JP6568813B2 - Information processing apparatus, voice recognition method, and program - Google Patents

Description

  The present invention relates to an information processing apparatus, a speech recognition method, and a program.

  In a call center such as a contact center or customer center, conversation between an operator and a customer is converted into characters (text), and the operator's response status is analyzed and monitored using the converted characters (text). . In addition, an operation is performed in which a manager (supervisor) monitors conversations between an operator and a customer in real time and takes appropriate measures in real time according to the operator's response status.

  For example, there is a technique disclosed in Patent Document 1 as a conventional technique related to a call center.

Japanese Patent Application Laid-Open No. 2015-211403

  The number of customers handled by the operator at the call center is not constant. For example, the number of inquiries from customers tends to be higher on weekdays than at night, and the number of inquiries from customers tends to be higher on holidays than on weekdays. Therefore, the load on the information processing apparatus that performs processing for converting the conversation between the operator and the customer into characters varies depending on the date and time.

  If the resources of the information processing apparatus are insufficient, there is a possibility that the process of converting the conversation between the operator and the customer in real time is not performed, and the call center manager cannot take appropriate measures. As a solution, there may be a method of securing sufficient resources of the information processing apparatus in accordance with the peak of the number of inquiries from customers, but there is a problem that the cost increases due to the addition of hardware.

  The present invention has been made in view of the above, and an object of the present invention is to provide a technique capable of suppressing an increase in processing load of an information processing apparatus that performs voice conversion processing.

  An information processing apparatus according to an embodiment of the present invention is an information processing apparatus having a plurality of speech recognition modules that perform processing for converting speech into characters, and includes an input unit that receives speech input, and a voice that is input by the input unit. Is received by the input means by the selection means for selecting a predetermined voice recognition module from the plurality of voice recognition modules and the predetermined voice recognition module based on the processing load in the information processing apparatus. Output means for outputting a character obtained by converting the converted voice.

  According to the embodiment of the present invention, there is provided a technology capable of suppressing an increase in processing load of an information processing apparatus that performs voice conversion processing.

It is a figure which shows the function structural example of the information processing apparatus in this Embodiment. It is a figure which shows the structure of DNN. It is a figure which shows an example of selection information. It is a flowchart which shows the process sequence which the information processing apparatus which concerns on embodiment performs. It is a figure which shows a mode that the voice recognition module selected changes according to the congestion degree of a call center. It is a figure which shows an example of selection information (modification).

  Embodiments of the present invention will be described below with reference to the drawings. The embodiment described below is only an example, and the embodiment to which the present invention is applied is not limited to the following embodiment. In the following embodiments, an information processing apparatus that performs a process of converting speech into characters in a call center will be described as an example. However, embodiments of the present invention are not limited to this, and perform a process of converting speech into characters. The present invention can be applied to all information processing apparatuses.

<Overview, functional configuration>
The information processing apparatus 10 according to the present embodiment has a plurality of speech recognition modules that perform processing for converting speech into characters. When the conversation between the operator and the customer is started, the information processing apparatus 10 selects an appropriate predetermined voice recognition module from among the plurality of voice recognition modules based on the processing load of the information processing apparatus 10 itself. Further, the information processing apparatus 10 performs processing for converting a conversation (voice) between the operator and the customer into characters using the selected predetermined voice recognition module.

  FIG. 1 is a diagram illustrating a functional configuration example of the information processing apparatus according to the present embodiment. As illustrated in FIG. 1, the information processing apparatus 10 includes an input unit 101, a load monitoring unit 102, a selection unit 103, a storage unit 104, a conversion processing unit 105, and an output unit 107. The conversion processing unit 105 includes a plurality of voice recognition modules 106. Each of these units is realized by processing that one or more programs installed in the information processing apparatus 10 cause the CPU of the information processing apparatus 10 to execute. Note that each of the input unit 101, the load monitoring unit 102, the selection unit 103, the storage unit 104, the conversion processing unit 105, and the output unit 107 may be realized by using different computers, or the computers are distributed in finer units. May be. That is, the information processing apparatus 10 may be realized using one or a plurality of computers. In addition, the one or more computers may be virtual servers using virtualization technology, or may be virtual servers mounted on the cloud.

