JP6636379B2 - Identifier construction apparatus, method and program - Google Patents

Description

  The present invention relates to a technology for constructing a discriminator for discriminating the presence or absence of a speech intention assumed by a dialog system.

  In the dialogue system, only the utterance of the system user is cut out using the utterance section detection technology (for example, see Non-Patent Document 1) and sent to the dialogue system. By extracting only the utterance of the system user with high accuracy, a smooth dialogue between the system and the user is realized.

Masayoshi Fujimoto, “Basics of Voice Activity Detection and Recent Research Trends”, IEICE Technical Report.,

  However, even if the utterance section can be correctly detected by the utterance section detection technique, the detected utterance may not be directed to the dialogue system. For example, there is a case where users in front of the interactive system are chatting with each other. In this case, the dialogue system may respond to an utterance not directed to the dialogue system.

  An object of the present invention is to construct a classifier for constructing a classifier for discriminating presence / absence of a speech intention assumed by a dialog system so as not to respond to an utterance not directed to the dialog system. It is to provide an apparatus, a method and a program.

According to one embodiment of the present invention, there is provided a discriminator construction apparatus that detects a speech section in an input speech signal, performs speech recognition on the detected speech section, and performs a speech recognition result and a speech recognition result. A speech recognition unit that obtains the reliability of the speech recognition unit; a speech intention determination unit that determines whether the speech section includes the speech intention assumed by the dialogue system based on the speech recognition result; a first one sound voice interval is a speech segment that contains an identification unit for identifying reliability and a second speech section is a speech section is not included is the utterance intention less than a predetermined threshold value, the first A feature amount calculation unit that calculates the feature amount of each of the voice section and the second voice section, and a discriminator construction that constructs a discriminator for discriminating the presence / absence of an utterance intention to the interactive system using the calculated feature amounts And a part There.

  It is possible to construct a discriminator for discriminating the presence / absence of a speech intention assumed by the dialogue system. By using this discriminator, a dialog system that does not respond to an utterance that is not directed to the dialog system can be constructed.

FIG. 2 is a block diagram illustrating an example of a classifier construction apparatus. 5 is a flowchart for explaining an example of a classifier construction method. 5 is a flowchart for explaining an example of processing of the identification unit 23. FIG. 4 is a diagram for explaining an example of processing of the voice section detection unit 1. FIG. 9 is a diagram for explaining an example of processing of the identification unit 23.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[Identifier construction apparatus and method]
As illustrated in FIG. 1, the discriminator construction apparatus includes a voice section detection unit 1, a dialogue system unit 2, a feature amount calculation unit 3, a learning data storage unit 4, and a discriminator construction unit 5. The dialogue system unit 2 includes, for example, a voice recognition unit 21, an utterance intention determination unit 22, and an identification unit 23. The utterance intention determination unit 22 includes, for example, an utterance intention corpus storage unit 211.

  The discriminator construction method is realized by each unit of the discriminator construction apparatus executing the processing of FIG. 2 and steps S1 to S5 described below.

<Speech section detection unit 1>
A voice signal is input to the voice section detection unit 1.

  The voice section detector 1 detects a voice section in the input voice signal (step S1). The detected voice section is output to the voice recognition unit 21 of the dialogue system unit 2.

  An existing technique such as Non-Patent Document 1 may be used as a method for detecting a voice section.

  As illustrated in FIG. 4, the input voice signal includes a voice section. The voice section detection unit 1 detects a voice section in the input voice signal, and outputs a voice signal of the detected voice section. In the example of FIG. 4, two voice sections are detected, and respective voice signals of the two detected voice sections are output to the voice recognition unit 21.

<Speech recognition unit 21>
The speech section detected by the speech section detection unit 1 is input to the speech recognition unit 21.

  The voice recognition unit 21 performs voice recognition on the voice section detected by the voice section detection unit 1, and obtains the voice recognition result and the reliability of the voice recognition result (step S21). The obtained speech recognition result is output to the utterance intention determination unit 22. The reliability of the obtained speech recognition result is output to the identification unit 23.

  An existing technology may be used for the voice recognition method.

<Utterance intention determination unit 22>
The speech recognition result obtained by the speech recognition unit 21 is input to the speech intention determination unit 22.

  The utterance intention determination unit 22 determines whether or not the utterance intended by the dialog system is included in the above-mentioned voice section based on the speech recognition result obtained by the speech recognition unit 21 (step S22). The determination result is output to the identification unit 23.

  For example, in the case of an interactive system for sightseeing guidance, if a keyword or expression related to a tourist attraction, a name of a sightseeing value, a meal, a transportation system, etc. is detected from a speech recognition result, the utterance intention assumed by the interactive system is included. It is determined that there is. On the other hand, when no keyword or expression leading to the utterance intention is included, it is determined that the utterance intention is not included and the utterance intention cannot be detected. The information on keywords and expressions for understanding these utterance intentions is generally available as a utterance intention corpus, and is prepared in advance by a dialog system builder. For example, it is stored in the utterance intention corpus storage unit 221 in advance. In this case, the utterance intention determination unit 22 performs the above-described determination using the keyword or expression read from the utterance intention corpus storage unit 221.

<Identifier 23>
The reliability of the voice recognition result obtained by the voice recognition unit 21 and the determination result obtained by the speech intention determination unit 22 are input to the identification unit 23.

