JP2011053563A - Collation system of voice keyword in voice data, method thereof, and collation program of voice keyword in voice data - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To achieve collation of voice keyword in voice data with a practical search precision without requiring voice recognition processing by a highly versatile and simple configuration. <P>SOLUTION: The collation system includes: a voice waveform-combining part which combines a standard voice waveform pattern and a plurality of derived voice waveform patterns different from the standard voice waveform pattern in speed and/or volume, from an input search keyword text; a similarity threshold calculation part which variably sets a similarity threshold for collation by raising or reducing a reference threshold; and a voice keyword collation part which refers to a voice keyword database and successively compares the standard voice waveform pattern and the plurality of derived voice waveform patterns with a voice waveform pattern of read voice data to calculate similarities and obtains positions of voice keywords of which the calculated similarities are equal to or higher than the similarity threshold. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a speech keyword matching system in speech data, a method thereof, and a speech keyword matching program. More specifically, for example, in a Customer Relationship Management (CRM) system that records and manages calls made between a customer's phone and an agent's phone, the entered keyword text for search is recorded and accumulated. The present invention relates to a technique for collating voice keywords in a call voice and extracting and reproducing the collated voice keywords.

  Various technologies have been proposed for recording and managing the voice of calls made between customers and businesses on the business side.

  For example, in a centralized call recording system for recording and searching for the contents of calls made by operators in a call center, which is a telephone reception department from a customer, in general, the premises of a call center etc. operated by a business operator are not included. An exchange (PBX) in which outgoing calls and incoming calls are concentrated from a public switched telephone network (PSTN) is installed, and the voice call is distributed to a plurality of fixed telephones in the call center. For this reason, if a call recording server branched from this exchange is provided, calls can be recorded and stored. On the operator side, a terminal device such as a PC may be provided together with a telephone answering extension telephone, and this operator terminal device has a function for searching customer information using the customer name given by the speaker as a key, and the customer A function may be provided for displaying the past call history.

JP-A-8-249343 JP 2009-58548 A

  By the way, it is requested that a specific word or phrase can be searched after the call ends from the call voice recorded and stored in the voice data file. This is because, for example, the operator himself / herself who handled the phone call to the customer or the manager thereof needs to quickly check the reproduction of the call voice specifying the important part of the call recording content, It is also necessary for the manager of quality and compliance in the telephone reception of the operator to promptly confirm whether the operator is not speaking a word or sentence prohibited by law or compliance to the customer. In this case, it takes a long time to reproduce the entire call voice having an enormous amount of data, or to specify the call time zone and to play the call voice, and it takes a lot of time to detect a specific word or phrase. This is because it involves a lot of labor.

  In the speech keyword search technology for searching for specific words or sentences from the speech data file, a technology for converting speech in the speech data file into text by speech recognition processing and extracting specific characters or sentences from the text data is known. It is.

  For example, Patent Document 1 converts external real-time audio data into digital audio data to obtain an audio file, converts audio data extracted from the audio file into text data by audio recognition processing, and includes keywords in the text data. Discloses a technique for automatically generating a search key for external data search by determining the appearance frequency of.

  However, converting voice data into text data by voice recognition processing is complicated because it requires registration of a voice recognition dictionary to be referred to by the voice recognition engine in advance, and further managing and updating the recognition dictionary. In addition, it lacks versatility, and the processing load for speech recognition and morphological analysis is high, so hardware facilities are inevitably expensive.

  In the first place, in call center operations, call voices are recorded and accumulated all day for each of a large number of operators. Therefore, it is difficult to convert all of the accumulated call recording data into text data.

  On the other hand, in a voice keyword search technique for searching for a specific word or phrase from a voice data file, a technique for extracting a specific character or sentence from the voice data itself as a search target is also known.

  For example, Patent Literature 2 calculates a time series shape of a feature amount (speech height, size, length) of a speech keyword input as a search condition, and a time series shape of the feature amount of the input keyword; Disclosed is a technique for obtaining a difference between a voice feature amount stored in a voice database and a time-series shape and outputting a voice whose difference is equal to or less than a predetermined threshold as a search result.

  However, the comparison between the time-series shapes of the voice feature amounts is actually Exact Matching that does not allow ambiguous search, especially when assuming a variety of customer-side speakers and utterance situations, or replacing short phrases When a sentence longer than this is used as a search keyword, the search accuracy is significantly impaired.

  Also, as described above, in call center operations, all day call voices are recorded and stored for each of a large number of operators. In order to reduce the file size, the recorded and stored data is compressed and compressed call voice. Usually, it is stored as a file in a large-scale storage device, and the audio file before compression is not separately stored for a long time. However, in order to enjoy sufficient compression efficiency, it is necessary to perform so-called irreversible compression processing in which a part of the pre-compression voice data is lost to obtain the post-compression voice data. Even if the Exact Matching process is performed, there is a problem in that the speech keyword to be searched is not extracted and the search accuracy is lowered.

  SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned problems, and an object thereof is to provide a speech recognition process suitable for a CRM system for recording and storing and managing calls made between a customer's phone and a customer's phone. A speech keyword matching system in speech data, which is highly versatile and can obtain practical search accuracy with a simple configuration, its method, and a speech keyword matching program in speech data It is in.

  Another object of the present invention is to enable speech keyword matching for speech data with a simple configuration and practical search accuracy even when various customer-side speakers and speech situations are assumed. It is in.

  Another object of the present invention is to enable speech keyword matching for speech data with a simple configuration and practical search accuracy even when a longer sentence is used as a search keyword instead of a short word. There is in point to do.

