JP2008216618A - Speech discrimination device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To discriminate a speech with high precision. <P>SOLUTION: A speech discrimination device which discriminates a speech and a car navigator which performs various operations are connected through a network etc., to communicate with each other. Further, the speech discrimination device has a microphone receiving various speeches. In such a configurations, the speech discrimination device specifies an utterance section showing a section where an utterance is made from the power of an input speech signal to calculate an acoustic feature quantity showing a feature of the speech signal from the power and the pitch of the speech signal in the specified utterance section, specifies the section before and/or the section after the specified utterance section to calculate an acoustic feature quantity from the power and the pitch of the speech signal in the specified section, and discriminates whether the input speech signal is a system request to request a connected processor which performs various processings to execute processings from the acoustic feature quantity in the calculated utterance section and acoustic feature quantities in the sections before and after the calculated utterance section. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a voice discrimination device that discriminates whether an input voice signal is a request to a system or a chat.

  Conventionally, voice interfaces have been put into practical use in various fields. In particular, the application to the operation of devices that are difficult to use with hands, such as communication with robots and car navigation systems, is remarkable.

  However, currently used speech recognition systems (speech discrimination systems) cannot accurately discriminate whether the input speech is utterance to the system or the surrounding chat, so that unintended operations can be generated without using switches. It was. In other words, since the voice recognition system is inaccurate only by voice recognition, the inaccuracy is compensated by detecting additional information such as the operation of a switch.

  On the other hand, there is a method of changing prosodic features and language features with the user consciousness, such as a voice spotter (reference: Goto et al, ICSLP, 8, 1522-1536, 2004). Feels unnatural in their utterances, which places a heavy burden on the user. Therefore, methods that use the acoustic features of natural utterances (references: Yamada, SIG-SLP-61 (2), 7-12, 2006-5) and speech recognition results are used as methods that place less burden on the user. (Reference: Sako, SIG-SLP-64, 19-24, 2006-12) has been proposed.

  For example, in Non-Patent Document 1, an acoustic feature amount indicating the characteristics of an audio signal is calculated from the power and pitch of the audio signal in the uttered section, and audio discrimination is performed as to whether the audio signal is a request utterance to the system or chat. Techniques for realizing it are disclosed. Specifically, voice identification is performed by calculating eight dimensions of the difference between the average, standard deviation, maximum, maximum and minimum values from the input speech signal as acoustic feature quantities, and further, HMM (Hidden Markov Model) Speech recognition is performed from the likelihood of the input utterance by the acoustic model based on, and the speech discrimination between the requested utterance to the system and the chat is performed using the result of the speech recognition and the speech identification.

Natsuki Sugimoto, Norihide Kitaoka, Seiichi Nakagawa, Department of Information Engineering, Toyohashi University of Technology, “Distinguishing between-system utterances and interpersonal utterances using speech features”, 2006 IEICE General Conference

  By the way, the above-described conventional technique has a problem that the accuracy of voice discrimination is low. Specifically, only in the uttered section, the likelihood is calculated by using the 8 dimensions of the difference between the average, the standard deviation, the maximum, the maximum value, and the minimum value for each of the power and the pitch as an acoustic feature, Since recognition is performed, it is easily affected by noise and the like, and the accuracy of speech recognition is not good. Then, using the result recognized in this manner, the voice discrimination between the requested utterance to the system and the chat is performed, so that the voice discrimination accuracy is also deteriorated.

  Accordingly, the present invention has been made in order to solve the above-described problems of the prior art, and an object thereof is to provide a voice discrimination device capable of realizing voice discrimination with high accuracy.

  In order to solve the above-described problems and achieve the object, the invention according to claim 1 is characterized in that an utterance section specifying means for specifying an utterance section indicating an utterance section from an input voice signal, and the utterance section specifying means. First acoustic feature quantity calculating means for calculating an acoustic feature quantity indicating the characteristics of the voice signal from the power and pitch of the voice signal in the utterance section specified by, and the utterance section specified by the utterance section specifying means A second acoustic feature quantity calculating unit that identifies a preceding section and / or a subsequent section and calculates the acoustic feature quantity from the power and pitch of the audio signal in the identified section; and the first acoustic feature quantity calculation A process for executing various processes in which the input audio signal is connected from the acoustic feature quantity calculated by the means and the acoustic feature quantity calculated by the second acoustic feature quantity calculation means. A voice discriminating means for discriminating whether or not the system request for requesting execution of processing to the device, characterized by comprising a.

  Further, in the invention according to claim 2, in the above invention, the first acoustic feature amount calculating means is an average value of the power and pitch of the audio signal in the utterance section specified by the utterance section specifying means. Any one or more of the standard deviation value, the maximum value, and the difference between the maximum value and the minimum value is calculated as the acoustic feature amount.

