JP2013029868A - Conversation screening program, conversation screening device, and conversation screening method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve work efficiency of monitoring by screening problematic conversation efficiently and highly accurately.SOLUTION: A recording voice data is acquired, and prosody information extraction processing is performed by an extraction unit 401. Next, speech utterance period detection processing is performed by a detection unit 402. Voice recognition result information is acquired by an acquisition unit 405. Then, basic conversation analysis processing by a basic conversation analysis unit 403, conversation structure analysis processing by a conversation structure analysis unit 404, and speech utterance content analysis processing by a speech utterance content analysis unit 406 are performed. Furthermore, determination processing is performed by a determination unit 407, and output processing is performed by an output unit 408.

Description

  The present invention relates to a dialogue selection program, a dialogue selection device, and a dialogue selection method for selecting a dialogue between recorded speakers as a problem dialogue or a normal dialogue.

  In the call center, monitoring is performed to check the response of an agent (also called an operator) to a customer. Monitoring means actually listening to the dialogue between the agent and the customer and checking the content of the dialogue. In monitoring, several dialogues of each agent are selected at random, and the selected dialogues are listened to from the head and checked.

  There are two main purposes for monitoring. One is to evaluate the agent's response skills (speaking, wording, etc.) and use it for guidance. The second is to investigate and respond to the cause of the problem call. When monitoring for the second purpose, it is possible to listen efficiently by selecting calls that may be problematic calls in advance. Here, the problem call will be specifically described. Problem calls are roughly divided into the following (1) to (3).

(1) Calls that take more time to answer than necessary (2) Calls that make customers angry (3) Calls that do not engage with customer interactions

In the problem call (1), a long time is spent waiting for a customer for an answer search or the like.
In the problem call (2), the customer is angry before the dialogue, or the customer is angry because of inappropriate response or misunderstanding.
The problem call of (3) is because the dialogue is prolonged or the story is not engaged because the agent engages the dialogue due to “the agent cannot answer and explain well”, “the agent does not understand the customer's question”, etc. In most cases, the customer gives up and ends the conversation.

  The problem call of the above (1) can be examined by using the technique of the following Patent Document 1 to check whether or not the agent is waiting for the customer by answer search or the like, and the problem call can be extracted. . In addition, the problem call in (2) above can be extracted by using the following Patent Document 2 to check the emotional state of the customer utterance and find the call that may be angry by the customer. It is.

  In addition, it is a fact that many call centers are selecting and checking calls with a long answering time. As a result of this check, it is possible to extract cases in which the conversation continues for a long time until the talk is engaged, among the problem calls of (3) above.

JP 2007-33754 A Special table 2003-508805 gazette

  However, among the problem calls in (3) above, the case where the conversation is terminated without talking is a problem that the call cannot be selected because the response is not prolonged.

  In order to solve the above-described problems caused by the conventional technology, the present invention can efficiently improve the monitoring work efficiency by efficiently and accurately selecting dialogues that become problem calls (problem dialogues). It is an object of the present invention to provide a sorting program, a dialogue sorting device, and a dialogue sorting method.

  In order to solve the above-described problems and achieve the object, a first dialogue selection program, a dialogue selection device, and a dialogue selection method are provided by performing a series of utterances from phonetic information for each speaker from speech data related to dialogue between speakers. Detecting a section for each speaker, identifying an initiative speaker for each specific conversation section of the two speakers based on a section length of a series of utterance sections for each detected speaker, and recording the recording The section in which the initiative speaker is a speaker who asks the other party at the beginning of the voice data is set as a question phase, and the initiative speaker is a speaker who receives a question from the other party after the question phase. A certain section is set as an answer phase, and after the answer phase, the designated speaker among the two speakers calculates the initiative holding time that is the initiative speaker, and is based on the time length of the initiative speaker holding time of the designated speaker. The Wherein analyzing the deviation of the specified speaker conversation, based on the analysis result, it determines the dialogue interaction problems and requirements to output the determination result.

  According to the dialog screening program, the dialog screening device, and the dialog screening method, it is possible to select a dialog that has been unilaterally spoken by any one of the speakers as a problem dialog that has ended without engaging the dialog.

  Further, the second dialogue selection program, the dialogue selection device, and the dialogue selection method detect a series of utterance sections for each speaker from phonological information for each speaker from recorded voice data related to a dialogue between speakers, Based on the detected section length of a series of utterance intervals for each speaker, the initiative speaker is identified for each specific conversation interval of the two speakers, and the initiative story is started at the beginning of the recorded voice data. The section where the speaker is the speaker who is in the position to ask the other party the question is the question phase, and after that question phase the section where the initiative speaker is the speaker who receives the question from the other party is the answer phase, and the predetermined A recognition result for each speaker regarding the recorded voice data including a keyword that matches or relates to a keyword and its appearance time is acquired, and a key selected from the recognition results of the designated speaker is acquired. Based on the appearance status of the specified speaker in the series of utterance intervals detected by the detection means, the smoothness of the progress after the answer phase of the dialogue is analyzed, and based on the analysis result It is a requirement to determine the dialogue as a problem dialogue and output the decision result.

  According to the dialogue selection program, the dialogue selection device, and the dialogue selection method, a dialogue in which the same utterance content is repeated can be selected as a problem dialogue that has ended without engaging the dialogue.

  According to the dialogue selection program, the dialogue selection device, and the dialogue selection method, it is possible to improve the monitoring work efficiency by efficiently and highly accurately selecting problem dialogues.

It is explanatory drawing which shows an example of the problem call used as the extraction object in this Embodiment. It is a block diagram which shows the hardware constitutions of the dialogue selection apparatus concerning this Embodiment. It is explanatory drawing which shows the memory content of a call information table. It is a block diagram which shows the functional structure of the dialogue selection apparatus concerning this Embodiment. It is explanatory drawing which shows the input / output of an extraction part. It is explanatory drawing which shows the input / output of a detection part. It is explanatory drawing which shows the dynamic specific example of an initiative speaker. It is explanatory drawing which shows the input / output of a basic dialog analysis part. It is explanatory drawing which shows the input / output of a dialog structure analysis part. It is explanatory drawing which shows the acquisition process of the speech recognition result information by an acquisition part. It is explanatory drawing which shows the input / output of an utterance content analysis part. It is explanatory drawing which shows the memory content of a similarity calculation table. It is explanatory drawing which shows a problem list file. It is a flowchart which shows the dialogue selection process procedure concerning this Embodiment. It is a flowchart which shows the detailed process sequence of the utterance area detection process (step S1403) by a detection part. It is a flowchart which shows the detailed process sequence of the basic dialog analysis process (step S1406) by a basic dialog analysis part. It is a flowchart (the 1) which shows the detailed process sequence of the initiative speaker specific process (step S1601) by an initiative speaker specific part. It is a flowchart (the 2) which shows the detailed process sequence of the initiative speaker specific process (step S1601) by a initiative speaker specific part. It is a flowchart which shows the detailed process sequence of the dialog structure analysis process (step S1407) by a dialog structure analysis part. It is a flowchart which shows the detailed process sequence of the utterance content analysis process (step S1408) by the utterance content analysis part. It is a flowchart which shows the detailed process sequence of the determination process (step S1409) by a determination part.

Exemplary embodiments of the dialog selection program, the dialog selection apparatus, and the dialog selection method will be described below in detail with reference to the accompanying drawings. In order to select dialogues, it is first necessary to define what kind of dialogues are normal dialogues and abnormal dialogues that become problem dialogues, and to enable automatic selection by a computer. The dialogue at the call center or window is a dialogue of “question from customer → answer by agent → plus α utterance”. Regardless of the normal dialogue or problem dialogue, “question from customer ⇒ answer by agent” is almost the same, and “normal utterance” determines whether it is normal or abnormal.

