and λ ₂ is given as a weight of likelihood. As shown in FIG. 5, when the operation cost weight λ ₁ is 0 in the equations (1) and (2), the same final command C ^ is output.

In the above example, the case where an operation cost of “1” is generated for each screen operation has been described as an example. However, a recognition cost or a misrecognition that is a cost generated when selecting a function in the same hierarchy is described. The cost of misrecognition, which is a cost arising from ease, the return cost, which is the cost when returning to a higher function, the reselection cost, which is the cost of returning to the menu screen and following the same hierarchy, etc., is used alone or in combination The operation cost may be calculated.
In the above example, the command is selected based on the recognition result obtained as a result of the speech recognition process. However, the command may be selected based on the text obtained as a result of the text input.

Embodiment 1 FIG.
Next, a command processing apparatus that performs the above-described command processing will be described. FIG. 6 is a block diagram showing the configuration of the command processing apparatus according to Embodiment 1 of the present invention.
The command processing device 10 includes a voice input unit 1, a voice recognition processing unit (recognition processing unit) 2, a command selection unit 3, a likelihood correction unit 4, a recognition dictionary database memory 5, a conversion database memory 6, and an operation cost database memory 7. It is configured.

  The voice input unit 1 inputs voice uttered by the user, performs A / D conversion on the voice, and outputs voice data. The voice recognition processing unit 2 inputs the voice data output from the voice input unit 1, refers to the recognition dictionary database memory 5, performs vocabulary recognition on the voice data, and outputs a recognition result. At this time, the speech recognition processing unit 2 obtains a vocabulary that can be expected from the most probable vocabulary as the first candidate, for example, a candidate list in which the vocabulary can be predicted up to the fifth candidate and arranged in a probable order, and a recognition value that represents the probability of each candidate. The likelihood is output as a recognition result (see Reference 2).

Reference 2.
JP-A-60-166997

  The command selection unit 3 inputs the recognition result, refers to the conversion database memory 6, calculates the likelihood, and outputs the intermediate command and the likelihood. At this time, the command selection unit 3 obtains a command that can be predicted as the most likely command as the first candidate, for example, up to the fifth candidate, and outputs a candidate list that is arranged in the most likely order as an intermediate command. The likelihood correction unit 4 inputs an intermediate command and likelihood, refers to the operation cost database memory 7, calculates an average value of the operation cost using the operation cost, and calculates a command having the minimum average operation cost. Output.

  The recognition dictionary database memory 5 is a memory for storing a database of acoustic models (models) and language models (models) used when the speech recognition processing unit 2 calculates the likelihood. The conversion database memory 6 is a memory for storing a database in which correspondences between command names and related words for converting words uttered by the user into commands are described. That is, the conversion database memory 6 is constructed in a table structure that stores a plurality of pieces of conversion information.

  The operation cost database memory 7 represents a hierarchical structure relationship between commands, and is a memory for storing a database of operation costs associated with transitions between commands for the likelihood correction unit 4 to calculate an average value of operation costs. That is, the operation cost database memory 7 is a memory for storing operation costs associated with transitions between commands, as shown in FIG.

Next, the operation of the command processing apparatus according to the first embodiment will be described. FIG. 7 is a flowchart showing the operation of the command processing apparatus according to the first embodiment.
The voice input unit 1 is generally installed in the vicinity of a user who is a speaker, or is held by the user, and the user's voice is input. The voice input unit 1 performs A / D conversion with, for example, 16 kHz sampling and 16 bits on this input, and outputs voice data to the voice recognition processing unit 2 (step ST1).

  The voice recognition processing unit 2 determines a voice section for the voice data input in step ST1 in the same manner as a known method, and the acoustic model / language model of the recognition dictionary database memory 5 is determined for the voice data of the determined voice section. Are matched, and the most probable word / phrase is output as a recognition result to the command selection unit 3 (step ST2).

The command selection unit 3 refers to the conversion database memory 6 based on the recognition result input in step ST2, and outputs the intermediate command and likelihood to the likelihood correction unit 4 (step ST3). FIG. 8 is a diagram illustrating a conversion database used when the command selection unit 3 according to Embodiment 1 calculates the likelihood. Hereinafter, the operation of the command selection unit 3 will be described in detail with reference to FIG.
In FIG. 8, for example, when the user's utterance input to the speech recognition processing unit 2 is “fast route”, the probability value calculated using the acoustic model or language model described in the recognition dictionary is used as the likelihood. Then, “TOLL_FIRST_ROUTE” (toll road priority route search) and “RECOMMENDED_ROUTE” (recommended route search) have a likelihood of 6, “SHORSTEST_ROUTE” (shortest distance search) has a likelihood of 5, and “ROUTE_SEARCH” has a likelihood of 3, Elected as. “” Indicates the recognition result, and “” indicates the command name.