In this embodiment, DNN (Deep Neural Network) is used as the speech recognition module 106 to perform conversion processing with high accuracy (recognition rate). The plurality of voice recognition modules 106 have different voice recognition rates and conversion speeds. In FIG. 1, four speech recognition modules 106 of DNN (1) 106 ₁ , DNN (2) 106 ₂ , DNN (3) 106 ₃ , DNN (4) 106 ₄ are shown as speech recognition modules 106. Yes. DNN (1) 106 ₁ has the highest speech recognition rate (conversely the conversion speed is slowest), and the voice recognition rate decreases in the order of DNN (1-4) 106 _1-4 (conversely, the conversion speed becomes high). ). Note that a DNN having a higher voice recognition rate (faster conversion speed) has a higher processing load necessary for the voice conversion process. In other words, the processing load decreases in the order of DNN (1-4) 106 _1-4 . It is assumed that each of the plurality of speech recognition modules 106 has at least a certain speech conversion accuracy. Although four DNNs (1 to 4) 106 ₁ to ₄ are illustrated in FIG. 1, the number of DNNs is not limited and may include five or more DNNs. Hereinafter, when DNN (1-4) 106 _1-4 is not distinguished, it is called "DNN106".

As shown in FIG. 2, the DNN (1-4) 106 _{1-4 according} to the present embodiment includes an input layer for inputting speech, an output layer for outputting characters, and an input layer and an output layer. It consists of multiple layers of existing hidden layers. Each layer is composed of a plurality of units.

  Basically, when the number of units in each layer is reduced, the amount of calculation is reduced. By reducing the amount of calculation, the processing load required for processing is reduced accordingly. On the other hand, it is assumed that the speech recognition rate will decrease as the number of units in each layer decreases, but it has been found as an experience value that there is almost no deterioration in the speech recognition rate with respect to the reduction ratio of the number of units in each layer (For example, even if the calculation amount is halved, the deterioration of the speech recognition rate is about several percent). In the present embodiment, by using such characteristics, when the processing load of the information processing apparatus 10 itself is high, the speech recognition module 106 with a small amount of calculation is selected, while maintaining a certain speech conversion accuracy. This makes it possible to suppress an increase in the load on the information processing apparatus 10 itself. Returning to FIG.

The input unit 101 has a function of accepting voice input related to the conversation when the conversation is started between the operator and the customer. Further, when the input unit 101 accepts voice input, the voice data should be passed to which voice recognition module 106 (any one of DNN (1 to 4) 106 _{1 to} 4) to the selection unit 103. And the voice data is transmitted to the voice recognition module 106 instructed by the selection unit 103. Note that the input unit 101 once transmits voice data to the voice recognition module 106 and then continues until the conversation between the operator and the customer ends (that is, until the voice session ends). Continue to transmit audio data to 106. That is, in the information processing apparatus 10 according to the present embodiment, once selected voice recognition module 106 is not changed until the conversation between the operator and the customer ends.

  The load monitoring unit 102 has a function of monitoring the processing load of the information processing apparatus 10 itself. Further, the load monitoring unit 102 notifies the processing load of the information processing apparatus 10 itself in response to an inquiry from the selection unit 103.

In response to the inquiry from the input unit 101, the selection unit 103 should pass the voice data to which voice recognition module 106 (in the example of FIG. 1, one of DNN (1 to 4) 106 _{1 to N} ). Instruct. More specifically, when receiving an inquiry from the input unit 101, the selection unit 103 inquires of the load monitoring unit 102 about the processing load of the information processing apparatus 10 itself. Subsequently, the selection unit 103 selects a voice recognition module 106 corresponding to the processing load notified from the load monitoring unit 102 according to selection information described later, and instructs the input unit 101 of the selected voice recognition module 106.

  The storage unit 104 stores selection information. The storage unit 104 can be realized by using a storage device (memory, HDD, or the like) included in the information processing apparatus 10, a storage device that can be connected to the information processing apparatus 10 via a network, or the like.