  For example, additional questions and provision of reference information are the contents that can be issued because the dialogue is engaged, and in this case, the dialogue is normal. In addition, when there is no bias in the utterance time between the customer and the agent, it is presumed that communication is taking place, so the dialogue is normal. Furthermore, if the same utterance or word is not repeated, the dialogue is proceeding smoothly, so it can be said that the dialogue is normal.

  On the other hand, if one speaker is speaking continuously or the same utterance or word is repeated, it is presumed that the dialogue is not engaged. Therefore, it can be said that the dialogue is biased toward one speaker or the same utterance or word is repeated.

  In the present embodiment, the target of monitoring is narrowed down by automatically selecting problem dialogs that have ended without engaging such dialogs. This increases the efficiency of the monitoring work and reduces the work load.

(Example of problem call)
FIG. 1 is an explanatory diagram showing an example of a problem call to be extracted in the present embodiment. In FIG. 1, reference symbol Ta # (# is a number) is an agent utterance, and reference symbol Tc # is a customer utterance.

  In the customer utterance Tc1, the customer feels that the driver has not been installed, and makes an inquiry to the agent. On the other hand, in the utterance Ta1, since the agent cannot read the floppy for that reason, the agent tells the reading method.

  Further, in the utterance Tc3, the customer wants to obtain advice / answer for the fact that the driver could not be installed following the manual procedure, so the same is repeated.

  Also, in utterance Ta5, the agent cannot respond to the question and escapes because the floppies can be read by the method taught anyway. On the other hand, in the utterance Tc6, the customer ends the dialogue by saying “I understand” whether he / she realized that it is meaningless to continue further.

  The dialogue does not give an accurate answer to the customer's question, “I was unable to install the driver, but is there a problem?” The fact that the call is a problem call can be understood by actually listening to the call.

  In the above example, the recorded voice data of the telephone conversation (call) between the customer and the agent such as a call center has been described. However, the recorded voice data of the direct conversation with the customer at the store window is not limited to the call. Good.

(Hardware configuration of dialog screening device)
FIG. 2 is a block diagram showing a hardware configuration of the dialog screening apparatus according to the present embodiment. In FIG. 2, the dialogue selection device includes a CPU (Central Processing Unit) 201, a ROM (Read-Only Memory) 202, a RAM (Random Access Memory) 203, a magnetic disk drive 204, a magnetic disk 205, and an optical disk drive. 206, an optical disk 207, a display 208, an I / F (Interface) 209, a keyboard 210, a mouse 211, a scanner 212, and a printer 213. Each component is a bus 2
00, respectively.

  Here, the CPU 201 governs the overall control of the dialogue selection device. The ROM 202 stores a program such as a boot program. The RAM 203 is used as a work area for the CPU 201. The magnetic disk drive 204 controls reading / writing of data with respect to the magnetic disk 205 according to the control of the CPU 201. The magnetic disk 205 stores data written under the control of the magnetic disk drive 204.

  The optical disk drive 206 controls reading / writing of data with respect to the optical disk 207 according to the control of the CPU 201. The optical disk 207 stores data written under the control of the optical disk drive 206, or causes the computer to read data stored on the optical disk 207.

  The display 208 displays data such as a document, an image, and function information as well as a cursor, an icon, or a tool box. As the display 208, for example, a CRT, a TFT liquid crystal display, a plasma display, or the like can be adopted.

  An interface (hereinafter abbreviated as “I / F”) 209 is connected to a network 214 such as a LAN (Local Area Network), a WAN (Wide Area Network), and the Internet through a communication line. Connected to other devices. The I / F 209 controls an internal interface with the network 214 and controls data input / output from an external device. For example, a modem or a LAN adapter may be employed as the I / F 209.

  The keyboard 210 includes keys for inputting characters, numbers, various instructions, and the like, and inputs data. Moreover, a touch panel type input pad or a numeric keypad may be used. The mouse 211 performs cursor movement, range selection, window movement, size change, and the like. A trackball or a joystick may be used as long as they have the same function as a pointing device.

  The scanner 212 optically reads an image and takes in the image data into the dialogue sorting device. The scanner 212 may have an OCR (Optical Character Reader) function. The printer 213 prints image data and document data. As the printer 213, for example, a laser printer or an ink jet printer can be employed.

(Call information table)
FIG. 3 is an explanatory diagram showing the stored contents of the call information table. The call information table 300 realizes its function by a storage device such as the ROM 202, the RAM 203, and the magnetic disk 205 shown in FIG. The call information table 300 has field items such as a call ID, a dialog time, a dialog structure analysis result flag Fa, an utterance content analysis result flag Fb, and an AND result flag Fc, and is call information indicating a dialog for each record. The call ID is identification information that identifies a dialog (or call information thereof). The call ID becomes a pointer to the storage destination of the recorded voice data of the dialog.

  Here, the recorded voice data is voice data obtained by recording dialogue voice, and is stored in a database (not shown). When the recorded voice data is stereo, one channel is customer voice data and the other channel is agent voice data. In the case of monaural, it is assumed that the speakers are separated.

  The dialogue time is the time from the start to the end of recording of the recorded voice data. The dialogue time is information extracted from the data length of the recorded voice data. The dialog structure analysis result flag Fa is binary information indicating the dialog structure analysis result of the dialog, and the default value is Fa = 0. Dialog structure analysis will be described later. When problem dialogue selection is performed using only the dialogue structure analysis result flag Fa, when the dialogue structure analysis result flag Fa is Fa = 0, the dialogue specified by the recorded voice data of the call information is a normal dialogue. On the other hand, when the dialog structure analysis result flag Fa is Fa = 1, the dialog specified by the recorded voice data of the call information is a problem dialog.

  The utterance content analysis result flag Fb is binary information indicating the utterance content analysis result of the dialogue, and the default value is Fb = 0. The utterance content analysis will be described later. When problem dialogue selection is performed using only the utterance content analysis result flag Fb, when the utterance content analysis result flag Fb is Fb = 0, the dialogue specified by the recorded voice data of the call information is a normal dialogue. On the other hand, when the utterance content analysis result flag Fb is Fb = 1, the dialogue specified by the recorded voice data of the call information is a problem dialogue.

  The AND result flag Fc is binary information indicating an AND result of the dialog structure analysis result flag Fa and the utterance content analysis result flag Fb. The default value is Fc = 0. When the dialogue structure analysis result flag Fa and the utterance content analysis result flag Fb are Fa = 1 and Fb = 1, Fc = 1. When problem dialogue selection is performed using the dialogue structure analysis result flag Fa and the utterance content analysis result flag Fb, when the AND result flag Fc is Fc = 0, the dialogue specified by the recorded voice data of the call information is a normal dialogue. It becomes. On the other hand, when the AND result flag Fc is Fc = 1, the dialogue specified by the recorded voice data of the call information is a problem dialogue.

(Functional configuration of dialog screening device)
FIG. 4 is a block diagram showing a functional configuration of the dialogue selection apparatus according to the present embodiment. The dialogue selection device 400 includes an extraction unit 401, a detection unit 402, a basic dialogue analysis unit 403, a dialogue structure analysis unit 404, an acquisition unit 405, an utterance content analysis unit 406, a determination unit 407, and an output unit 408. The basic dialog analysis unit 403 includes an initiative speaker specifying unit 431 and an opening phase specifying unit 432. The utterance content analysis unit 406 includes a calculation unit 461 and a determination unit 462.

  Specifically, each of the functions 401 to 408 is executed by causing the CPU 201 to execute a program stored in a storage device such as the ROM 202, the RAM 203, and the magnetic disk 205 illustrated in FIG. 2 or by the I / F 209. Realize its function. Hereinafter, each function will be described individually.