  The likelihood correcting unit 4 refers to the operation cost database memory 7 based on the intermediate command and the likelihood input in step ST3, and outputs a final command (step ST4). The final command C ^ output in step ST4 is expressed as shown in the above-described equation (2). The operation of the likelihood correction unit 4 is as described above based on FIG.

  As described above, according to the first embodiment, the operation cost database memory 7 that stores a database of operation costs associated with transitions between commands and the operation cost database of the operation cost database memory 7 are referred to, and Is used to calculate the average value of the operation costs, and the likelihood correction unit 4 that outputs the command having the smallest average value as the final command is provided. Therefore, the command is selected in consideration of the hierarchical relationship of the commands. It is possible to reduce the labor associated with the screen operation, which is a modification of the selected command.

  In the first embodiment, the average value of the operation cost is described as an example of the evaluation value. However, the expected value of the operation cost and the weight of likelihood are added to the average value of the operation cost represented by the equation (5). A value in consideration may be used as the evaluation value. The same applies to the second embodiment below.

Embodiment 2. FIG.
In the second embodiment, in addition to the configuration of the first embodiment, a configuration is shown in which a specific command is selected and an average value of operation costs is calculated.
FIG. 9 is a block diagram showing a configuration of the command processing device according to the second embodiment, and a command selection unit 8 is additionally provided in the command processing device 10 of the first embodiment. In the following, the same or corresponding parts as those of the command processing apparatus according to the first embodiment are denoted by the same reference numerals as those used in the first embodiment, and the description thereof is omitted or simplified.

The command selection unit 8 inputs the intermediate command selected by the command selection unit 3 and the likelihood, and refers to the operation cost database memory 7, for example, the operation cost with the command having the maximum likelihood among the intermediate commands is constant. The command below the value is set as a selection command, and the selection command and likelihood are output to the likelihood correction unit 4a.
The likelihood correction unit 4a inputs a selection command and likelihood, refers to the operation cost database memory 7, calculates an average value of the operation cost using the operation cost, and a command that minimizes the average value of the operation cost Is output.

Next, the operation of the command processing apparatus according to the second embodiment will be described with reference to the flowchart of FIG. Note that the same reference numerals as those used in FIG. 7 are attached to steps for performing the same processing as that of the command processing apparatus of the first embodiment, and description thereof will be omitted or simplified.
The command selection unit 3 refers to the conversion database memory 6 based on the recognition result input in step ST2, and outputs an intermediate command and likelihood to the command selection unit 8 (step ST11). The command selection unit 8 refers to the operation cost database memory 7 based on the intermediate command and likelihood input from the command selection unit 3 in step ST11, acquires the operation command and likelihood, and outputs the operation command and likelihood to the likelihood correction unit 4a. (Step ST12).

  The likelihood correcting unit 4a refers to the operation cost database memory 7 based on the selection command and the likelihood input in step ST12, and outputs a final command (step ST13). The final command C ^ output in step ST13 is expressed as shown in the above-described equation (2). The operation of the likelihood correcting unit 4a is as described above based on FIG.

  As described above, according to the second embodiment, the operation cost database memory 7 that stores a database of operation costs associated with the transition between commands and the operation cost database of the operation cost database memory 7 are referred to, and the operation cost is stored. Is used to calculate the average value of the operation costs, and the likelihood correction unit 4a that outputs the command having the smallest average value as the final command is provided. Therefore, the command is selected in consideration of the hierarchical relationship of the commands. It is possible to reduce the labor associated with the screen operation, which is a modification of the selected command.

  In addition, according to the second embodiment, the operation cost database memory 7 is referred to and the command selection unit 8 that selects a command that satisfies a predetermined condition is provided. It is possible to prevent the upper command from being selected on the operation hierarchy by ignoring. Furthermore, the calculation amount for calculating the evaluation value can be suppressed.

  DESCRIPTION OF SYMBOLS 1 Voice input part, 2 Voice recognition process part, 3 Command selection part, 4, 4a Likelihood correction part, 5 Recognition dictionary database memory, 6 Conversion database memory, 7 Operation cost database memory, 8 Command selection part, 10 Command processing apparatus .

Claims (2)

A recognition processing unit that performs vocabulary recognition on input data input by a user and outputs a probable vocabulary as a recognition result;
For the recognition result, when a result appears according to the model, a likelihood that is a likelihood that the model is estimated by looking from the input data is calculated, and an intermediate command that is a probable command candidate is calculated based on the likelihood. A command selection section to select;
An operation cost database memory that defines an operation when the user makes an operation error as an operation cost, and stores an operation cost associated with an operation between commands,
A command processing apparatus comprising: a likelihood correcting unit that selects a command that minimizes an evaluation value calculated based on the operation cost from the intermediate command.   The command processing apparatus according to claim 1, further comprising a command selection unit that refers to the operation cost and the likelihood of the intermediate command and selects an intermediate command to be output to the likelihood correction unit from the intermediate command.

Images