FIG. 3 is a diagram illustrating an example of selection information. The selection information is associated with the processing load range of the information processing apparatus 10 and the voice recognition module 106 to be selected for each range. FIG. 3 shows an example of selection information when the CPU usage rate is used as the processing load of the information processing apparatus 10. More specifically, in the example of FIG. 3, when the CPU usage rate is 70% or less (or less), DNN (1) 106 ₁ should be selected, and the CPU usage rate is 70% to 80%. should be selected DNN (2) 106 _2, CPU utilization is 80% to 90%, should be selected DNN (3) 106 _3, CPU utilization is 90% (or greater than) for, it has been shown that should select DNN ₍₄₎ 106 4.

The “conversion speed” is a conversion speed per DNN 106 unit, and means that the higher the number, the faster the conversion speed when converting speech into characters. “Voice recognition rate” means the voice recognition rate (conversion accuracy from voice to characters) in the DNN 106. “Layer, unit” indicates the structure of DNN 106 (how many layers are formed, and how many units are included in each layer). For example, DNN (1) 106 ₁ indicates that it is composed of 10 layers and 2048 units. The numbers shown in “layer, unit” are merely examples, and any number can be taken in the present embodiment. Note that “conversion speed”, “voice recognition rate”, and “layer, unit” are reference information regarding the DNN 106, and thus may not be included in the selection information. The selection information illustrated in FIG. 3 is an example. When the conversion processing unit 105 includes five or more DNNs 106, the range of the processing load of the information processing apparatus 10 set in the selection information according to the number of DNNs 106 is as follows. It may be further finely divided.

The conversion processing unit 105 performs processing for converting the voice received from the input unit 101 into characters. As described above, the conversion processing unit 105 includes a plurality of speech recognition modules 106 (DNN (1-4) 106 _1-4 ). For example, the voice recognition module 106 analyzes the voice data received from the input unit 101 to extract an acoustic feature, and converts the extracted acoustic feature into a sentence, thereby converting the voice into a character. May be.

  The output unit 107 has a function of outputting characters converted by the voice recognition module 106.

<Operation example>
FIG. 4 is a flowchart illustrating a processing procedure performed by the information processing apparatus according to the embodiment. A processing procedure performed by the information processing apparatus 10 according to the present embodiment will be described with reference to FIG.

  In step S <b> 101, the input unit 101 receives an input of voice related to a conversation when an inquiry from the customer is generated and a conversation is started between the operator and the customer.

  In step S <b> 102, the input unit 101 inquires of the voice recognition module 106 to which voice data should be transmitted to the selection unit 103. Subsequently, the selection unit 103 inquires of the load monitoring unit 102 about the processing load of the information processing apparatus 10 itself.

In step S <b> 103, the selection unit 103 uses the selection information to correspond to the processing load of the information processing apparatus 10 itself notified from the load monitoring unit 102, which corresponds to the voice recognition module 106 (DNN (1-4) 106 _1-4 . Any one). For example, when the information shown in FIG. 3 is stored in the selection information, if the processing load notified from the load monitoring unit 102 is 60%, the selection unit 103 uses the DNN (1) as the voice recognition module 106. 106 ₁ is selected. For example, when notified process load from the load monitoring unit 102 was 95%, the selection unit 103 will select the DNN (4) 106 ₄ as the speech recognition module 106.

  In step S104, the selection unit 103 notifies the input unit 101 of the voice recognition module 106 selected in the processing procedure of step S103. The input unit 101 transmits voice data to the notified voice recognition module 106. The voice recognition module 106 performs processing of analyzing voice data received from the input unit 101 and converting it into characters.

  In step S <b> 105, the output unit 107 outputs the characters converted by the voice recognition module 106.

  The processing procedure from step S101 to step S105 described above is performed every time an inquiry from a customer occurs and a conversation is started between the operator and the customer, that is, every time a new voice input is received by the input unit 101. Repeatedly.

  FIG. 5 is a diagram showing how the selected voice recognition module changes according to the congestion degree of the call center.