<Extractor 401>
FIG. 5 is an explanatory diagram showing input / output of the extraction unit 401. The extraction unit 401 has a function of extracting phoneme information for each speaker from the recorded voice data D relating to the conversation between the speakers. Specifically, for example, as illustrated in FIG. 5, the extraction unit 401 inputs the recorded voice data D, and extracts the agent prosody data Ma and the customer prosody data Mc from the input recorded voice data D. The prosodic data Ma and Mc are information in which the power value for each channel (for each speaker) of the recorded voice data D for each predetermined time is recorded. A prosodic ID is assigned to each power value. The prosodic ID is an ascending number in time series. The extraction of the prosodic data Ma and Mc is realized by a known method, and therefore details thereof are omitted in this specification.

<Detection unit 402>
FIG. 6 is an explanatory diagram showing input / output of the detection unit 402. The detection unit 402 has a function of detecting a series of utterance sections for each speaker from the phoneme information for each speaker extracted by the extraction unit 401. Specifically, for example, as shown in FIG. 5, agent utterance section information Sa in which a series of utterance sections of the agent is time-sequentially generated from the agent prosodic data Ma. Similarly, customer utterance section information Sc in which a series of utterance sections of the customer are time-sequentially generated from the customer prosody data Mc.

  Specifically, the detection unit 402 reads the power values of the prosodic data Ma and Mc in time series, the continuous power value is equal to or greater than a predetermined threshold value, and the continuous time is equal to or greater than the predetermined minimum utterance interval length. Is detected as an utterance interval. In the utterance section information Sa and Sc shown in FIG. 6, each utterance section has an utterance ID, a start time, and an end time. The utterance ID is identification information for specifying the utterance section. Here, the utterance ID is information obtained by adding a branch number to the call ID in order to associate with the call information.

<Basic dialogue analysis unit 403>
As shown in FIG. 4, the basic dialog analysis unit 403 includes an initiative speaker specifying unit 431 and an opening phase specifying unit 432. The initiative speaker identification unit 431 identifies the initiative speaker for each specific conversation section of both speakers based on the section length of a series of utterance sections for each speaker detected by the detection unit 402. Specifically, for example, it is specified which speaker is the speaker who has the initiative in the conversation in a certain conversation section based on the length and frequency of the utterance section information Sa and Sc.

  Here, the dialogue section is a section in which both speakers speak alternately. The conversation section may be set dynamically during the initiative speaker specifying process, or may be statically set so that a predetermined section is set as the conversation section. When set dynamically, the interval from the first utterance start time to the predetermined utterance end time when the number of utterances of both speakers or one of the speakers reaches the predetermined number of utterances is the conversation interval It becomes. Then, the speaker with the longer utterance time in the dialogue section is set as the initiative speaker.

  FIG. 7 is an explanatory diagram illustrating a dynamic specific example of initiative speakers. Here, a description will be given using the dialogue shown in FIG. In FIG. 1, the dialogue ends with the utterance Tc6, but here it is assumed that the dialogue continues thereafter. Further, the predetermined number of utterances serving as a threshold is three, and the initiative speaker is specified when the number of utterances of any one of the speakers becomes three.

  When counting from the utterance Tc1, first, a section from the start time of the customer's utterance Tc1 to the end time of the customer's utterance Tc3 is the conversation section R1. The customer utterance time in the conversation section R1 is the total section length of the utterances Tc1 to Tc3, and the agent utterance time in the conversation section R1 is the total section length of the utterances Ta1 to Ta2. In this case, since the total section length of the utterances Tc1 to Tc3 is larger, the initiative speaker in the conversation section R1 is a customer. Similarly, in the conversation period R2, the initiative speaker is an agent.

  FIG. 8 is an explanatory diagram showing input / output of the basic dialog analysis unit 403. As shown in FIG. 8, the basic dialogue analysis unit 403 inputs the utterance interval information Sa and Sc, and identifies the initiative speaker for each dialogue interval as shown in FIG. Q is output. The initiative speaker information Q includes a call ID, initiative speaker, and conversation section (start time and end time) as field items. Each record indicates who is the initiative speaker in the dialogue section in the dialogue specified by the call ID.

In FIG. 4, the opening phase specifying unit 432 sets a section in which the initiative speaker is a speaker (a customer in this example) asking the other party at the beginning of the recorded audio data D as a question phase. After the question phase, the answering phase is a section in which the initiative speaker is a speaker (agent in this example) in a position to receive a question from the other party. In particular,
As shown in FIG. 8, the dialogue section R1 in which the customer first becomes the initiative speaker among the total dialogue sections is set as the question phase. Further, after the question phase, the dialogue section R2 in which the agent first becomes the initiative speaker is set as the answer phase.

  As shown in FIG. 8, the opening phase specifying unit 432 generates and outputs opening phase information Ph. The opening phase information Ph includes a call ID, a phase type, and a phase section (start time, end time) as field items. Each record indicates a section between the question phase and the answer phase at the beginning of the dialogue specified by the call ID.

  As described above, the question phase and the answer phase are specified by the opening phase specifying unit 432, whether it is a normal dialogue or a problem dialogue at the beginning of the dialogue, but there is at least one question and answer. This is because whether or not the dialogue is normal depends on the contents of the subsequent “plus α dialogue”.

<Dialogue structure analysis unit 404>
FIG. 9 is an explanatory diagram showing input / output of the dialog structure analysis unit 404. The dialog structure analysis unit 404 inputs the agent utterance section information Sa, the customer utterance section information Sc, the initiative speaker information Q, and the opening phase information Ph, and outputs a dialog structure analysis result. Specifically, the dialog structure analysis result flag Fa of the call information table 300 is updated.

  In FIG. 4, the dialog structure analysis unit 404 has a function of structurally analyzing dialogs after the answer phase specified by the recorded voice data D. Specifically, the initiative holding time in which the designated speaker is the initiative speaker among the two speakers after the answer phase identified by the opening phase identification unit 432 is calculated. Then, based on the time length of the designated speaker's initiative holding time, the bias of the dialogue toward the designated speaker is analyzed.

  Here, the designated speaker is a speaker set in advance. If the designated speaker is set to “customer”, the bias of the customer's utterance after the answer phase is analyzed, and if the designated speaker is set to “agent”, the bias of the agent's utterance after the answer phase is analyzed. Will be analyzed.

  The dialogue structure analysis unit 404 has a dialogue time length equal to or greater than a predetermined dialogue time length that is a threshold value after the answer phase, and the average speaker time length of the designated speaker is accumulated in a database (not shown). Is equal to or longer than the average utterance time length of all customer utterances, and is equal to or longer than a predetermined retention time length that is the threshold of the retention time of the designated speaker's initiative (the ratio of the two speakers to the initiative retention time is a predetermined ratio) If this is the case, it is determined that the conversation is biased toward the designated speaker.

  When the dialogue time after the answer phase is longer than the predetermined dialogue time, which is the threshold, if the dialogue time after the answer phase is short, cases such as just the greeting at the end of the call are excluded. Because. The reason why the average utterance time length is greater than or equal to the average utterance time length is to exclude cases where utterances such as “Yes” and “Yes” are continuous.

<Acquisition unit 405>
In FIG. 4, the acquisition unit 405 has a function of acquiring a recognition result for each speaker regarding the recorded voice data D including a keyword that matches or relates to a predetermined recognition keyword and its appearance time. Prior to this acquisition, voice recognition of the recorded voice data D is necessary.

FIG. 10 is an explanatory diagram illustrating a process of acquiring speech recognition result information by the acquisition unit 405. The voice recognition process reads the recorded voice data D and the voice recognition keyword list L, executes the voice recognition process of the recorded voice data D, and outputs voice recognition result information W. The acquisition unit 405 acquires the voice recognition result information W.

  In the speech recognition process, words uttered in the recorded speech data D are converted into words of text data by an existing method such as word spotting or a subword method. The speech recognition keyword list L is text data created from a target product, a business manual, or FAQ (Frequently Asked Questions) used by an agent.