  According to a third aspect of the present invention, in the above invention, the second acoustic feature quantity calculating means specifies a section before and / or a section after the utterance section specified by the utterance section specifying means. Calculating one or more of an average value, a standard deviation value, a maximum value, and a difference between the maximum value and the minimum value of the power and pitch of the audio signal in the specified section as an acoustic feature amount. And

  According to a fourth aspect of the present invention, in the above invention, the language feature for performing speech recognition for extracting the content of the utterance as character data from the input speech signal and generating a language feature from the speech-recognized character data. A volume generation unit, wherein the voice determination unit includes the acoustic feature amount calculated by the first acoustic feature amount calculation unit, the acoustic feature amount calculated by the second acoustic feature amount calculation unit, Whether or not the input audio signal is a system request for requesting execution of processing to a connected processing device that executes various processes based on the language feature generated by the language feature generating means. It is characterized by discriminating.

  According to the first aspect of the present invention, an utterance section indicating an utterance section is specified from the input voice signal, and an acoustic feature amount indicating the characteristics of the voice signal from the power and pitch of the voice signal in the specified utterance section. Is calculated, the section before and / or after the specified utterance section is specified, the acoustic feature amount is calculated from the power and pitch of the voice signal in the specified section, and the acoustic feature of the calculated utterance section is calculated. The processing unit is requested to execute processing for various processes to which the input audio signal is connected, based on the amount and the acoustic feature amount of the section before and / or after the calculated speech section Therefore, it is possible to realize highly accurate voice discrimination.

  For example, if it is a system request, the acoustic feature amount is calculated from each of the utterance section of the utterance section and the section before and after the utterance based on the fact that there is often silence in the section before and after the utterance section. Since it is determined whether or not the input signal is a system request, it is possible to calculate and distinguish many feature amounts compared to the case where the acoustic feature amount is calculated and determined only from the utterance period. It is possible to realize high voice discrimination.

  In addition, low-cost and sufficiently accurate without requiring an expensive device for recognizing the input speech signal and calculating a language feature that can be expected to be highly accurate speech discrimination Voice discrimination can be realized.

  According to the invention of claim 2, any one of an average value, a standard deviation value, a maximum value, and a difference between the maximum value and the minimum value of the power and pitch of the voice signal in the specified speech section Alternatively, since a plurality are calculated as acoustic feature amounts, it is possible to realize more accurate speech discrimination.

  For example, by using all of “average value, standard deviation value, maximum value, difference between maximum value and minimum value” as acoustic features, the features can be clarified, resulting in more accurate voice discrimination. It is possible to realize. Further, by excluding parameters that do not clearly show features, it is possible to reduce the processing load of the system and increase the processing speed.

  According to the invention of claim 3, the section before the specified speech section and / or the section after the specified section is specified, and the average value and standard deviation value of the power and pitch of the audio signal in the specified section are specified. Since one or more of the maximum value and the difference between the maximum value and the minimum value is calculated as the acoustic feature amount, more accurate speech discrimination can be performed compared to the case where the acoustic feature amount of only the utterance section is used. It is possible to realize.

  For example, by using all of the “average value, standard deviation value, maximum value, difference between maximum value and minimum value” of the power and pitch as the acoustic features in each of the sections before and after the utterance section, the maximum in total Since 24 dimensions can be used as the acoustic feature amount and the feature can be further clarified, it is possible to realize voice discrimination with higher accuracy. In addition, as a result of being able to calculate only the previous section and the rear section as the acoustic feature amount as necessary, by excluding parameters that do not show the feature so clearly, the processing load of the system can be reduced, It is possible to increase the processing speed.

  According to the invention of claim 4, speech recognition is performed to extract the content of the utterance from the input speech signal as character data, a language feature is generated from the speech-recognized character data, and the calculated speech interval From the acoustic feature amount, the calculated acoustic feature amount before and / or after the utterance interval, and the generated language feature amount, the input speech signal executes various processes to be connected. Since it is determined whether or not it is a system request for requesting the processing apparatus to execute processing, it is possible to realize more accurate voice determination.

  For example, by using an expensive function or product that can calculate language features, speech discrimination with higher accuracy is possible compared to speech discrimination using only acoustic features in the utterance section and the preceding and following sections. Can be realized.

  Exemplary embodiments of a speech discrimination device, a speech discrimination method, and a speech discrimination program according to the present invention will be described below in detail with reference to the accompanying drawings. In the following, the main terms used in this embodiment, the outline and features of the speech discrimination apparatus according to this embodiment, the configuration of the speech discrimination apparatus and the flow of processing will be described in order, and finally various modifications to this embodiment will be described. An example will be described.