When the number of inquiries from customers is small, such as at night (the state on the left side of FIG. 5), the processing load on the information processing apparatus 10 is low, so the highest speech recognition rate is achieved in the process of converting each conversation into characters. The DNN (1) 106 ₁ that is the voice recognition module 106 is selected.

Next, when the number of inquiries from customers has increased (the state in the center of FIG. 5), the processing load of the information processing apparatus 10 gradually increases, so the conversation related to new inquiries from customers is converted into characters. in the process, DNN (1) 106 is smaller calculation amount than ₁ (DNN (1) 106 low speech recognition rate than _1, and processing speed is high) DNN (2) is a speech recognition module 106 106 ₂ or DNN (3) 106 ₃ is selected. Incidentally, the example of the center of FIG. 5, DNN (1) 106 ₁ is illustrated as a speech recognition module 106 which is selected. As described above, since the voice recognition module 106 once selected is not changed until the conversation between the operator and the customer is finished, it is selected when the voice recognition module 106 is in the state on the left side of FIG. This shows that the DNN (1) 106 ₁ thus made may continue to operate even after shifting to the center state of FIG.

Next, when the number of inquiries from customers has increased further, such as on holidays or during the daytime (the state on the right side of FIG. 5), the processing load on the information processing apparatus 10 further increases, so new inquiries from customers are made. in the process of converting the conversation of the character, the amount of calculation than DNN (3) 106 ₃ is less (DNN ₍₃₎ 106 3 lower speech recognition rate than, and higher processing speed) DNN a speech recognition module 106 (4) so that the 106 ₄ is selected.

Note that the selection unit 103 includes two combinations of the selected DNN 106 (for example, two DNN (1) 106 _1, DNN (2) 106 _2, and DNNN (3) 106 ₃ as shown in the center of FIG. And the processing load of the information processing apparatus 10 itself may be associated with each other and stored in the history as needed. In addition, when a change in the processing load of the information processing apparatus 10 itself is small when the selection unit 103 newly selects any of the DNNs 106 (for example, less than a predetermined threshold value), the selection unit 103 next selects the DNN 106. Alternatively, the DNN 106 having a higher voice recognition rate may be selected. On the contrary, when the selection unit 103 newly selects any of the DNNs 106 and the change in the processing load of the information processing apparatus 10 itself is large (for example, when it is equal to or greater than a predetermined threshold), the selection unit 103 next selects the DNNs 106. Alternatively, the DNN 106 having a low voice recognition rate may be selected. As a result, the selection unit 103 can learn various correspondence relationships between the combinations of the DNNs 106 and the processing loads, and can realize a DNN selection method with higher accuracy.

  As described above, the information processing apparatus 10 according to the present embodiment has a smaller amount of calculation as the number of conversations simultaneously performed between the operator and the customer (the number of customers who are simultaneously supported) increases (speech) The speech conversion module 106 operates so as to perform speech conversion processing (with a low recognition rate and a high processing speed). Thereby, the information processing apparatus 10 according to the present embodiment can suppress an increase in the processing load of the information processing apparatus 10 itself while maintaining a certain voice conversion accuracy. In addition, the information processing apparatus 10 according to the present embodiment can suppress an increase in its processing load and can keep the average processing speed of the voice conversion processing performed by itself at a certain speed or higher.

<Modification>
In the embodiment described above, the CPU usage rate is used as the processing load of the information processing apparatus 10, but other parameters may be used instead of the CPU usage rate. For example, the information processing apparatus 10 may select the voice recognition module 106 based on the number of processes of the DNN 106 instead of the CPU usage rate. Since the information processing apparatus 10 according to the present embodiment assumes that one process of the DNN 106 is activated for each conversation between the operator and the customer, the number of processes = the number of customers who are simultaneously supporting It can also be said.

  In this modification, as shown in FIG. 6A, the selection information stores the number of processes instead of the CPU usage rate. Further, the load monitoring unit 102 according to the present modification operates so as to monitor the number of processes of the DNN 106 instead of the CPU usage rate. Further, the selection unit 103 according to the present modification operates to select the voice recognition module 106 corresponding to the number of processes notified from the load monitoring unit 102 using the selection information.