  The speech recognition result information W has the recognition result ID, the recognition keyword, the speaker type, and the appearance time as field items. In the recognition result ID, a unique number is written each time it matches or relates to the recognition keyword in the speech recognition keyword list L. In the recognition keyword, a recognition keyword in the speech recognition keyword list L that matches or relates to the speech-recognized word is written. In the speaker type, the name of the speaker who has issued a word that matches or relates to the recognition keyword is written. In the appearance time, a time (a position on the recorded voice data D) at which a word that matches or is related to the recognition keyword is written.

  This acquisition unit 405 may execute voice recognition processing inside, or may receive from outside the dialogue sorting apparatus 400. In any case, at least the speech recognition result information W only needs to be held in the storage device in the dialogue selection device 400.

<Speech content analysis unit 406>
In FIG. 4, the utterance content analysis unit 406 has a function of analyzing the content of the dialogue after the answer phase. FIG. 11 is an explanatory diagram showing input / output of the utterance content analysis unit 406. The utterance content analysis unit 406 inputs the specified utterance section information (here, the specified speaker is the customer utterance section information Sc) and the speech recognition result information W, and outputs the utterance content analysis result. Specifically, the utterance content analysis result flag Fb in the call information table 300 is updated.

  The utterance content analysis unit 406 analyzes the smoothness of the progress of the dialogue based on the appearance status of the recognition keyword of the designated speaker selected from the recognition result list table in a series of utterance sections of the designated speaker. . The smoothness of the progress of the dialogue is determined by whether or not the same utterance content is repeated. Dialogues that are determined to be repeated become problem dialogues.

  As for the usage status of the recognition keyword, the change of the phrase in each utterance section (appearance status of the same recognition keyword) is selected to select the one having no change. In this case, not a short recognition keyword such as “Yes”, but a recognition keyword that becomes an utterance longer than a certain utterance length is targeted. The recognition keywords are selected for each utterance section and written into the similarity calculation table.

  FIG. 12 is an explanatory diagram showing the contents stored in the similarity calculation table. The similarity calculation table 1200 uses the utterance ID, the recognition keyword, and the similarity as field items. The utterance content analysis unit 406 extracts the utterance section of each utterance from the utterance section information of the designated speaker. Then, the recognition keyword extracted from the speech recognition result information W during the extracted utterance time is read out using the appearance time as a clue and is written in the record of the extracted utterance section of the similarity calculation table 1200.

  For example, when attention is paid to the utterance ID: call001-10, the recognition keywords “floppy”, “driver”, and “install” of the customers with the recognition result IDs: 1 to 3 are voiced as the recognition keywords having the appearance time in the utterance section. Extracted from the recognition result information W. Then, the recognition keywords “floppy”, “driver”, and “install” are written in the record of the utterance ID: call001-10 of the similarity calculation table 1200.

  Also, since the recognition keywords “install” and “floppy” of the recognition result IDs 4 and 5 appear in the utterance section of the utterance ID: call001-11, they are written in the record of the utterance ID: call001-11 of the similarity calculation table 1200. It is. Note that the recognition keyword “operation” with the recognition result ID: 6 appears in the utterance section with the utterance ID: call001-11, but is not a designated speaker (customer), and thus is not written in the similarity calculation table 1200.

  As shown in FIG. 4, the utterance content analysis unit 406 includes a calculation unit 461 and a determination unit 462. The calculation unit 461 has a function of calculating the similarity shown in FIG. Specifically, the number of occurrences (word count) of the same or similar keyword as the recognition keyword selected from the recognition results of the designated speaker in the continuous utterance section of the designated speaker and the designation in the continuous utterance section of the designated speaker Based on the number of appearances (number of phrases) of all recognition keywords in the recognition result of the speaker, the similarity between consecutive utterance sections is calculated.

Here, the similarity is calculated by A / B.
A: Number of keywords (number of phrases) that are the same as or similar to the recognition keyword selected from the speech recognition result information W in the continuous speech section of the designated speaker
B: Number of all recognized keywords (number of phrases) in the speech recognition result information W of the designated speaker in the continuous utterance section of the designated speaker

  In the numerator A of the similarity, the utterance section for which the similarity is calculated is a continuous utterance section. The continuous utterance sections are two utterance sections of a designated speaker (in this case, a customer) that is continuous in time series. These two utterance sections may be literally continuous, but an utterance section of a non-designated speaker (here, an agent) may exist between them. In addition, similar keywords may be issued in the continuous utterance period, and the identity of the keywords may depend on the accuracy of the speech recognition processing, so similar keywords may be counted. Whether or not they are similar may be determined by referring to a synonym database (not shown).

  Hereinafter, the similarity calculation method will be described with reference to FIG. Since the utterance section of the utterance ID: call001-10 is the head utterance section, the similarity cannot be calculated. For the utterance section of the utterance ID: call001-11, the similarity is calculated with the preceding utterance section (utterance ID: call001-10).

  First, a keyword that is the same as or similar to the recognition keywords “install” and “floppy” in the utterance section (utterance ID: call001-11) is searched from the recognition keyword group in the utterance section (utterance ID: call001-10). Since the recognition keywords “install” and “floppy” also exist in the recognition keyword group in the utterance section (utterance ID: call001-10), the numerator A = 2. On the other hand, the denominator B has a denominator B = 3 because the recognition keywords appearing in the continuous speech section are “install”, “floppy”, and “driver”. Therefore, the similarity between the utterance section (utterance ID: call001-10) and the utterance section (utterance ID: call001-11) is A / B = 0.66.

  In FIG. 4, the determination unit 462 has a function of determining whether or not the similarity calculated by the calculation unit 461 is equal to or higher than a predetermined similarity. Specifically, for example, when a predetermined similarity level serving as a threshold is 0.5, between the above-described utterance section (utterance ID: call001-10) and the utterance section (utterance ID: call001-11) The similarity (= 0.66) is equal to or greater than a predetermined threshold. Therefore, it can be estimated that the same utterance is repeated in the continuous utterance section. Therefore, when the number of such continuous utterance sections is equal to or greater than a predetermined number, it is analyzed that the progress is not smooth.

  In the example of FIG. 12, only one set of continuous utterance sections has been described. However, the calculation unit 461 may calculate the similarity for each continuous utterance section, and the determination unit 462 may determine for each continuous utterance section. . That is, a plurality of sets of continuous utterance sections may be analyzed. In this case, the calculation unit 461 further counts the number of sets of continuous utterance sections having a predetermined similarity or higher, and determines whether or not the counted number of sets is equal to or higher than a predetermined number serving as a threshold value. It is good as well. In this case, when the number of pairs counted exceeds a predetermined number, it is analyzed that the progress is an unsuccessful dialogue. Therefore, the reliability of analysis is improved as compared with the case of analyzing a set of continuous speech sections.

<Determining unit 407>
In FIG. 4, the determination unit 407 has a function of determining a dialogue as a problem dialogue based on the analysis result analyzed by the dialogue structure analysis unit 404 and / or the utterance content analysis unit 406. Specifically, it is determined by the values of the dialog structure analysis result flag Fa, the utterance content analysis result flag Fb, and the AND result flag Fc of the call information table 300. Then, the call ID of the call information that becomes the problem dialogue is extracted.

  For example, when the problem dialogue is determined based on the utterance bias of the designated speaker (customer) after the answer phase, when the dialogue structure analysis result flag Fa is 1, the problem dialogue is determined. When Fa = 0, there is no bias in the dialogue, so the problem dialogue is not decided and treated as a normal dialogue. Similarly, when the problem dialogue is determined by the smoothness of the progress of the dialogue of the designated speaker (customer) after the answer phase, when the utterance content analysis result flag Fb is Fb = 1, the question dialogue is determined. When Fb = 0, since the dialogue is smooth, it is not determined as a problem dialogue and is treated as a normal dialogue.