  Another object of the present invention is to enable voice keyword matching for voice data without impairing search accuracy even when the voice data after compression is targeted.

  The inventors of the present application conducted a speech keyword search using speech data whose target is speech produced by various speakers, and performed speech data matching using a plurality of speech waveform patterns according to the diversity of the speaker population. The knowledge that it is efficient was obtained.

  Based on this knowledge, in the present invention, when synthesizing a speech waveform pattern from a search keyword input as text, a plurality of different speech waveform patterns are synthesized. Preferably, for example, a plurality of speech waveform patterns whose speech speeds are high speed, normal speed, and low speed may be synthesized from the input search keyword text. A plurality of speech waveform patterns whose utterance volume is low volume, normal volume, and high volume may be synthesized from the input search keyword text. Moreover, you may synthesize | combine a some audio | voice waveform pattern from the input search keyword text according to the difference in voice quality, such as sex and age group, for example.

  According to one aspect of the present invention, an audio database that stores audio data so as to be reproducible, and a search keyword input unit that inputs search keyword text and search conditions for collating audio keywords to be reproduced in the audio database; A speech waveform synthesizer that synthesizes a standard speech waveform pattern from the input search keyword text and synthesizes a plurality of derived speech waveform patterns whose speech speed and / or volume are different from the standard speech waveform pattern; A similarity threshold calculation unit that variably sets a threshold of similarity in matching keyword text and the voice keyword by increasing or decreasing a reference threshold, and a voice that matches the search condition with reference to the voice keyword database Reading data, and the standard speech waveform pattern and the plurality of Speech keyword collation that sequentially compares a raw speech waveform pattern with the speech waveform pattern of the speech data that has been read, calculates the similarity, and obtains the location of the speech keyword for which the similarity equal to or greater than the similarity threshold is calculated And an output unit that transmits the position of the obtained voice keyword to the terminal device, thereby enabling reproduction of voice data from the position of the voice keyword obtained on the terminal device. Is provided.

  The voice keyword matching unit, when either inbound utterance or outbound utterance is selected and input as the search condition, refers to call control information corresponding to the voice data and is selectively input The speech waveform pattern of only one of the utterances and the outbound utterances may be limited as the target of collation.

  The speech keyword matching unit performs a first comparison process between a derived speech waveform pattern having a higher speech speed and / or lower volume than the standard speech waveform pattern and the speech waveform pattern of the read speech data, If no voice keyword candidate is obtained in the first comparison, a second comparison process is performed between the standard waveform pattern and the waveform pattern of the read voice data, and the second comparison If no voice keyword candidate is obtained in the comparison process, the derived speech waveform pattern having a lower speech speed and / or louder volume than the standard speech waveform pattern and the speech waveform pattern of the speech data read out. A third comparison process may be performed.

  The similarity threshold calculation unit decreases the reference threshold when the number of characters in the search keyword text is large, and increases the threshold from the reference threshold when the number of characters in the search keyword text is small. May be set.

  The similarity threshold calculation unit is configured to decrease from the reference threshold when the compression rate of the audio file in the audio database is high, and to increase from the reference threshold when the compression rate of the audio file is low. A similarity threshold may be set.

  The voice waveform synthesis unit may synthesize a plurality of derived voice waveform patterns having different voice qualities characterized by sex and / or age group from the input search keyword text.

  According to another aspect of the present invention, a speech keyword matching server device comprising a speech database, a search keyword input unit, a speech waveform synthesis unit, a similarity threshold calculation unit, a speech keyword collation unit, and an output unit. Is a speech keyword matching method executed by the search keyword input unit for inputting search keyword text and search conditions for matching a speech keyword to be reproduced in a speech database storing speech data reproducibly. Then, the voice waveform synthesizer synthesizes a standard voice waveform pattern from the input search keyword text, and synthesizes a plurality of derived voice waveform patterns having different speech speed and / or volume from the standard voice waveform pattern. Step and the similarity calculation unit, the search keyword text and the voice key. A step of variably setting a threshold value of similarity in collation with a word by increasing or decreasing a reference threshold value, and referring to the speech keyword database by the speech keyword collation unit, for speech data that matches the search condition In addition, the similarity is calculated by sequentially comparing the standard speech waveform pattern and the plurality of derived speech waveform patterns with the speech waveform pattern of the speech data that has been read, and the similarity is equal to or greater than the similarity threshold. The step of obtaining the position of the calculated voice keyword, and the output unit transmits the position of the obtained voice keyword to the terminal device, whereby the voice data from the position of the voice keyword obtained on the terminal device is transmitted. A speech keyword matching method comprising the steps of enabling playback.

According to another aspect of the present invention, there is provided a speech keyword collation program for causing a computer to execute speech keyword collation processing, wherein the program stores speech data in a speech database so that speech data can be reproduced. A storage process, a search keyword input process for inputting a search keyword text and a search condition for matching a voice keyword to be reproduced in the voice database, a standard voice waveform pattern is synthesized from the input search keyword text, A threshold value of similarity in speech waveform synthesis processing for synthesizing a plurality of derived speech waveform patterns whose speech speed and / or volume is different from the standard speech waveform pattern, and matching between the search keyword text and the speech keyword, Set to variable by increasing or decreasing the threshold With reference to the similarity threshold value calculation process and the voice keyword database, the voice data matching the search condition is read out, and the standard voice waveform pattern and the plurality of derived voice waveform patterns are read out. A speech keyword collation process for calculating the similarity by sequentially comparing with a speech waveform pattern and obtaining a position of the speech keyword for which a similarity equal to or greater than the threshold of the similarity is calculated; A voice keyword collation for executing a process including an output process that enables reproduction of voice data from the position of the voice keyword obtained on the terminal device A program is provided.