  Further, the CPU usage rate and other parameters may be used together as the processing load of the information processing apparatus 10. For example, the information processing apparatus 10 may select the voice recognition module 106 based on both the CPU usage rate and the number of processes of the DNN 106.

  In this case, as shown in FIG. 6B, both the CPU usage rate and the number of processes are stored in the selection information. In addition, the load monitoring unit 102 according to the present modification operates to monitor both the CPU usage rate and the number of processes. In addition, the selection unit 103 according to this modification uses the selection information to correspond to the voice recognition module 106 corresponding to the CPU usage rate notified from the load monitoring unit 102 and the number of processes notified from the load monitoring unit 102. The speech recognition module 106 to be selected, and, of the two selected speech recognition modules 106, the speech recognition module 106 having a small calculation amount (low speech recognition rate and high processing speed), or the calculation amount It operates to select the speech recognition module 106 with a large number (high speech recognition rate and low processing speed).

<Supplement of embodiment>
The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims. The processing procedures described in the embodiments may be switched in order as long as there is no contradiction.

  As described above, each functional unit of the information processing apparatus 10 according to the embodiment executes a program corresponding to the process executed by the information processing apparatus 10 by using hardware resources such as a CPU and a memory included in the information processing apparatus 10. It is possible to realize. Further, the program can be stored in a storage medium.

DESCRIPTION OF SYMBOLS 10 Information processing apparatus 101 Input part 102 Load monitoring part 103 Selection part 104 Storage part 105 Conversion processing part 106 Voice recognition module 107 Output part

Claims (6)

An information processing apparatus having a plurality of speech recognition modules that perform processing for converting speech into characters,
Input means for receiving voice input;
A selection unit that selects a voice recognition module from the plurality of voice recognition modules based on a processing load in the information processing apparatus when receiving an input of voice by the input unit;
An output means for outputting a character obtained by converting the voice received by the input means by the voice recognition module selected by the selection means ;
The selection means includes
When the voice is converted by the selected voice recognition module, if the change in processing load in the information processing apparatus is less than a predetermined threshold, a voice recognition module with a higher voice recognition rate is selected as the next voice recognition module And
When the voice is converted by the selected voice recognition module, if the change in the processing load in the information processing apparatus is equal to or greater than a predetermined threshold, a voice recognition module with a lower voice recognition rate is selected as the next voice recognition module An information processing apparatus. The plurality of voice recognition modules perform conversion processing at different voice recognition rates and different conversion speeds, respectively.
The selection means selects a speech recognition module that performs conversion processing at a speech recognition rate and a conversion speed according to the range of the processing load height.
The information processing apparatus according to claim 1. The selection unit selects a speech recognition module that performs conversion processing at a speech recognition rate and a conversion speed according to a range of the processing load height and the number of processes in which the speech recognition module performs conversion processing. Item 3. The information processing device according to Item 2. The information processing apparatus according to any one of claims 1 to 3, wherein the voice recognition module performs a process of converting voice into characters by a DNN (Deep Neural Network). A speech recognition method executed by an information processing apparatus having a plurality of speech recognition modules that perform processing for converting speech into characters,
An input step for receiving voice input;
A selection step of selecting a speech recognition module from the plurality of speech recognition modules based on a processing load in the information processing apparatus when receiving an input of speech in the input step;
An output step of outputting a character obtained by converting the voice received in the input step by the voice recognition module selected in the selection step ;
The selection step includes
When the voice is converted by the selected voice recognition module, if the change in processing load in the information processing apparatus is less than a predetermined threshold, a voice recognition module with a higher voice recognition rate is selected as the next voice recognition module And
When the voice is converted by the selected voice recognition module, if the change in the processing load in the information processing apparatus is equal to or greater than a predetermined threshold, a voice recognition module with a lower voice recognition rate is selected as the next voice recognition module to, voice recognition method. The program for functioning a computer as each means of the information processing apparatus as described in any one of Claims 1 thru | or 4 .

Images