[Explanation of terms]
First, main terms used in this embodiment will be described. The “speech discrimination device (corresponding to“ speech discrimination device ”recited in the claims”) used in the present embodiment is an input voice signal that is a request to the system or other chat. It is a device that determines whether or not. As a specific example, the “voice discrimination device” is connected to a car navigation system (hereinafter referred to as “car navigation system”), receives voices spoken from a driver's seat passenger, passengers, etc., and the spoken contents. Is a system request or chat. For example, the “voice discrimination device” outputs an instruction to the car navigation system to implement the spoken content when it determines that the input voice signal is a system request, and outputs an instruction when it is determined as chat. Discard the utterance without doing it.

  The system request indicates, for example, various operations of the car navigation such as “activate the car navigation”, “change the car navigation display to an enlarged view”, “change the destination”, etc. Includes explanations of car navigation operations and daily conversations that are not related to car navigation operations, such as “I will explain the operation of the car navigation system”, “You can change the display by pressing this button”, “The volume of the TV is this button”. In the present embodiment, the case where the “voice discrimination device” discriminates whether it is a system request to a car navigation system or a chat will be described. However, the present invention is not limited to this and is not limited to a car navigation system. It can be applied to various fields such as system requirements for robots.

  Further, the “accuracy rate” used in the present embodiment is “accuracy rate” when “A” and “B” are set to be determined to be valid, and “C” is set to be determined not to be valid. = A / (A + C) ", that is, an accuracy index for the occurrence of false detection, and" recall rate "is calculated as" recall rate = A / (A + B) " That is, it is an accuracy index for non-detection. The F value is the harmonic average of precision and recall, and the value increases when both precision and recall are high in a balanced manner. In other words, the precision and recall are integrated. It is an index to be evaluated.

[Outline and features of voice identification device]
Next, the outline and features of the speech discrimination apparatus according to the first embodiment will be described with reference to FIG. FIG. 1 is a system configuration diagram illustrating an overall configuration of a system including the voice discrimination device according to the first embodiment.

  As shown in FIG. 1, in this system, a voice discrimination device that discriminates voice and a car navigation system that executes various operations are connected to each other via a network or the like. The voice discrimination device also includes a microphone that accepts various voices. In addition, the car navigation system executes the instruction content input from the voice discrimination device. For example, when an instruction indicating “change the display of the car navigation to an enlarged view” is received, the car navigation changes the figure displayed according to the received instruction content to the enlarged view.

  In such a configuration, as described above, the speech discrimination apparatus according to the first embodiment has an outline of discriminating whether the input speech signal is a request to the system or other chat. In particular, the main feature is that it is possible to realize highly accurate voice discrimination.

  The main feature will be described in detail. The voice discrimination device receives spoken voice via a microphone (see (1) in FIG. 1). To give specific examples, the voice discrimination device utters “change the car navigation display to an enlarged view”, “change the destination”, “ An audio signal such as “The display changes” is received via the microphone.

  Then, the speech discriminating device specifies an utterance interval indicating an utterance interval from the input audio signal, and averages, standard deviation values, and maximum values of the power and pitch of the audio signal in the specified utterance interval Then, any one or more of the difference between the maximum value and the minimum value is calculated as the acoustic feature amount (see (2) in FIG. 1). Specifically, in the above example, the voice discrimination device determines in advance the power of the input voice signal when the voice signal is input as “change the car navigation display to an enlarged view”. A predetermined threshold that is easy to handle in voice discrimination is set, margins are extracted before and after this threshold, and the extracted section is specified as an utterance section. Then, the speech discriminating apparatus determines that the power of the speech signal in the specified speech section is “average value, standard deviation value, maximum value, difference between maximum value and minimum value” and the pitch of the speech signal “average value, The standard deviation value, the maximum value, and the difference between the maximum value and the minimum value are calculated, and these are used as acoustic feature amounts.

  Subsequently, the speech discriminating apparatus identifies the preceding section and / or the following section of the identified speech section, and the average value, standard deviation value, and maximum value of the power and pitch of the speech signal in the identified section. Then, any one or a plurality of differences between the maximum value and the minimum value are calculated as acoustic feature amounts (see (3) and (4) in FIG. 1). More specifically, in the example described above, the speech discriminating apparatus specifies the first section of the specified utterance section to 0.7 seconds before as the previous section of the utterance section (section A), and the end of the specified utterance section. To 0.7 seconds later is identified as the latter section (section B) of the utterance section. Then, the speech discriminating apparatus performs “the difference between the average value, the standard deviation value, the maximum value, the maximum value and the minimum value” of the power of the sound signal in the specified section A and the power of the sound signal in the specified section B. Of “average value, standard deviation value, maximum value, and difference between maximum value and minimum value” are calculated as acoustic feature values.