  According to the present invention, the speech keyword matching server synthesizes a plurality of speech waveform patterns having different utterance speed, speech volume, voice quality, and the like from the input keyword, and further, the plurality of speech waveform patterns and searched speech data. Search target voice data in which the similarity threshold used as the threshold for matching is variably set according to the number of characters of the input keyword and the compression rate of the voice data, and the similarity equal to or greater than the calculated similarity threshold is calculated. Is specified, and the specified location of the searched audio data is output so that the audio reproduction of this location is possible.

  As a result, it is suitable for a CRM system that records and manages calls made between the customer's phone and the customer's phone, and is practical with a highly versatile and simple configuration that does not require voice recognition processing. Voice keyword matching in voice data with search accuracy is realized.

  Further, even when various customer-side utterers and utterance situations are assumed, it is possible to perform speech keyword collation for speech data with a simple configuration and practical search accuracy.

  Further, even when a sentence longer than this is used as a search keyword instead of a short phrase, it is possible to perform speech keyword collation for speech data with a simple configuration and practical search accuracy.

  Furthermore, even when the compressed voice data is targeted, it is possible to perform voice keyword matching on the voice data without impairing the search accuracy.

  Therefore, according to the speech keyword collation system in speech data, the method thereof, and the speech keyword collation program in speech data according to the present invention, the input keyword is used as a search key without additional equipment in call center operations. Voice files can be searched quickly and with high accuracy, and voice data can be reproduced from the searched voice parts as necessary, which contributes to the improvement of the CRM of the operator.

It is a block diagram which shows an example of the network structure of the speech keyword collation system in the speech data which concerns on one Embodiment of this invention. In the voice keyword matching system according to the embodiment of the present invention, the recorded voice from the start position of the voice keyword checked on the compressed call voice file 31 from the login of the operator to the voice keyword search PC terminal 9b and the input of the search keyword. Is a diagram showing an example of a processing sequence until the user is reproduced on the speech keyword search PC terminal 9b. It is a block diagram which shows an example of a function structure of the speech keyword collation system which concerns on one Embodiment of this invention. It is a flowchart which shows an example of the process sequence of the speech keyword collation system which concerns on one Embodiment of this invention. It is a flowchart which shows an example of the collation processing procedure with the several audio | voice waveform pattern and call recording audio | voice file which the audio | voice keyword collation part of FIG. 3 performs. It is a schematic diagram which shows an example of the standard audio | voice waveform pattern synthesize | combined from the search keyword text input into the search keyword input part. It is a schematic diagram which shows the example of the derived speech waveform pattern converted so that the conversation speed may differ from the standard speech waveform pattern of FIG. It is a schematic diagram which shows the example of the derived audio | voice waveform pattern converted so that the conversation volume may differ from the standard audio | voice waveform pattern of FIG. It is a figure explaining the collation process of the search keyword and audio | voice file by the sound pressure (power) strength pattern in an audio | voice keyword collation part. It is a figure explaining the collation process of the search keyword and audio | voice file by the periodic pattern which pitch-analyzed in the audio | voice keyword collation part. It is a figure explaining the collation process of the search keyword and audio | voice file by the phonetic utterance pattern by a formant analysis in an audio | voice keyword collation part. It is a figure explaining the collation process of the search keyword and audio | voice file by the specific pattern of the utterance location by voiceprint analysis in an audio | voice keyword collation part. It is a schematic diagram explaining the labeling process by the phoneme unit with respect to a sound recording. It is a schematic diagram explaining the pitch unification process with respect to a sound recording. It is a schematic diagram explaining the sound volume and speed unification processing with respect to the recorded sound. It is a schematic diagram explaining the process which applies a fast Fourier transform to the time signal of a sound recording voice, and obtains an audio spectrogram. It is a schematic diagram explaining the process which detects a formant from the envelope with respect to an audio | voice spectrogram. It is a schematic diagram explaining the change on the time axis of an audio spectrogram. It is a figure which shows an example of the hardware constitutions of each server apparatus which concerns on this embodiment. It is a block diagram which shows an example of a function structure of the speech keyword collation system in the 2nd Embodiment of this invention. It is a flowchart which shows an example of the speech waveform pattern estimation processing procedure in the 2nd Embodiment of this invention. It is a flowchart which shows an example of the collation processing procedure with the some audio | voice waveform pattern and call recording audio | voice file in the 2nd Embodiment of this invention.

  Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In addition, in this specification and drawing, about the component which has the substantially same function and structure, the duplicate description is abbreviate | omitted by attaching | subjecting the same code | symbol.