  The identification unit 23 includes a first voiced section that is a voice section whose reliability is equal to or more than a predetermined threshold value and includes a speech intention, and a voice section whose reliability is equal to or less than a predetermined threshold value and does not include the speech intention. The voice section is identified as the second voice section (step S23). The identification result is output to the feature value calculation unit 3.

  For example, a flag for identifying whether a voice section is a first voice section or a second voice section is added. Here, the flag to be given may be a numerical expression such as giving "0" to the first voice section and "1" to the second latter. In FIG. 5, the flag of the voice section U1 is set to 0, the flag of the voice section U2 is set to 1, the flag of the voice section U3 is set to 1, the flag of the voice section U4 is set to 0, and the flag of the voice section U5 is set to 0. The flag of the section U6 is set to 0.

  The first voice section can be assumed to be a voice correctly input to the dialogue system. It can be assumed that the second voice section is not a voice that was correctly input to the dialog system.

  For example, as illustrated in FIG. 3, the identification unit 23 determines whether or not the reliability is equal to or more than ε and the utterance intention is included in the voice section, with a predetermined threshold being ε. When the reliability is equal to or more than ε and includes the utterance intention, the identification unit 23 sets the voice section as the first voice section. In other cases, the identification unit 23 determines whether the reliability is equal to or less than ε and does not include an utterance intention. When the reliability is equal to or less than ε and does not include a speech intention, the identification unit 23 sets the voice section as a second voice section. In other cases, the identification unit 23 outputs the audio signal of the audio section to the utterance intention determination unit 22.

  It should be noted that the identification unit 23 also determines that the voice section is the second voice even when the reliability is equal to or more than ε and does not include the utterance intention, and when the reliability is equal to or less than ε and includes the utterance intention. It may be a section.

<Feature amount calculation unit 3>
The voice signal of the first voice section and the voice signal of the second voice section are input to the feature amount calculation unit 3.

  The feature amount calculation unit 3 calculates the feature amounts of the first voice section and the second voice section using the voice signal of the first voice section and the voice signal of the second voice section (step S3). The calculated feature amount is stored in the learning data storage unit 4.

  As the characteristic amount, at least one of the average of power and pitch, the standard deviation, their Δ component, and the voice section length can be used. What is necessary is just to calculate these characteristic amounts using an existing technique.

<Learning data storage unit 4>
The learning data storage unit 4 stores the feature amounts of the first voice section and the second voice section.

  Until the feature amounts of the first and second voice sections are accumulated for a predetermined time T, the processing from step S1 to step 3 is repeatedly executed. T is a predetermined time length, for example, T can be set to about 0.5 [h].

<Identifier construction unit 5>
The feature amount of each of the first voice section and the second voice section is input to the discriminator construction unit 5. The discriminator construction unit 5 reads the respective feature amounts of the first speech section and the second speech section from, for example, the learning data storage unit 4.

  The discriminator construction unit 5 constructs a discriminator for discriminating the presence / absence of an utterance intention to the dialogue system using the calculated feature amount (step S5).

  Existing technology may be used to construct the classifier. For example, a recognition model such as a support vector machine may be used.

  By using the discriminator constructed by the discriminator construction apparatus and method, a system capable of determining an intended utterance and an unintended utterance can be provided. As a result, unnatural dialog responses can be reduced, and the quality of the dialog system can be improved.

[Program and recording medium]
When each process in the classifier construction device is realized by a computer, the processing contents of the functions that the classifier construction device should have are described by a program. Then, by executing this program on a computer, each processing is realized on the computer.

  A program describing this processing content can be recorded on a computer-readable recording medium. As a computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  Further, each processing means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

[Modification]
The processes described in the discriminator construction apparatus are not only executed in chronological order according to the order of description, but may be executed in parallel or individually according to the processing capability of the apparatus that executes the processing or as necessary.

  In addition, it goes without saying that changes can be made as appropriate without departing from the spirit of the present invention.

Claims (3)

A voice section detection unit that detects a voice section in the input voice signal;
A voice recognition unit that performs voice recognition on the detected voice section and obtains a voice recognition result and the reliability of the voice recognition result;
An utterance intention determining unit that determines whether or not the utterance intention assumed by the dialogue system is included in the voice section based on the voice recognition result;
In the reliability is the first one sound voice interval is a speech segment that contains is the utterance intention not less than the predetermined threshold value, the reliability speech interval does not contain the speech intention is equal to or less than a predetermined threshold value An identification unit that identifies a certain second voice section;
A feature amount calculation unit that calculates a feature amount of each of the first voice section and the second voice section,
A classifier constructing unit configured to construct a classifier for identifying the presence or absence of an utterance intention with respect to the dialogue system using the calculated feature amount;
A classifier construction apparatus including: A voice section detection step of detecting a voice section in the input voice signal,
A voice recognition step of performing voice recognition on the detected voice section to obtain a voice recognition result and the reliability of the voice recognition result;
An utterance intention determining unit that determines whether or not the utterance intended by the dialogue system is included in the voice section based on the result of the voice recognition,
Identifying portion, the reliability and the second one sound voice interval is a speech segment that contains is the utterance intention not less than the predetermined threshold value, the reliability contains the utterance intention is equal to or less than a predetermined threshold value An identification step for identifying a second voice section that is not a voice section;
A feature value calculating step of calculating a feature value of each of the first voice section and the second voice section,
A classifier constructing unit configured to construct a classifier for identifying presence / absence of an utterance intention with respect to the dialogue system using the calculated feature amount;
A classifier construction method including:   A program for causing a computer to function as each unit of the discriminator construction apparatus according to claim 1.

Images