  Further, when the problem dialogue is determined based on the utterance bias of the designated speaker (customer) after the answer phase and the smoothness of the progress of the dialogue, when the AND result flag Fc is Fc = 1, the question dialogue is decided. When Fc = 0, it is handled as a normal dialogue.

<Output unit 408>
The output unit 408 has a function of outputting the determination result determined by the determination unit 407. Specifically, for example, a problem list file in which call IDs extracted as problem dialogues are listed is output. FIG. 13 is an explanatory diagram showing a problem list file. The problem list file 1300 is output to an output device such as the display 208, the printer 212, or the I / F 209. The call ID described in the problem list file 1300 may be linked to the recorded voice data D. Thereby, when the problem list file 1300 is displayed on the display 208, the recorded audio data D can be reproduced by designating the call ID.

(Interaction screening process)
FIG. 14 is a flowchart showing a procedure for dialogue selection processing according to the present embodiment. First, the target recorded audio data D is acquired (step S1401), and the prosody information extraction process is executed by the extraction unit 401 (step S1402). Next, the speech section detection process is executed by the detection unit 402 (step S1403). Then, the voice recognition result information W is acquired by the acquisition unit 405 (step S1404).

  Thereafter, the setting process is executed (step S1405). In this setting process, setting of a designated speaker, setting of whether or not to use dialog structure analysis and / or utterance content analysis, setting of the number of target groups of continuous utterance sections in utterance content analysis, and the like are performed by user input. This setting process may be performed prior to steps S1401 to S1404.

Thereafter, a basic dialog analysis process (step S1406) by the basic dialog analysis unit 403, a dialog structure analysis process (step S1407) by the dialog structure analysis unit 404, and an utterance content analysis process (step S1408) by the utterance content analysis unit 406 are executed. . The dialog structure analysis process (step S1407) and the utterance content analysis process (step S1408) are executed according to the setting of the setting process. Thereafter, the determination process is executed by the determination unit 407 (step S1409), and the output process is executed by the output unit 408 (step S1410). As a result, a series of dialogue selection processing is completed.

<Speech section detection processing procedure>
FIG. 15 is a flowchart showing a detailed processing procedure of the speech segment detection processing (step S1403) by the detection unit 402. First, it is determined whether there is unprocessed prosody data among the prosody data Ma and Mc (step S1501). If there is unprocessed prosody data (step S1501: Yes), unprocessed prosody data is selected. (Step S1502). Then, s = 1 and r = 0 are set (step S1503). Here, s is a prosodic ID that identifies the power value. r is a counter that increments the prosodic ID: s.

  Then, it is determined whether or not s ≦ S (step S1504). S is the maximum value of s which is a prosodic ID. If s ≦ S is not satisfied (step S1504: NO), the process returns to step S1501. On the other hand, if s ≦ S (step S1504: Yes), it is determined whether P (s + r) ≧ Pt with respect to the lowest power value Pt at which the power value Ps of the prosody ID: s is a threshold value. (Step S1505).

  If P (s + r) ≧ Pt (step S1505: Yes), r is incremented (step S1506), and it is determined whether s + r ≦ S is satisfied (step S1507). If s + r ≦ S (step S1507: Yes), the continuous time t (s, s + r) is calculated (step S1508), and the process returns to step S1505. The continuous time t (s, s + r) is the total time from prosodic ID: s to prosodic ID: (s + r).

  On the other hand, if s + r ≦ S is not satisfied (step S1507: NO), it is determined whether or not the continuous time t (s, s + r−1) has been calculated (step S1514). If not calculated (step S1514: NO), the process proceeds to step S1501. On the other hand, if calculated (step S1514: Yes), is t (s, s + r-1)? T with respect to the predetermined continuous time T at which the continuous time t (s, s + r-1) is a threshold value? It is determined whether or not (step S1515). If t (s, s + r−1) ≧ T is not satisfied (step S1515: NO), the process proceeds to step S1501. On the other hand, if t (s, s + r−1) ≧ T (step S1515: Yes), the continuous time t (s, s + r−1) is stored as an utterance interval (step S1516).

  If P (s + r) ≧ Pt is not satisfied in step S1505 (step S1505: No), it is determined whether the continuous time t (s, s + r) has been calculated (step S1509). If not calculated (step S1509: No), the process proceeds to step S1512. On the other hand, if calculated (step S1509: Yes), it is determined whether or not t (s, s + r) ≧ T with respect to the predetermined continuous time T at which the continuous time t (s, s + r) is a threshold value. (Step S1510).

If t (s, s + r) ≧ T is not satisfied (step S1510: NO), the process proceeds to step S1512. On the other hand, if t (s, s + r) ≧ T (step S1510: Yes), the continuous time t (s, s + r) is stored as an utterance interval (step S1511). Thereafter, the prosodic ID: s is set to s = s + r + 1 (step S1512), the counter r is reset (r = 0) (step S1513), and the process returns to step S1505. On the other hand, if there is no unprocessed prosodic data in step S1501 (step S1501: No)
), The process proceeds to step S1404.

<Basic dialogue analysis processing procedure>
FIG. 16 is a flowchart showing a detailed processing procedure of the basic dialog analysis process (step S1406) by the basic dialog analysis unit 403. First, the initiative speaker identification unit 431 executes initiative speaker identification processing (step S1601), and the beginning phase identification unit 432 executes opening phase identification processing (step S1602). Then, control goes to a step S1407.

<Initiative speaker identification processing procedure>
FIG. 17 is a flowchart (part 1) illustrating a detailed processing procedure of the initiative speaker specifying process (step S1601) by the initiative speaker specifying unit 431. First, the section ID of the utterance section: x is set to x = 1 (step S1701), the section ID counter: i is set to i = 0, the customer utterance count counter: j is set to j = 0, the agent utterance count counter: k is set to k = 0. (Step S1702).

  Next, it is determined whether or not x ≦ X with respect to the total number X of utterance sections (step S1703). If x ≦ X is not satisfied (step S1703: NO), the process proceeds to step S1602. On the other hand, if x ≦ X is satisfied (step S1703: YES), it is determined whether La (x + i) <Lc (x + i) is satisfied (step S1704). La (x + i) is the section length of the utterance section of the agent with section ID: x + i, and Lc (x + i) is the section length of the utterance section of the customer with section ID: x + i. That is, it is determined which speaker's utterance is longer in the same section ID.

  If La (x + i) <Lc (x + i) (step S1704: Yes), the customer is longer, so the customer utterance number counter: j is incremented (step S1705), and the process proceeds to step S1707.

  On the other hand, if La (x + i) <Lc (x + i) is not satisfied (step S1704: No), the agent is longer, so the agent utterance number counter: k is incremented (step S1706), and the process proceeds to step S1707. In step S1707, the section ID counter: i is incremented (step S1707), and it is determined whether i ≦ n is satisfied (step S1708).

  Here, n is a predetermined number of sections for specifying the initiative speaker. If i ≦ n (step S1708: YES), the process returns to step S1704. On the other hand, if i ≦ n is not satisfied (step S1708: NO), the process proceeds to step S1801 in FIG.

  FIG. 18 is a flowchart (part 2) illustrating a detailed processing procedure of the initiative speaker specifying process (step S1601) by the initiative speaker specifying unit 431. In step S1801, if j> k (j ≧ k may be acceptable) (step S1801: Yes), the initiative speaker in the section [x, x + n−1] is determined as the customer (step S1802), and the process proceeds to step S1804. Transition.

  On the other hand, if j> k (j ≧ k may be acceptable) (step S1801: No), the initiative speaker in the section [x, x + n−1] is determined as the agent (step S1802), and the process proceeds to step S1804.