First Embodiment <Network Configuration of this Embodiment>
FIG. 1 shows a non-limiting example of a network configuration of a speech keyword matching system in speech data according to the first embodiment of the present invention. The voice keyword matching system in the voice data includes a PBX (switch) 1, a voice acquisition server 2, a call recording server 3, a control server 4, a voice keyword matching server 5, a customer telephone terminal 7, a PSTN (public telephone network) 8, an operator A telephone terminal 9a and a voice keyword search PC terminal 9b are provided. During voice keyword matching, all or part of PBX (switch) 1, voice acquisition server 2, call recording server 3, control server 4, voice keyword matching server 5, operator telephone terminal 9a, and voice keyword search PC terminal 9b are And may be interconnected by an IP (Internet Protocol) network such as an intranet 11d such as a LAN / WAN. Alternatively, the voice acquisition server 2, the call recording server 3, the control server 4, the voice keyword matching server 5, and the compressed call voice file 31 and the call information database 32 included in these servers are all or part of the Internet It may be installed outside the call center as appropriate via a remote IP connection. In particular, when an administrator other than the operator of the call center operates the voice keyword search PC terminal 9b to perform the voice keyword extraction process in the compressed call voice file 31, the voice keyword search PC terminal 9b is the operator phone terminal 9a. It is not necessary to be installed in the vicinity of the call center, and it is preferable that it is installed outside the call center as appropriate via a remote IP connection.

  The PBX 1 connects the extension telephones in the call center to each other, and connects each operator telephone terminal 9a to the PSTN (public telephone network) 8 via the local lines 11a, 11b, 11c. A telephone call between the telephone terminal 9a and the customer telephone terminal 7 is realized.

  The voice acquisition server 2 is branched and connected to the PBX 1, acquires call voices between the operator telephone terminals 9a and the customer telephone terminals 7, and associates the acquired voices with numbers (for example, extension numbers) of the operator telephone terminals 9a. Supply to each server.

  Alternatively, the voice acquisition server 2 may be branched and connected to a line between the terminating device (DSU) of the PSTN 8 and the PBX 1.

  The call recording server 3 compresses the acquired voice supplied from the voice acquisition server 2 after the incoming call as necessary under the control of the control server 4 and converts the compressed voice data into, for example, NAS (Network Application Storage). Are stored in a database composed of a large-scale external storage device such as

  Preferably, when the voice recording server 3 is supplied with analog voice from the voice acquisition server 2, the voice recording server 3 samples the analog voice waveform in voltage with a predetermined bit depth and a predetermined sampling frequency, thereby digital audio. More preferably, the digital voice data is compressed and stored in the compressed call voice file 31. By using the compressed call voice file 31 as a target for voice keyword detection, it becomes unnecessary to separately store and store the voice data before compression, and the required storage resource capacity can be reduced. Various known methods can be used for compressing the recorded sound at various compression rates. As a non-limiting example, recording is performed by monaural 1/5 compression, monaural 1/10 compression, or stereo no compression. Audio is compressed. Alternatively, the call recording server 3 stores and saves the voice data supplied from the voice acquisition server 2 in a call voice file without converting and compressing the voice data, and the voice keyword matching server 10 stores the uncompressed call voice file as a voice. It is good also as an object of keyword collation.

  The call recording server 3 also writes the call information of the voice file recorded in the call information file 32 in association with the call voice file stored and stored in the compressed call voice file 31. This call information is acquired by the PBX 1 when an incoming call is made to the PBX 1. Call information acquired includes, for example, incoming call start information (including an incoming call start time stamp), outgoing call start information (including a outgoing call start time stamp), call start information (including a call start time stamp), and call end information ( Call control information (including call end time stamps) and calls such as caller phone number, callee phone number, caller channel number, caller number, caller channel number, callee phone number (destination extension number, etc.) Including identification information.

  This call information further includes speaker identification information that identifies the polarity of whether the utterance in the recorded call is inbound, ie, from the customer side, or outbound, ie, from the operator side. The speaker identification information can be acquired by the PBX 1. For example, in the case of ISDN, the speaker identification information can be grasped as a physical pin position of a line termination unit (Digital Service Unit: DSU). In the case of the SIP (Session Initiation Protocol) protocol, it is possible to grasp at the time of session configuration at the time of call generation. Specifically, for example, an Invite command transmitted from the calling side to the called side at the time of session configuration. In the SDP (Session Description Protocol) that describes the information required to start a session, the IP address and port number used for reception by the calling party are specified, and in response, from the called party to the calling party The IP address and port number used by the called party for reception are specified in the SDP in the 200 OK message to be transmitted, and the voice on the RTP (Realtime Transport Protocol) protocol is used using the specified IP address and port number. Day There are sent and received. Therefore, by acquiring the IP address and port number used for reception by each of the calling side and the called side, it is possible to obtain the speaker identification information for each utterance in one call, and for the customer in one call. The utterance and the operator's utterance can be distinguished or separated as necessary.

  The call information is preferably displayed in real time on these display devices in conjunction with a CTI program running on the control server 4 or the operator PC terminal device that implements the CTI (Computer Telephony Integration) protocol. May be displayed.

  The call recording server 3 is also provided with a customer information database 33 in which customer information centered on customers who have already received a response history is pre-registered. This customer information is personal information for identifying the customer, and includes, for example, the customer name, address, registered customer telephone number, date of birth, age group, gender, other customer attributes, product purchase history, response history, etc. In addition, it can be appropriately displayed and output on a terminal device that can be operated by the operator in response to an instruction input by the operator.

  The call recording server 3 may alternatively be connected between, for example, the PSTN 8 and the PBX 1 instead of being connected to the local line 11d. With this configuration, the call recording server 3 Can be obtained directly. Further alternatively, the call recording server 3 may be connected to the local line without directly installing the voice acquisition server 2 and directly acquire the call voice supplied to the local line.

  Based on the data and control information supplied from the voice acquisition server 2, the call recording server 3, and the voice keyword matching server 5, the control server 4 performs processing executed by these servers, and transmits and receives data traffic and control information between these servers. To control. Alternatively, the voice keyword matching server 5 can access the compressed call voice file 31 and the call information file 32 held by the call recording server 3 and interface to the voice keyword search PC terminal 9b without using the control server 4. May be provided directly.