  Then, the speech discrimination device is a processing device that executes various processes in which a speech signal input from an acoustic feature amount calculated from an utterance interval and an acoustic feature amount calculated from an interval before and after the utterance interval is connected. It is determined whether or not the request is a system request for requesting execution of processing (see (5) in FIG. 1). More specifically, in the above example, the SVM (Support Vector Machine) of the voice discrimination device calculates the acoustic feature amount “average value, standard deviation value, maximum value, maximum value of power and pitch” calculated from the speech section. The difference between the value and the minimum value ”and the acoustic feature amount calculated from the section A“ average value, standard deviation value, maximum value, difference between the maximum value and the minimum value of power and pitch ”and the section B The order of the trees generated by the decision tree learning tool C4.5 with the acoustic feature quantity “average value, standard deviation value, maximum value, difference between maximum value and minimum value of power and pitch” calculated from (Reference: CWHsu et al, http://www.csie.ntu.edu.tw/~cjlin/papers/guide.pdf), and the weight for each feature (which feature is important) Is calculated). Then, the SVM determines whether or not the audio signal input from the calculated weight for each feature amount is a system request. If the voice signal input here is “change the car navigation display to an enlarged view”, the voice discrimination device determines that the voice signal is a system request.

  Thereafter, the voice discrimination device outputs an instruction to the car navigation system so as to execute the operation content instructed by the voice signal (see (6) in FIG. 1). Specifically, in the above example, the voice discrimination device performs the operation content “change display to enlarged view” instructed by the voice signal “change the display of car navigation to an enlarged view”. Outputs instructions. Then, the car navigation system processes the instructed content. When it is determined that the voice signal is not a system request but a chat, the voice discrimination device discards the voice signal without outputting an instruction to the car navigation system.

  As described above, the speech discriminating apparatus according to the first embodiment calculates the 8-dimensional acoustic feature amount from each of the utterance section and the section before and after the utterance, that is, calculates the total 24-dimensional acoustic feature amount, As a result of discriminating whether or not is a system request, it is possible to realize highly accurate speech discrimination as the main feature described above.

[Configuration of voice discrimination device]
Next, the configuration of the voice discrimination device shown in FIG. 1 will be described with reference to FIG. FIG. 2 is a block diagram illustrating the configuration of the speech discrimination apparatus according to the first embodiment. As shown in FIG. 2, the voice discrimination device 10 includes a communication control I / F unit 11, an input unit 12, a storage unit 13, and a control unit 20.

  The communication control I / F unit 11 controls communication related to various types of information exchanged with a connected device. Specifically, in the above example, the communication control I / F unit 11 determines that the input voice signal is a system request by the control unit 20 to be described later, and transmits an instruction to the system to the car navigation system. To do.

  The input unit 12 includes a microphone and receives an audio signal uttered by a user. For example, the input unit 12 utters “changes the car navigation display to an enlarged view”, “changes the destination”, or “presses this button to change the display” spoken by the user. A voice signal such as a system request, chat, conversation or the like is received, and the voice signal is stored in the input signal DB 14 via an utterance section specifying unit 21 described later.

  The storage unit 13 stores data and programs necessary for various processes performed by the control unit 20 and includes an input signal DB 14 and a discrimination result DB 15 that are particularly closely related to the present invention. In addition, the storage unit 13 stores a predetermined threshold that is easy to handle by a predetermined voice discrimination used for specifying the utterance section, a predetermined number of seconds used for specifying the sections before and after the utterance section, and the like. Remember. The information including various data and parameters can be arbitrarily changed unless otherwise specified.

  The input signal DB 14 temporarily stores the audio signal input from the input unit 12. As a specific example, the input signal DB 14 stores an audio signal as shown in FIG. The determination result DB 15 stores the result determined by the SVM 24 described later. As a specific example, as shown in FIG. 4, the discrimination result DB 15 includes “system request” that is a voice signal discriminated as a system request and “chat” that is a voice signal discriminated as a chat. “Activate the car navigation system, increase the volume, explain the operation of the car navigation system, press this button to change the display”, etc. 3 is a diagram illustrating an example of an input audio signal stored in the input signal DB, and FIG. 4 is a diagram illustrating an example of information stored in the determination result DB. In addition, information including various data and parameters can be arbitrarily changed unless otherwise specified.

  The control unit 20 has a control program such as an OS (Operating System), a program that defines various processing procedures, and an internal memory for storing necessary data. In particular, the control unit 20 is closely related to the present invention. The speech section specifying unit 21, the first acoustic feature calculation unit 22, the second acoustic feature calculation unit 23, the SVM 24, and the instruction output unit 25 are provided, and various processes are executed by these.

  The utterance interval specifying unit 21 specifies an utterance interval indicating an utterance interval from the power of the input voice signal. Specifically, in the example described above, the speech segment specifying unit 21 determines in advance the power of the input voice signal when the voice signal “Change the car navigation display to an enlarged view” is input. A predetermined threshold value that is easy to handle in the determined voice discrimination is set, margins are extracted before and after the threshold value, and the extracted section is specified as the speech section. The utterance section specifying unit 21 corresponds to “speech section specifying means” recited in the claims.