In step S1804, it is determined whether or not x> 1 (step S1804). If x> 1 is not satisfied (step S1804: No), the process proceeds to step S1808. On the other hand, if x> 1 (step S1804: Yes), it is determined whether or not the speaker determined as the initiative speaker (determined speaker) is the same speaker as the determined speaker in the immediately preceding section ( Step S1805).

  If they are the same (step S1805: Yes), the end point of the section immediately before the initiative speaker information Q is corrected to the end time of the section ID: x + n-1 (step S1808), and the process proceeds to step S1807. On the other hand, if they are not the same (step S1805: No), a new record of initiative speaker information Q is written with the start point as the start time of section x and the end point as the end time of section x + n-1 (step S1806). In step S1807, x is incremented, and the process returns to step S1702.

<Dialogue structure analysis processing procedure>
FIG. 19 is a flowchart showing a detailed processing procedure of the dialog structure analysis process (step S1407) by the dialog structure analysis unit 404. First, it is determined whether there is target call information (step S1901). Which call information is the target call information is set in the setting process (step S1405). By default, unprocessed call information may be sequentially targeted.

  If there is target call information (step S1901: Yes), the target call information is selected (step S1902), and the interaction time after the answer phase is calculated (step S1903). Then, it is determined whether or not the dialogue time after the answer phase is equal to or longer than a predetermined time as a threshold (step S1904). When it is not longer than the predetermined time (step S1904: No), the process returns to step S1901.

  On the other hand, if it is equal to or longer than the predetermined time (step S1904: Yes), the initiative holding time of each speaker after the answer phase is calculated (step S1905). The initiative holding time is the total section length for each speaker of the conversation section that has become the initiative speaker. Then, it is determined whether or not the designated speaker's initiative retention rate is equal to or higher than a predetermined retention rate that is a threshold value (step S1906). The initiative retention rate is a value obtained by dividing the initiative retention time of a designated speaker (for example, a customer) by the total initiative retention time of both speakers. If it is not equal to or higher than the predetermined retention rate (step S1906: NO), the process returns to step S1901.

  On the other hand, if it is equal to or higher than the predetermined retention rate (step S1906: Yes), the average utterance time length of the utterance of the designated speaker after the answer phase is calculated (step S1907). And it is judged whether it is more than the average utterance time length of all the designated speakers (step S1908). When it is not longer than the average utterance time length (step S1908: No), the process returns to step S1901.

  On the other hand, if it is equal to or longer than the average utterance time length (step S1908: Yes), the dialog structure analysis result flag Fa is set to Fa = 1 (step S1909) and the process returns to step S1901. On the other hand, in step S1901, if there is no target call information (step S1901: No), the process proceeds to step S1408.

<Speech content analysis processing procedure>
FIG. 20 is a flowchart showing a detailed processing procedure of the utterance content analysis process (step S1408) by the utterance content analysis unit 406. First, it is determined whether there is target call information (step S2001). Which call information is the target call information is set in the setting process (step S1405). By default, unprocessed call information may be sequentially targeted.

If there is target call information (step S2001: Yes), the target call information is selected (step S2002), and it is determined whether or not there is an unprocessed designated speaker after the answer phase (step S2003). If there is an utterance of an unprocessed predetermined speaker (step S2003: Yes), an unprocessed utterance is selected (step S2004), and whether or not the utterance time length of the selected utterance is greater than or equal to a predetermined time length that becomes a threshold Judgment is made (step S2005). When it is not longer than the predetermined time length (step S2005: No), the process returns to step S2003.

  On the other hand, when it is more than the predetermined time length (step S2005: Yes), the recognition keyword in the said speech area is extracted from the speech recognition result information W (step S2006). Then, the extracted keyword is written in the similarity calculation table 1200 (step S2007), and the process returns to step S2003.

  On the other hand, if there is no unprocessed designated speaker utterance after the answer phase in step S2003 (step S2003: No), the calculation unit 461 calculates the similarity between consecutive utterances of the designated speaker (step S2008). ). Then, for each continuous utterance section, the determination unit 462 determines whether the similarity is equal to or higher than a predetermined similarity, which is a threshold value, and the calculation unit 461 counts the number of continuous utterance sections that are equal to or higher than the predetermined similarity (step S2009).

  Then, the determination unit 462 determines whether or not the count value is a predetermined number or more (step S2010). Since this predetermined number is a threshold value, it is set in the setting process (step S1405). If the number is greater than or equal to the predetermined number (step S2010: Yes), the utterance content analysis result flag Fb is set to Fb = 1 (step S2011). On the other hand, when it is not more than the predetermined number (step S2010: No), the process returns to step S2001.

<Decision processing procedure>
FIG. 21 is a flowchart illustrating a detailed processing procedure of the determination process (step S1409) by the determination unit 407. First, it is determined whether there is target call information (step S2101). Which call information is the target call information is set in the setting process (step S1405). By default, unprocessed call information may be sequentially targeted.

  If there is target call information (step S2101: Yes), the target call information is selected (step S2102), and it is determined whether or not the flag value is 1 (step S2103). Here, whether the flag to be used is the dialog structure analysis result flag Fa, the utterance content analysis result flag Fb, or the AND result flag Fc is set in the setting process (step S1405). Refers to the value of the specified flag.

  When the value of the flag is not “1” (step S2103: No), the process returns to step S2101. On the other hand, if it is “1” (step S2103: Yes), the target call information is determined to be a problem dialogue (step S2104). Specifically, the call ID is written in the problem list file 1300. Then, the process returns to step S2101. On the other hand, when there is no target call information in step S2101 (step S2101: No), the process proceeds to an output process (step S1410).

  As described above, in this embodiment, it is possible to extract problem dialogues (candidates) whose dialogue is biased toward a designated speaker by analyzing the dialogue structure after the answer phase. In such a problem dialogue, it is presumed that the talks between the speakers are not engaged. Therefore, problem dialogues can be selected efficiently and with high accuracy, and monitoring work efficiency can be improved.

Further, by analyzing the utterance content of the designated speaker, it is possible to extract problem dialogues (candidates) in which the same utterance is repeated. In such a problem dialogue, it is presumed that the talks between the speakers are not engaged. Therefore, problem dialogues can be selected efficiently and with high accuracy, and monitoring work efficiency can be improved.

  Further, by performing both analysis of dialog structure analysis and utterance content analysis, it is possible to extract problem dialogs (candidates) in which the dialog is biased toward the designated speaker and the same utterance is repeated. In such a problem dialogue, it is estimated that the probability that the talk is not engaged between the speakers is higher than the case of either analysis. Therefore, problem dialogues can be selected efficiently and with high accuracy, and monitoring work efficiency can be improved.

  The dialog selection 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 is recorded on a computer-readable recording medium such as a hard disk, a flexible disk, a CD-ROM, an MO, and a DVD, and is executed by being read from the recording medium by the computer. The program may be a medium that can be distributed through a network such as the Internet.

  The following additional notes are disclosed with respect to the embodiment described above.

(Appendix 1) Computer
Detecting means for detecting a series of utterance sections for each speaker from phonological information for each speaker obtained from speech data relating to dialogue between speakers;
Based on the section length of a series of utterance sections for each speaker detected by the detection means, the initiative speaker identification means for identifying the initiative speaker for each specific conversation section of the two speakers,
A section in which the initiative speaker specified by the initiative speaker specifying means at the beginning of the voice data is a speaker who is in a position to ask a question to the other party is set as a question phase, and the initiative speaker after the question phase The initial phase identification means, where the answer phase is the section where the speaker is in the position of receiving a question from the other party,
After the answer phase specified by the opening phase specifying means, the initiative holding time in which the designated speaker is the initiative speaker among the two speakers is calculated, and the time length of the initiative holding time of the designated speaker A dialog structure analysis means for analyzing the bias of the dialog toward the designated speaker,
A determination unit that determines the dialogue as a problem dialogue based on the analysis result analyzed by the dialogue structure analysis unit;
Output means for outputting the determination result determined by the determination means;
Dialogue screening program characterized by functioning as

(Appendix 2)
An acquisition means for acquiring a recognition result for each speaker related to the voice data including a keyword matching or related to a predetermined recognition keyword and its appearance time;
Analyzing the smoothness of the progress of the dialogue based on the appearance status of the keyword selected from the recognition result of the designated speaker acquired by the acquiring means in the series of utterance sections of the designated speaker. Function as an utterance content analysis means,
The determining means includes
The dialogue selection program according to appendix 1, wherein the dialogue is determined to be a question dialogue based on the analysis result analyzed by the dialogue structure analysis means and the utterance content analysis means.