  The speech keyword matching server 5 receives a search keyword that is text-inputted from, for example, hiragana or katakana from the speech keyword search PC terminal 9b, converts the search keyword into a plurality of speech waveform patterns, and uses the speech waveform pattern. The voice keyword in the compressed call voice file 31 is checked, and the position information of the checked voice keyword, that is, the start position on the compressed call voice file 31 is supplied to the voice keyword search PC terminal 9 as a search result.

  The supplied position information of one or more collated voice keywords is preferably displayed in a list on the display screen of the voice keyword search PC terminal 9. The position information of the speech keyword selected by the selection input for voice reproduction by the operator from the position information of one or more collated speech keywords displayed in the list is the speech keyword collation server 5 or the control server. 4, the voice keyword matching server 5 or the control server 4 reads the call voice data from the position information of the selected voice keyword in the compressed call voice file 31 and supplies it to the voice keyword search PC terminal 9 b for voice keyword search. The reproduced sound is output on the PC terminal 9b.

  The PBX 1 in FIG. 1 is connected to the customer call terminal 4 through a public switched telephone network such as PSTN 1, but instead of or in addition to this, by providing an IP network connection function, The voice acquisition server 2 may be connected to a customer IP call terminal and an operator phone terminal via a voice packet communication network such as a VoIP (Voice Over Internet Protocol) network. A voice call between 9a can be acquired. The customer phone terminal 7 may be a fixed phone or a mobile phone.

  Further, the configuration of the network and hardware shown in FIG. 1 is merely an example, and each server and database may be integrated as necessary, and each component may be installed outside an ASP (Application Service Provider) or the like. .

<Voice Keyword Matching Processing Sequence and User Interface in First Embodiment>
FIG. 2 illustrates an operation of logging in an operator to the spoken keyword search PC terminal 9b and inputting a search keyword in the spoken keyword matching system according to the first embodiment, which is executed under the control of the control server 4 as necessary. A non-limiting example of a processing sequence from the start position of the voice keyword collated on the compressed call voice file 31 to the reproduction of the recorded voice on the voice keyword search PC terminal 9b is shown.

  In FIG. 2, first, the voice keyword search PC terminal 9 b prompts the voice keyword matching server 5 to input the user ID and password on the display screen, typically, to log in (step S <b> 1). The voice keyword matching server 5 authenticates the input user ID and password with reference to authentication information set for each operator and stored in the storage device, and transmits a search condition input request when the authentication is successful. (Step S2).

  The speech keyword search PC terminal 9b that has received the search condition input request outputs a search condition input request screen on the display device. This search condition input request screen includes an input field for inputting text at least in the search keyword using hiragana, katakana, or the like. As a non-limiting example, the operator himself / herself may enter words such as “chumon (order)” and “bait (compensation)” for the purpose of confirming the contents of the telephone response. Therefore, for example, a sentence such as “I will definitely pay” or “Please leave it” may be entered. Alternatively, Chinese characters or alphabets may be input in this input field, and instead, the operator's voice may be directly input via a voice input means such as a microphone. A process for synthesizing the retrieved search keyword into the speech waveform pattern is not necessary.

  Preferably, the search condition input request screen further includes, for example, a call start date and time, a call end date and time, a recording time, a caller telephone number, a callee telephone number, and a caller to narrow down a compressed call voice file to be searched. An input field for inputting a search condition such as a name of a person, an identifier of whether the call is an outgoing call or a received call from the operator side, and a base name of a call center or a business office may be provided.

  More preferably, the search condition input screen may include an input field for selecting whether a customer's utterance or an operator's utterance during a call is to be searched as a search condition.

  When the search keyword text and the search condition are input via the search condition input request screen of the voice keyword search PC terminal 9b, the search keyword text and the search condition are transmitted to the voice keyword matching server 5 (step S3). .

  The voice keyword matching server 5 generates a search target voice acquisition request from the received search conditions (step S4), and searches the compressed call voice file 31 held by the call recording server 3 and the call information database 32 as necessary. As a result, the compressed call voice file 31 to be searched is acquired (step S5).

  The speech keyword matching server 5 synthesizes a plurality of speech waveform patterns from the input search keyword text for the acquired compressed call speech file 31, and compresses the synthesized speech waveform patterns into the compressed call speech file 31. The voice keyword is collated (step S6). Details of the voice keyword matching process will be described later.

  The voice keyword matching server 5 transmits at least the position information in the compressed call voice file 31 of the voice keyword obtained as a result of the voice keyword matching process to the voice keyword search PC terminal 9b as a search result in response to step S3. The voice keyword search PC terminal 9b displays and outputs the position information and the voice keyword in the compressed call voice file 31 of the received voice keyword on the display device via the search result screen (step S7).

  Preferably, when there are a plurality of pieces of position information in the compressed call voice file 31 of the voice keywords obtained as a result of the voice keyword collating process, the search result screen displays in the compressed call voice file 31 of the plurality of voice keywords. It is possible to display a list of position information. Alternatively, on the initial search result screen, a call voice file corresponding to each call including a voice keyword obtained as a result of the voice keyword matching process is displayed in a list together with corresponding call information, and the call voice file of the operator is displayed. The position information in the selected call voice file may be displayed in a list in response to an input for selecting.