  The first acoustic feature calculation unit 22 calculates the average value, the standard deviation value, the maximum value, and the difference between the maximum value and the minimum value of the power and pitch of the speech signal in the utterance section specified by the utterance section specifying unit 21. Any one or more of them are calculated as acoustic feature quantities. More specifically, in the above example, the first acoustic feature calculation unit 22 calculates the “average value, standard deviation value, maximum value, maximum value” of the power of the voice signal in the speech segment specified by the speech segment specifying unit 21. The “difference from the minimum value” and the “average value, standard deviation value, maximum value, difference between the maximum value and minimum value” of the pitch of the audio signal are calculated as acoustic feature amounts, and these are described later as SVM 24. Output to. The first acoustic feature calculation unit 22 corresponds to “first acoustic feature amount calculation means” described in the claims. The calculation of “average value, standard deviation value, maximum value, difference between maximum value and minimum value” can be performed using a known method.

  The second acoustic feature calculation unit 23 specifies a section before and / or after the utterance section specified by the utterance section specifying unit 21, and calculates an acoustic feature amount from the power and pitch of the audio signal in the specified section. calculate. Specifically, the second acoustic feature calculation unit 23 refers to the number of seconds stored in the storage unit 13 and the input voice signal, and the second acoustic feature calculation unit 23 refers to the first speech segment specified. From 0.7 to 0.7 seconds before is specified as the previous section of the utterance section (section A), and the last section from the end of the specified utterance section is specified as the subsequent section of the utterance section (section B). Then, the speech discriminating apparatus performs “the difference between the average value, the standard deviation value, the maximum value, the maximum value and the minimum value” of the power of the sound signal in the specified section A and the power of the sound signal in the specified section B. “Average value, standard deviation value, maximum value, difference between maximum value and minimum value” are calculated as acoustic feature amounts, and these are output to the SVM 24 described later. The 0.7 second used here is merely an example, and is not limited. The second acoustic feature calculation unit 23 corresponds to “second acoustic feature amount calculation means” recited in the claims.

  The SVM 24 uses a sound feature amount calculated from the utterance section and a sound feature amount calculated from the sections before and after the utterance section to a processing device that executes various processes to which the input audio signal is connected. It is determined whether it is a system request for requesting execution of processing. Specifically, the SVM 24 sets an identification boundary at a position where the Eugrid distance becomes the maximum with reference to the support vector closest to the other class in the data, and based on the boundary Determine whether it is a system request or chat.

  For example, the SVM 24 calculates the acoustic feature amount “average value, standard deviation value, maximum value, difference between maximum value and minimum value of power and pitch” calculated from the utterance section and the acoustic feature calculated from the section A. Feature value “average value, standard deviation value, maximum value, difference between maximum value and minimum value of power and pitch” and acoustic feature value calculated from section B “average value of power and pitch The standard deviation value, the maximum value, and the difference between the maximum value and the minimum value ”are scaled based on the order of the trees generated by the decision tree learning tool C4.5, and the weight for each feature amount is calculated. Then, as shown in FIG. 5, the SVM 24 distributes the respective weights and provides a boundary so that the margin is maximized. FIG. 5 is a diagram for explaining request determination by the SVM.

  After that, the SVM 24 determines that the input audio signal is a system request when the calculated number of weights for each feature amount is higher than the lower side of the provided boundary, and determines the audio signal as the determination result DB 15. The instruction output unit 25 is instructed to output an instruction corresponding to the system request. On the other hand, when many are located below, the SVM 24 determines that the input audio signal is chat, and stores the audio signal in the determination result DB 15. The SVM 24 corresponds to “voice discrimination means” described in the claims.

  The instruction output unit 25 outputs an operation instruction related to the system to a connected device such as a car navigation system. Specifically, in the above example, when the instruction output unit 25 receives a signal to instruct a system request from the SVM 24, the instruction output unit 25 corresponds to the input voice signal “change the car navigation display to an enlarged view”. An instruction is output to the car navigation system so as to perform the instruction operation “change display to enlarged view”. The car navigation system that has received such an instruction implements the instruction content.

[Processing by voice discrimination device]
Next, processing performed by the voice discrimination device will be described with reference to FIG. FIG. 6 is a flowchart illustrating the flow of the speech discrimination process in the speech discrimination apparatus according to the first embodiment.

  As shown in FIG. 6, when a voice is input (Yes at Step S <b> 601), the utterance section specifying unit 21 of the voice discrimination device 10 specifies a utterance section indicating a utterance section from the power of the input voice signal. (Step S602).