(Appendix 3)
The number of appearances of the same or similar keyword as the keyword selected from the recognition results of the designated speaker in the continuous utterance section of the designated speaker and the recognition result of the designated speaker in the continuous utterance section of the designated speaker Calculating means for calculating the similarity of the continuous utterance sections based on the number of appearances of all the keywords of
Functioning as a determination means for determining whether the similarity calculated by the calculation means is equal to or greater than a predetermined similarity;
The utterance content analysis means includes:
The dialogue selection program according to appendix 2, wherein the dialogue is analyzed as a dialogue whose progress is not smooth when the judgment means judges that the degree of similarity is equal to or higher than the predetermined similarity.

(Supplementary Note 4) The calculating means includes:
Counting the number of continuous speech sections determined by the determining means to be equal to or greater than the predetermined similarity,
The determination means includes
4. The dialogue selection program according to appendix 3, wherein it is determined whether or not the number calculated by the calculating means is a predetermined number or more, and the dialogue is analyzed as a dialogue whose progress is not smooth.

(Supplementary note 5)
In any one of Supplementary notes 1 to 4, wherein when the dialogue is analyzed by the dialogue structure analyzing unit as a dialogue biased toward the designated speaker, the dialogue is determined as a problem dialogue. The dialogue selection program described.

(Appendix 6) The determination means includes:
The dialogue selection according to any one of appendices 2 to 4, wherein when the dialogue is analyzed by the utterance content analysis means as a dialogue whose progress is not smooth, the dialogue is determined as a question dialogue. program

(Supplementary note 7)
When the dialog is analyzed by the dialog structure analysis means as a dialog biased toward the designated speaker, and the dialog is analyzed by the utterance content analysis means as a dialog whose progress is not smooth, The dialogue selection program according to any one of appendices 2 to 4, wherein the dialogue is determined as a problem dialogue.

(Appendix 8)
Detecting means for detecting a series of utterance sections for each speaker from phonological information for each speaker obtained from speech data relating to dialogue between speakers;
Based on the section length of a series of utterance sections for each speaker detected by the detection means, the initiative speaker identification means for identifying the initiative speaker for each specific conversation section of the two speakers,
A section in which the initiative speaker specified by the initiative speaker specifying means at the beginning of the voice data is a speaker who is in a position to ask a question to the other party is set as a question phase, and the initiative speaker after the question phase The initial phase identification means, where the answer phase is the section where the speaker is in the position of receiving a question from the other party,
An acquisition means for acquiring a recognition result for each speaker related to the voice data including a keyword matching or related to a predetermined recognition keyword and its appearance time;
Based on the appearance status of the designated speaker in the series of utterance sections of the keyword selected from the recognition result of the designated speaker acquired by the acquiring unit, by the opening phase specifying unit of the dialog Utterance content analysis means for analyzing the smoothness of progress after the identified answer phase,
A determination unit that determines the dialogue as a problem dialogue based on the analysis result analyzed by the utterance content analysis unit;
Output means for outputting the determination result determined by the determination means;
Dialogue screening program characterized by functioning as

(Supplementary note 9)
The number of appearances of the same or similar keyword as the keyword selected from the recognition results of the designated speaker in the continuous utterance section of the designated speaker and the recognition result of the designated speaker in the continuous utterance section of the designated speaker Calculating means for calculating the similarity of the continuous utterance sections based on the number of appearances of all the keywords of
Functioning as a determination means for determining whether the similarity calculated by the calculation means is equal to or greater than a predetermined similarity;
The utterance content analysis means includes:
9. The dialogue selection program according to appendix 8, wherein the dialogue is analyzed as a dialogue whose progress is not smooth when the judgment means judges that the degree of similarity is equal to or higher than the predetermined similarity.

(Supplementary Note 10) The calculating means includes:
Counting the number of continuous speech sections determined by the determining means to be equal to or greater than the predetermined similarity,
The determination means includes
The dialogue selection program according to appendix 9, wherein it is determined whether or not the number calculated by the calculation means is a predetermined number or more, and the dialogue is analyzed as a dialogue whose progress is not smooth.

(Supplementary Note 11) The determining means includes:
The dialogue selection according to any one of appendices 8 to 10, wherein when the dialogue is analyzed by the utterance content analysis means as a dialogue whose progress is not smooth, the dialogue is determined as a problem dialogue. program.

(Supplementary Note 12) Detection means for detecting a series of utterance sections for each speaker from phonological information for each speaker obtained from voice data related to conversation between speakers;
Based on the section length of a series of utterance sections for each of the speakers detected by the detection means, initiative speaker identification means for identifying the initiative speaker for each specific conversation section of the two speakers,
A section in which the initiative speaker specified by the initiative speaker specifying means at the beginning of the voice data is a speaker who is in a position to ask a question to the other party is set as a question phase, and the initiative speaker after the question phase An initial phase identification means in which an answer phase is a section in which the speaker is a speaker in a position to receive a question from the other party,
After the answer phase specified by the opening phase specifying means, the initiative holding time in which the designated speaker is the initiative speaker among the two speakers is calculated, and the time length of the initiative holding time of the designated speaker A dialog structure analyzing means for analyzing the bias of the dialog toward the designated speaker,
Determining means for determining the dialogue as a problem dialogue based on the analysis result analyzed by the dialogue structure analyzing means;
Output means for outputting the determination result determined by the determination means;
A dialog screening apparatus comprising:

(Supplementary note 13) Detection means for detecting a series of utterance sections for each speaker from phonological information for each speaker obtained from voice data related to conversation between speakers;
Based on the section length of a series of utterance sections for each of the speakers detected by the detection means, initiative speaker identification means for identifying the initiative speaker for each specific conversation section of the two speakers,
A section in which the initiative speaker specified by the initiative speaker specifying means at the beginning of the voice data is a speaker who is in a position to ask a question to the other party is set as a question phase, and the initiative speaker after the question phase An initial phase identification means in which an answer phase is a section in which the speaker is a speaker in a position to receive a question from the other party,
An acquisition means for acquiring a recognition result for each speaker related to the voice data including a keyword that matches or relates to a predetermined recognition keyword and its appearance time;
Based on the appearance status of the designated speaker in the series of utterance sections of the keyword selected from the recognition result of the designated speaker acquired by the acquiring unit, by the opening phase specifying unit of the dialog An utterance content analysis means for analyzing the smoothness of progress after the identified answer phase;
Determining means for determining the dialogue as a question dialogue based on the analysis result analyzed by the utterance content analyzing means;
Output means for outputting the determination result determined by the determination means;
A dialog screening apparatus comprising:

(Supplementary note 14)
A detection step for detecting a series of utterance sections for each speaker from phonological information for each speaker obtained from speech data relating to dialogue between speakers;
Based on the section length of a series of utterance intervals for each speaker detected by the detection step, the initiative speaker identification step for identifying the initiative speaker for each specific conversation interval of the two speakers,
A section in which the initiative speaker identified by the initiative speaker identification step at the beginning of the voice data is a speaker who is in a position to ask a question to the other party is defined as a question phase, and the initiative speaker after the question phase The initial phase identification process in which the answering phase is the section where the speaker is in a position to receive questions from the other party,
After the answer phase specified by the opening phase specifying step, the initiative holding time that the designated speaker is the initiative speaker among the two speakers is calculated, and the time length of the initiative holding time of the designated speaker Based on the dialog structure analysis step of analyzing the bias of the dialog to the designated speaker,
A determination step of determining the dialogue as a problem dialogue based on the analysis result analyzed by the dialogue structure analysis step;
An output step of outputting the determination result determined by the determination step;
Dialogue screening method characterized by executing

(Supplementary Note 15) A detection step of detecting a series of utterance sections for each speaker from phonological information for each speaker obtained from voice data related to dialogue between speakers;
Based on the section length of a series of utterance intervals for each speaker detected by the detection step, the initiative speaker identification step for identifying the initiative speaker for each specific conversation interval of the two speakers,
A section in which the initiative speaker identified by the initiative speaker identification step at the beginning of the voice data is a speaker who is in a position to ask a question to the other party is defined as a question phase, and the initiative speaker after the question phase The initial phase identification process in which the answering phase is the section where the speaker is in a position to receive questions from the other party,
An acquisition step of acquiring a recognition result for each speaker related to the voice data including a keyword that matches or relates to a predetermined recognition keyword and its appearance time;
Based on the appearance status of the designated speaker in the series of utterance sections of the keyword selected from the recognition result of the designated speaker acquired by the acquiring step, by the opening phase specifying step of the dialogue Utterance content analysis process that analyzes the smoothness of the progress after the identified answer phase,
A determination step of determining the dialogue as a question dialogue based on the analysis result analyzed by the utterance content analysis step;
An output step of outputting the determination result determined by the determination step;
Dialogue screening method characterized by executing

400 Dialogue Sorting Device 401 Extraction Unit 402 Detection Unit 403 Basic Dialogue Analysis Unit 404 Dialogue Structure Analysis Unit 405 Acquisition Unit 406 Utterance Content Analysis Unit 407 Determination Unit 408 Output Unit 431 Initiative Speaker Specification Unit 432 Opening Phase Specification Unit 461 Calculation Unit 462 Judgment part

Claims (6)

Computer
Detecting means for detecting a series of utterance sections for each speaker from phonological information for each speaker obtained from speech data relating to dialogue between speakers;
Based on the section length of a series of utterance sections for each speaker detected by the detection means, the initiative speaker identification means for identifying the initiative speaker for each specific conversation section of the two speakers,
A section in which the initiative speaker specified by the initiative speaker specifying means at the beginning of the voice data is a speaker who is in a position to ask a question to the other party is set as a question phase, and the initiative speaker after the question phase The initial phase identification means, where the answer phase is the section where the speaker is in the position of receiving a question from the other party,
After the answer phase specified by the opening phase specifying means, the initiative holding time in which the designated speaker is the initiative speaker among the two speakers is calculated, and the time length of the initiative holding time of the designated speaker A dialog structure analysis means for analyzing the bias of the dialog toward the designated speaker,
A determination unit that determines the dialogue as a problem dialogue based on the analysis result analyzed by the dialogue structure analysis unit;
Output means for outputting the determination result determined by the determination means;
Dialogue screening program characterized by functioning as The computer,
An acquisition means for acquiring a recognition result for each speaker related to the voice data including a keyword matching or related to a predetermined recognition keyword and its appearance time;
Analyzing the smoothness of the progress of the dialogue based on the appearance status of the keyword selected from the recognition result of the designated speaker acquired by the acquiring means in the series of utterance sections of the designated speaker. Function as an utterance content analysis means,
The determining means includes
2. The dialogue selection program according to claim 1, wherein the dialogue is determined as a question dialogue based on the analysis result analyzed by the dialogue structure analyzing unit and the utterance content analyzing unit. The computer,
The number of appearances of the same or similar keyword as the keyword selected from the recognition results of the designated speaker in the continuous utterance section of the designated speaker and the recognition result of the designated speaker in the continuous utterance section of the designated speaker Calculating means for calculating the similarity of the continuous utterance sections based on the number of appearances of all the keywords of
Functioning as a determination means for determining whether the similarity calculated by the calculation means is equal to or greater than a predetermined similarity;
The utterance content analysis means includes:
3. The dialogue selection program according to claim 2, wherein if the judgment means judges that the degree of similarity is equal to or higher than the predetermined similarity, the dialogue is analyzed as a dialogue whose progress is not smooth. Computer
Detecting means for detecting a series of utterance sections for each speaker from phonological information for each speaker obtained from speech data relating to dialogue between speakers;
Based on the section length of a series of utterance sections for each speaker detected by the detection means, the initiative speaker identification means for identifying the initiative speaker for each specific conversation section of the two speakers,
A section in which the initiative speaker specified by the initiative speaker specifying means at the beginning of the voice data is a speaker who is in a position to ask a question to the other party is set as a question phase, and the initiative speaker after the question phase The initial phase identification means, where the answer phase is the section where the speaker is in the position of receiving a question from the other party,
An acquisition means for acquiring a recognition result for each speaker related to the voice data including a keyword matching or related to a predetermined recognition keyword and its appearance time;
Based on the appearance status of the designated speaker in the series of utterance sections of the keyword selected from the recognition result of the designated speaker acquired by the acquiring unit, by the opening phase specifying unit of the dialog Utterance content analysis means for analyzing the smoothness of progress after the identified answer phase,
A determination unit that determines the dialogue as a problem dialogue based on the analysis result analyzed by the utterance content analysis unit;
Output means for outputting the determination result determined by the determination means;
Dialogue screening program characterized by functioning as Detecting means for detecting a series of utterance sections for each speaker from phonological information for each speaker obtained from speech data relating to conversation between speakers;
Based on the section length of a series of utterance sections for each of the speakers detected by the detection means, initiative speaker identification means for identifying the initiative speaker for each specific conversation section of the two speakers,
A section in which the initiative speaker specified by the initiative speaker specifying means at the beginning of the voice data is a speaker who is in a position to ask a question to the other party is set as a question phase, and the initiative speaker after the question phase An initial phase identification means in which an answer phase is a section in which the speaker is a speaker in a position to receive a question from the other party,
After the answer phase specified by the opening phase specifying means, the initiative holding time in which the designated speaker is the initiative speaker among the two speakers is calculated, and the time length of the initiative holding time of the designated speaker A dialog structure analyzing means for analyzing the bias of the dialog toward the designated speaker,
Determining means for determining the dialogue as a problem dialogue based on the analysis result analyzed by the dialogue structure analyzing means;
Output means for outputting the determination result determined by the determination means;
A dialog screening apparatus comprising: Computer
A detection step for detecting a series of utterance sections for each speaker from phonological information for each speaker obtained from speech data relating to dialogue between speakers;
Based on the section length of a series of utterance intervals for each speaker detected by the detection step, the initiative speaker identification step for identifying the initiative speaker for each specific conversation interval of the two speakers,
A section in which the initiative speaker identified by the initiative speaker identification step at the beginning of the voice data is a speaker who is in a position to ask a question to the other party is defined as a question phase, and the initiative speaker after the question phase The initial phase identification process in which the answering phase is the section where the speaker is in a position to receive questions from the other party,
After the answer phase specified by the opening phase specifying step, the initiative holding time that the designated speaker is the initiative speaker among the two speakers is calculated, and the time length of the initiative holding time of the designated speaker Based on the dialog structure analysis step of analyzing the bias of the dialog to the designated speaker,
A determination step of determining the dialogue as a problem dialogue based on the analysis result analyzed by the dialogue structure analysis step;
An output step of outputting the determination result determined by the determination step;
Dialogue screening method characterized by executing

Images