  In any case, the position information in the compressed call voice file 31 of a plurality of voice keywords displayed in a list includes, for example, the elapsed time from the start of the call, preferably the elapsed time in seconds, the input keyword, Can be displayed as a pair. Preferably, the position information in the compressed call voice file 31 of a plurality of voice keywords displayed in a list is displayed in descending order from the highest similarity score calculated in the voice keyword matching process. If displayed in this way, it is possible to easily select a text with a high likelihood of matching with the input search keyword text.

  When the operator selects and inputs one or a plurality of pieces of position information to be reproduced from the position information in the compressed call voice file 31 of the plurality of voice keywords displayed in the list, the voice keyword search PC terminal 9b is used as necessary. A voice reproduction request is generated from the designated input position of the compressed call voice file 31 held by the call recording server 3 via the voice keyword matching server 5 and transmitted to the call recording server 3 (step S8). Preferably, a field for selecting and inputting the position information in the compressed call voice file 31 of a plurality of voice keywords displayed in a list and indicating how many seconds before the selected position to play back is displayed on the screen. It may be provided above.

  The call recording server 3 that has received this voice reproduction request, from the corresponding position of the compressed call voice file 31 according to the target compressed call voice file identifier described in the voice reproduction request and the position information of the selected voice keyword, The file is read out so as to be reproducible and transmitted as voice data to the voice keyword search PC terminal 9b (step S9).

  The voice keyword search PC terminal 9b that has received the voice data plays back the received voice data. After the desired voice reproduction is completed, a logoff request is transmitted from the voice keyword search PC terminal 9b to the voice keyword matching server 5, and the communication session established in step S1 is disconnected (step S10).

<Functional Configuration of Speech Keyword Matching Server 5 According to First Embodiment and Speech Keyword Matching Processing Details>
FIG. 3 shows a non-limiting example of a functional configuration in the speech keyword matching server 5 according to the first embodiment shown in FIG.

  In FIG. 3, the speech keyword collation server 5 includes a search keyword input unit 51, a speech waveform synthesis unit 52, a speech waveform storage unit 53, a similarity threshold control unit 54, a speech collation unit 55, and a speech collation result output. Part 56.

  FIG. 4 shows, as a non-limiting example, details of the voice keyword matching process according to the first embodiment, which is executed by each component in the voice keyword matching server 5 shown in FIG.

  Referring to FIGS. 3 and 4, the search keyword input unit 51 has an interface function to the search keyword search PC terminal 9b, processes log-on and logoff request authentication from the search keyword search PC terminal 9b, and performs a search. The search keyword text and search conditions input from the keyword search PC terminal 9b are received from the speech keyword search PC terminal 9b and supplied to the speech waveform synthesis unit 52 and the similarity threshold control unit 54 (step S41).

  The speech waveform synthesis unit 52 synthesizes a speech waveform pattern having a normal speech rate and a normal speech volume as a standard speech waveform pattern from the supplied search keyword text.

  FIG. 6 shows, as a non-limiting example, a standard speech waveform pattern that is synthesized when “Kabushiki Gaishai Boi” is input as, for example, a search keyword text. In FIG. 6, a frequency and a waveform are shown with the vertical axis representing volume (amplitude) and the horizontal axis representing time.

  3 and 4, the speech waveform synthesizer 52 further changes the frequency of the standard speech waveform pattern from the standard speech waveform pattern, and the speech waveform pattern whose speech rate is higher than that of the standard speech waveform pattern (see FIG. 3). 7A), a speech waveform pattern (FIG. 7B) having a low speech rate, a speech waveform pattern (FIG. 8A) having a speech volume smaller than the standard waveform pattern by changing the amplitude of the standard speech waveform pattern, and a speech volume The voice waveform pattern (FIG. 8B) or the like having a high volume is appropriately synthesized, and the synthesized voice waveform patterns are stored in the temporary storage area of the voice waveform storage unit 53 (step S42). Since the plurality of speech waveform pattern synthesis processes are digital waveform conversion processes from the standard speech waveform pattern, they can be easily combined as appropriate. For example, both the speech rate and the speech volume are adjusted, and the speech waveform pattern with a speech rate that is faster and the speech volume is lower than the standard speech waveform pattern, or the speech waveform pattern that has a lower speech rate and a higher speech volume May be synthesized.

  As a modified example, a speech waveform pattern synthesized from a standard speech waveform pattern with different utterance speed and speech volume is increased or decreased stepwise by increasing the utterance speed and speech volume of the standard speech waveform pattern, respectively. A plurality of voice waveform patterns may be synthesized. Alternatively, a speech waveform pattern in which only one of the speech speed and the speech volume is different from the standard speech waveform pattern may be synthesized.

  Preferably, the speech waveform synthesizer 52 further selects, for example, a male and a younger group, a male and a middle-aged and higher class from the standard speech waveform pattern according to characteristics of the voice quality of the speaker, for example, gender and age group. Speech waveform patterns such as male and elderly, female and young, female and middle-aged, female and elderly may be synthesized. The customer information in the customer database 33 that can be referred to by the voice keyword matching server 5 is a customer identifier obtained by using the caller name, the caller telephone number or the receiver telephone number as a key as a search condition inputted in the voice keyword search terminal. If it is known by referring to, the speech waveform pattern having the characteristics of the sex and the age group specified according to the customer information may be used only or preferentially in the speech keyword matching process.

  Preferably, when synthesizing a speech waveform pattern from an input search keyword, only speech speech is obtained after eliminating noises other than speech, for example, hold sound, disconnection sound, push button sound (PB sound), etc. A speech waveform pattern may be synthesized.