  Subsequently, the first acoustic feature calculation unit 22 calculates the average value, the standard deviation value, the maximum value, the maximum value, and the minimum value of the power and pitch of the voice signal in the utterance section specified by the utterance section specifying unit 21. Any one or more of these differences are calculated as acoustic feature amounts, and the calculated acoustic feature amounts are output to the SVM 24 (step S603).

  And the 2nd acoustic feature calculation part 23 specifies the area before the utterance area specified by the utterance area specific | specification part 21, and the back area (step S604), From the power and pitch of the audio | voice signal in the specified area. The acoustic feature amount is calculated, and the calculated acoustic feature amount is output to the SVM 24 (step S605).

  After that, the SVM 24 performs processing for executing various processes to which the audio signal input from the acoustic feature amount calculated from the speech section and the acoustic feature amount calculated from the sections before and after the speech section is connected. In step S606, it is determined whether the request is a system request for executing the process.

  If it is determined that the request is a system request (Yes at step S607), the SVM 24 instructs the instruction output unit 25 to output an instruction corresponding to the system request, and the instruction output unit 25 performs an instruction operation. In this way, an instruction is output to the car navigation system (step S608). Subsequently, the SVM 24 stores the input audio signal and the discrimination result in the discrimination result DB 15 (step S609).

  On the other hand, when it is determined that the request is not a system request (No at Step S607), the SVM 24 stores the input audio signal and the determination result in the determination result DB 15 (Step S609).

[Effects of Example 1]
As described above, according to the first embodiment, an utterance period indicating an utterance period is specified from the power of the input audio signal, and the characteristics of the audio signal are determined from the power and pitch of the audio signal in the specified utterance period. Calculating the acoustic feature amount to be indicated, identifying the preceding and / or following segment of the identified utterance segment, calculating the acoustic feature amount from the power and pitch of the audio signal in the identified segment, and calculating the utterance From the acoustic feature value of the section and the calculated acoustic feature values of the preceding and following sections, the input audio signal is a system request that requests the processing device that executes various processes to be connected to execute the process. Therefore, it is possible to realize highly accurate voice discrimination.

  For example, the acoustic feature quantity is calculated from each of the utterance section and the sections before and after the utterance, and it is determined whether or not the input signal is a system request. Therefore, the acoustic feature quantity is calculated only from the utterance section. Compared to the case, more feature amounts can be calculated and discriminated, and as a result, highly accurate voice discrimination can be realized.

  In addition, low-cost and sufficiently accurate without requiring an expensive device for recognizing the input speech signal and calculating a language feature that can be expected to be highly accurate speech discrimination Voice discrimination can be realized.

  Further, according to the first embodiment, any one or more of the average value, standard deviation value, maximum value, and difference between the maximum value and the minimum value of the power and pitch of the audio signal in the specified speech section Thus, it is possible to realize more accurate voice discrimination.

  For example, by using all of “average value, standard deviation value, maximum value, difference between maximum value and minimum value” as acoustic features, the features can be clarified, resulting in more accurate voice discrimination. It is possible to realize. Further, by excluding parameters that do not clearly show features, it is possible to reduce the processing load of the system and increase the processing speed.

  Further, according to the first embodiment, the section before and / or after the identified speech section is identified, and the “average value, standard deviation value, One or more of “maximum value, difference between maximum value and minimum value” is calculated as an acoustic feature amount. This makes speech discrimination more accurate than when using acoustic feature amounts for only the speech segment. Can be realized.

  For example, by using all of “average value, standard deviation value, maximum value, difference between maximum value and minimum value” as the acoustic features in the sections before and after the utterance section, a total of up to 24 dimensions can be used as the acoustic features. As a result, it is possible to realize voice discrimination with higher accuracy. In addition, as a result of being able to calculate only the previous section and the rear section as the acoustic feature amount as necessary, by excluding parameters that do not show the feature so clearly, the processing load of the system can be reduced, It is possible to increase the processing speed.

  By the way, in the first embodiment, a case has been described in which speech discrimination is performed using acoustic feature amounts of an utterance section and sections before and after the utterance section. However, the present invention is not limited to this, and the input speech signal The speech discrimination may be performed by further using the language feature amount.

  Therefore, in the second embodiment, in the case where the speech discrimination is further performed using the acoustic feature amount of the utterance section and the preceding and following sections and the language feature amount of the input speech signal using FIGS. The experimental results will be described.

[Experiment environment]
First, in Example 2, it is assumed that two or more people and a system exist simultaneously. This can be said to be a natural situation, such as when there is a person around when operating the robot, or when a passenger is present in the passenger seat when operating the car navigation system. A robot that moves by voice commands was used as the system.

  Then, while two or more people are talking to each other, the system requests such as “Take a picture” or “Come here” to the robot. In addition, utterances such as “Follow me” that the user uttered in anticipation of the robot's motion, which cannot be accepted by the current robot, were also set as system request utterances. Recording was done with microphones attached to the chests of the two speakers. The number of utterances was 330, 49 of which were system requests.