  The similarity threshold control unit 54 compares the speech that has been compared with the similarity calculated when the plurality of synthesized speech waveform patterns and the speech keywords in the search target call speech file narrowed down by the search condition are collated. A similarity threshold for determining whether or not to extract a keyword is variably set (step S43).

  Preferably, the threshold value of the similarity is decreased when the number of characters of the input search keyword is large, for example, and is increased when the number of characters of the search keyword is small. It may be set as follows. The number of characters of the input search keyword can be obtained from the search keyword input unit 51. If the input keyword is a large number of characters, such as text, for example, if the similarity threshold is fixed, the calculated similarity value decreases as the number of characters in the search keyword increases, and the speech that should be extracted originally Although it is difficult to extract a keyword, the accuracy of voice keyword matching is not impaired even when a search keyword with a relatively large number of characters such as a sentence is input by setting the threshold value of the similarity in this manner. .

  More preferably, the threshold value of the similarity is reduced when the compression rate of the compressed call audio file obtained as a search target is high, for example, while the threshold value of the predetermined similarity as a reference is reduced. May be set to increase when the compression rate is low, and may be variably adjusted based on the difference in sound quality depending on the compression method. Various well-known methods and various compression ratios can be used for call voice compression. As a non-limiting example, monaural 1/5 compression, monaural 1/10 compression, or stereo no compression is used. It's okay. The compressed call voice file obtained as a search target can be obtained by referring to the compression control information stored in the compressed call voice file 31 corresponding to the voice data.

  In general, since the capacity of a recording call at a call center is enormous, a high compression rate is realized by an irreversible compression method that performs irreversible compression by converting and decimating data from uncompressed voice data. And, as the compression rate of the compressed call audio file obtained as a search target is higher, there is a data loss from the uncompressed audio data, so the similarity value calculated as the similarity threshold is fixed becomes smaller, It becomes difficult to extract the voice keywords that should be extracted originally. For this reason, in the first embodiment, by setting the similarity threshold variably according to the difference in sound quality such as the compression rate, the accuracy of voice keyword matching is improved even for a call voice file with a high compression rate. It will not be damaged.

  The voice keyword matching unit 55 refers to the compressed call voice file 31 and the call information database 32, reads out the compressed call voice file that matches the search condition, and call information corresponding to the compressed call voice file as necessary, The voice keyword collating unit 55 writes it in a temporary storage area that can be referred to, and collates the plurality of voice waveform patterns supplied from the voice waveform synthesizing unit 52 with the voice keywords in the compressed call voice file obtained as a search target, The similarity is calculated (step S44).

  When either the customer's utterance or the operator's utterance is selected as a search condition, the speech keyword matching unit 55 refers to the call information database 32 and searches only the selected utterer portion as a search target. The voice keyword matching process is executed.

  FIG. 5 shows an example of processing executed by the voice keyword matching unit 55 for matching a plurality of voice waveform patterns with voice keywords in a compressed call voice file obtained as a search target.

  Referring to FIG. 5, the voice keyword matching unit 55 first selects a voice waveform pattern having a high utterance speed and / or a low utterance volume, and uses this voice waveform pattern to compress the compressed call voice file 31. A collation process with the voice keyword in the middle is executed (step S51).

  If the similarity calculated in the collation process in step S51 is equal to or greater than the similarity threshold supplied from the similarity threshold control unit 54 (step S52Y), the position information of the collated voice keyword is output ( Step S58).

  On the other hand, if the similarity is less than the similarity threshold supplied from the similarity threshold control unit 54 (step S52N), the speech waveform pattern whose utterance speed is normal speed and / or utterance volume is normal volume is selected next. Then, collation processing with the voice keyword in the compressed call voice file 31 is executed using this voice waveform pattern (step S53).

  If the similarity calculated in the collation process in step S53 is equal to or greater than the similarity threshold supplied from the similarity threshold control unit 54 (step S54Y), the position information of the collated voice keyword is output ( Step S58).

  On the other hand, if it is less than the similarity threshold supplied from the similarity threshold control unit 54 (step S54N), the speech waveform pattern with the next lowest utterance speed and / or higher utterance volume is selected. The voice waveform pattern is used to perform collation processing with the voice keyword in the compressed call voice file 31 (step S55).

  If the similarity calculated in the collation process in step S55 is equal to or greater than the similarity threshold supplied from the similarity threshold control unit 54 (step S56Y), position information of the collated voice keyword is output ( On the other hand, if it is less than the similarity threshold supplied from the similarity threshold controller 54 (step S56N), data indicating that there is no voice keyword to be extracted is output (step S57).

  If a plurality of voice waveform patterns are selected in the order shown in FIG. 5, the voice waveform pattern selected first has a smaller data capacity, and the voice waveform pattern selected later increases the data capacity. For this reason, a collation result can be obtained more efficiently and at high speed.

  FIG. 9 schematically shows a speech keyword matching technique performed by the speech keyword matching unit 55.

  Referring to FIG. 9, specifically, the speech keyword collating unit 55 uses each of a plurality of synthesized speech waveform patterns or a part divided by phoneme as a speech analysis frame, and starts a call start time. Or when the search start point is input as a search condition, the collation point is shifted from the input search start point to the call end point, and the degree of similarity is calculated for each shift.

  When a plurality of speech waveform patterns are supplied, the speech keyword matching unit 55 may generate a digital speech analysis frame for each of the plurality of speech waveform patterns and perform a shift process using these in turn. If application priority is given, it is preferable to perform shift processing in order from the highest priority.