[Combination of acoustic features and language features]
Language features are prepared as a vector space with the number of different words included in the speech recognition results of all utterances as dimensions, and the number of occurrences of words in one utterance is used as a vector element (reference: Sako, SIG) -SLP-64, 19-24, 2006-12). As acoustic features, especially in chatting, fillers, laughter, and grudges often enter the front and back sections of utterances, whereas in system-requested utterances, it is often silent before and after utterances. It is divided into three sections, a central part, a central part, and a rear part, and similarly, power and pitch are obtained and used for discrimination as a 24-dimensional acoustic feature amount. As for the combination of the acoustic feature quantity and the language feature quantity, as shown in FIG. 7, a new feature quantity is calculated by connecting the feature quantity vectors using the scaling coefficient. FIG. 7 is a diagram illustrating an example of combinations of acoustic feature amounts and language feature amounts. However, RMS (Root Mean Square) was used for power, and F0 was used for pitch.

[Experiment]
The RVM kernel was used for the SVM Kernel function that performs speech discrimination, and the evaluation was performed by leave-one-out. Moreover, since the same method as in the first embodiment is used for the speech discrimination using the acoustic feature amount, a detailed description thereof is omitted here. However, as a case where the F value is maximized, only the speech section is used. It can be seen that the accuracy is improved by using the three sections as compared to the case of performing (see FIG. 10).

  As a condition for speech recognition using language features, a baseline acoustic model was created using speech speech of 200 male speakers in the CSJ monitor. The acoustic analysis conditions and HMM specifications are shown in FIG. Furthermore, acoustic model adaptation was performed by MLLR + MAP, and the amount of adaptation data was about 10 minutes. The acoustic model adaptation is closed including the test set, and the language model adaptation is open by creating a language model for recognition of the speaker A using the speech of the speaker B. FIG. 8 is a diagram illustrating an example of acoustic analysis conditions and HMM specifications.

  As a result of performing speech recognition with julius under this condition, the correct word accuracy was 42.1%, and as a result of generating language features based on this speech recognition result, it became a 566-dimensional vector. The results of system request discrimination using this are shown in FIGS. As shown in FIG. 9 and FIG. 10, the system request determination can be performed with higher accuracy than the determination result using the acoustic feature amount. In addition, when the discrimination is performed with a total of 590 dimensions of the acoustic feature amount and the language feature amount, as shown in FIGS. 9 and 10, compared to the case where only the acoustic feature amount is used or the case where only the language feature amount is used, It can be seen that the system requirement can be determined with higher accuracy. However, the scaling of the linguistic feature and acoustic feature was selected by experimentally finding the one with the highest identification rate. 9 and 10 are diagrams showing the system request determination result.

[Effects of Example 2]
As described above, according to the second embodiment, the speech recognition is performed to extract the content of the utterance as the character data from the input speech signal, and the language feature (for example, the importance level of each character data or Weighting etc.), and the input speech signal is calculated from the calculated acoustic feature amount of the utterance section, the acoustic feature amount of the section before and after the calculated utterance section, and the generated language feature amount. Since it is determined whether or not it is a system request for requesting execution of processing to a processing apparatus that executes various connected processes, it is possible to realize voice discrimination with higher accuracy.

  For example, by using an expensive function or product that can calculate language features, speech discrimination with higher accuracy is possible compared to speech discrimination using only acoustic features in the utterance section and the preceding and following sections. Can be realized.

  Although the embodiments of the present invention have been described so far, the present invention may be implemented in various different forms other than the embodiments described above. Therefore, as shown below, (1) acoustic features, (2) speech discrimination, (3) utterance section identification method, (4) system configuration, etc., and (5) programs are divided into different examples, and different embodiments are described. To do.

(1) Acoustic feature amount For example, in the first and second embodiments, the case of using “average value, standard deviation value, maximum value, difference between maximum value and minimum value” of power and pitch as the acoustic feature amount is used. Although described, this invention is not limited to this, The combination can be arbitrarily performed, such as using "an average value and a standard deviation value."

  Further, in the first and second embodiments, the case where the acoustic features of the utterance section and the sections before and after the utterance section are used has been described. However, the present invention is not limited to this, and the utterance section, the subsequent section, the utterance section, and the like. Only the acoustic feature amount of one of the sections such as the previous section may be used.

(2) Voice discrimination In the first and second embodiments, the case where the SVM is used as the voice discrimination means has been described. However, the present invention is not limited to this. Adaboost, average distribution, average vector analysis, probability Various pattern identification methods such as statistics, GMM (Gaussian Mixture Model), and decision trees can be used.

(3) Spoken Interval Identification Method In the first and second embodiments, the case where the utterance interval is specified from the power of the input voice signal has been described. However, the present invention is not limited to this, and various methods are available. The speech section can be specified by Specifically, the utterance interval can be specified using a known method such as VAD (Voice Activity Detection).

  For example, speech features are extracted by MFCC (Mel Frequency Cepstrum Coefficient), etc., and the likelihood of speech is calculated by comparing the extracted speech features with an acoustic model describing features of speech and noise. A method of specifying an utterance section by determining whether it is speech or non-speech (noise) may be used.

(4) System Configuration In addition, all or a part of the processes (for example, feature weighting process or voice recognition process) described as automatically performed among the processes described in the present embodiment are manually performed. Alternatively, all or part of the processing described as being performed manually can be automatically performed by a known method. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters (for example, FIG. 3, FIG. 4, FIG. It can be arbitrarily changed except for.

  Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. It can be configured by integrating (for example, integrating the first acoustic feature calculation unit and the second acoustic feature calculation unit). Further, all or any part of each processing function performed in each device may be realized by a CPU and a program analyzed and executed by the CPU, or may be realized as hardware by wired logic.

(5) Program The voice discrimination method described in the present embodiment can be realized by executing a program prepared in advance on a computer such as a personal computer or a workstation. This program can be distributed via a network such as the Internet. The program can also be executed by being recorded on a computer-readable recording medium such as a hard disk, a flexible disk (FD), a CD-ROM, an MO, and a DVD and being read from the recording medium by the computer.

  As described above, the speech discrimination device according to the present invention is useful for discriminating whether an input speech signal is a request to the system or a chat, and particularly realizes highly accurate speech discrimination. Suitable for that.

1 is a system configuration diagram illustrating an overall configuration of a system including a voice discrimination device according to Embodiment 1. FIG. 1 is a block diagram illustrating a configuration of a voice discrimination device according to Embodiment 1. FIG. It is a figure which shows the example of the input audio | voice signal memorize | stored in input signal DB. It is a figure which shows the example of the information memorize | stored in discrimination | determination result DB. It is a figure for demonstrating the request discrimination | determination by SVM. 3 is a flowchart illustrating a flow of a voice discrimination process in the voice discrimination device according to the first embodiment. It is a figure which shows the example of the combination of an acoustic feature-value and a language feature-value. It is a figure which shows the example of the specification of acoustic analysis conditions and HMM. It is a figure which shows a system request | requirement determination result. It is a figure which shows a system request | requirement determination result.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Voice discrimination device 11 Communication control I / F part 12 Input part 13 Storage part 14 Input signal DB
15 Discrimination result DB
20 Control Unit 21 Speaking Section Identification Unit 22 First Acoustic Feature Calculation Unit 23 Second Acoustic Feature Calculation Unit 24 SVM
25 Instruction output section

Claims (4)

An utterance section specifying means for specifying an utterance section indicating an utterance section from the input voice signal;
First acoustic feature quantity calculating means for calculating an acoustic feature quantity indicating the characteristics of the voice signal from the power and pitch of the voice signal in the speech section specified by the utterance section specifying means;
A second acoustic feature quantity that identifies a section before and / or a section behind the speech section identified by the speech section identifying means, and calculates the acoustic feature quantity from the power and pitch of the audio signal in the identified section. A calculation means;
Various processes in which the input audio signal is connected from the acoustic feature quantity calculated by the first acoustic feature quantity calculation means and the acoustic feature quantity calculated by the second acoustic feature quantity calculation means. Voice discrimination means for discriminating whether or not it is a system request for requesting the processing device to execute processing;
A voice discrimination device comprising:   The first acoustic feature quantity calculating means includes an average value, a standard deviation value, a maximum value, and a maximum value and a minimum value of the power and pitch of the speech signal in the utterance section specified by the utterance section specifying means. The voice discrimination device according to claim 1, wherein any one or more of the differences are calculated as an acoustic feature amount.   The second acoustic feature quantity calculating means specifies a preceding section and / or a following section of the utterance section specified by the utterance section specifying means, and each of the power and pitch of the audio signal in the specified section is specified. The speech discrimination apparatus according to claim 1 or 2, wherein any one or more of an average value, a standard deviation value, a maximum value, and a difference between the maximum value and the minimum value is calculated as an acoustic feature amount. Speech recognition for extracting the content of the utterance as character data from the input speech signal, further comprising a language feature generating means for generating a language feature from the speech-recognized character data;
The voice discrimination means is generated by the acoustic feature quantity calculated by the first acoustic feature quantity calculation means, the acoustic feature quantity calculated by the second acoustic feature quantity calculation means, and the language feature quantity generation means. It is determined whether or not the input voice signal is a system request for requesting execution of processing to a processing device that executes various connected processes from the language feature amount The voice discrimination device according to any one of claims 1 to 3.

Images