  The voice keyword matching unit 55 further compares the calculated similarity with the similarity threshold supplied from the similarity threshold control unit 54, and the case where the calculated similarity is equal to or greater than the similarity threshold. The start and / or end position (typically the position grasped by the elapsed time from the call start time) in the compressed call voice file of the voice keyword is acquired and supplied to the voice collation result output unit 56 (step S55). ).

  Returning to FIG. 3, the voice collation result output unit 56 uses the voice keyword position information collated in the compressed call voice file to be searched, which is supplied from the voice collation unit 55, as a search result for the voice keyword search PC terminal 9b. To send.

<Details of similarity calculation in voice keyword matching process>
Various methods can be used for the specific calculation logic of the similarity in the collation. For example, the speech keyword matching unit 55 positions the digital speech analysis frame in the compressed speech file obtained as a search target for the digital speech analysis frames of a plurality of speech waveform patterns supplied from the speech waveform synthesis unit 52. The degree of similarity may be calculated by comparing the digital waveform at the location with digital waveform pattern matching (template matching). In the speech keyword matching process according to the first embodiment, since the ambiguity analysis is allowed, especially when the number of input keyword characters is long, the similarity threshold is set so that the matching with the search keyword is detected if some words are matched. May be set. If matching between the digital waveforms is performed, a collation result can be obtained with practical accuracy even with simple and low-load processing logic.

  As a modification, other collation methods described below can be used instead of or in combination with the pattern matching between the digital waveforms. By using these matching methods, the ambiguity in speech keyword matching can be improved, the influence of noise other than speech is reduced, and the speech quality and sensitivity of speech quality are moderately reduced. Keyword candidates can be extracted.

  As shown in FIG. 10, for example, by plotting the strength of sound pressure (output power) (dB) on the time axis of a digital time waveform signal of a speech waveform pattern, a speech power spectrum envelope (Power Spectrum Envelope: The degree of similarity may be calculated by obtaining a PSE) pattern and comparing the power spectrum envelope (PSE) patterns. According to this, since it is relatively difficult to be affected by the sound quality, voice quality, noise and the like of the sound, a more relative comparison is possible.

  Alternatively, as shown in FIG. 11, for example, an F0 frequency value (Hz) that is a voice fundamental frequency indicating a voice pitch (sound height) on the time axis is extracted from a digital time waveform signal of a voice waveform pattern. Thus, a pitch (F0) curve may be obtained by plotting, and the degree of similarity may be calculated by comparing the pitch (F0) curves. According to this, a silent section can be extracted with higher sensitivity, and since it is relatively less affected by the sound quality, voice quality, noise, and the like, more relative comparison is possible.

  Alternatively, as shown in FIG. 12, for example, a digital time waveform signal of a speech waveform pattern is extracted by plotting a formant that is a frequency spectrum unique to speech on the time axis, thereby generating a phoneme (vowel). Alternatively, the similarity may be calculated by obtaining an occurrence pattern of consonants) and comparing the phoneme generation patterns. Since this formant frequency differs depending on the shape of the vocal tract of the speaker, higher collation accuracy can be obtained in consideration of individual differences among individual speakers.

  Further alternatively, as shown in FIG. 13, for example, a speech spectrogram is calculated on the time axis of a digital time waveform signal of a speech waveform pattern. The similarity may be calculated by comparing the audio spectrograms. Since this voice spectrogram shows the specific pattern of the utterance location according to the voiceprint analysis, it is possible to perform voiceprint matching in consideration of individual differences among speakers, and higher matching accuracy can be obtained.

  With reference to FIGS. 14 to 18, supplementary description will be given of various conversion processes in the processes shown in FIGS. 10 to 13, particularly conversion processes taking into account individual differences among speakers. Speech is labeled on a phoneme basis. That is, the phoneme unit is “a”, “i”, “u”, “e”, “o” for vowels, and “k”, “s”, “t”, “o” for consonants. Any one of “n”, “h”, “m”, “j”, “r”, “w”, “g”, “z”, “d”, “b”, “p” is given as a label. The FIG. 14 is a graph in which the magnitude of air vibration is plotted on the vertical axis and time is plotted on the horizontal axis when pronounced “electronic”, and different waveform patterns appear for each phoneme.

  Next, the labeled voice is normalized (unified) according to the pitch, loudness, and speed of the sound as necessary. Since the sound pitch (speech pitch) varies depending on the individual, as shown in Fig. 15, the original signal is thinned and further shortened on the time axis to change the pitch and normalize the pitch. To do. Also, since the volume and speed of the voice differ for each utterance, the voice waveform is expanded and contracted to normalize the volume and speed of the voice so as to match the speed of the reference voice as shown in FIG. These normalization processes are executed on a phoneme basis.

  Next, voiceprint information is extracted by generating a sound spectrogram. The spectrum of speech frequency characterizes the voiceprint of the speaker. This frequency spectrum can be obtained by subjecting a time signal to Fourier transform (Fourier Transform). For example, fast Fourier transform (FFT) processing suitable for processing by a processor is applied to a time waveform signal of speech. Can be realized. FIG. 17 shows how much frequency a certain sound is obtained by applying FFT processing to an audio signal, the frequency (Hz) on the horizontal axis and the loudness (dB) on the vertical axis. This is a speech spectrum shown in the graph, that is, a speech spectogram.

  FIG. 18 shows an envelope on the speech spectrogram, and a plurality of peaks appear in the envelope. Each of the peaks on the voice spectrum is called a formant, and a change on the time axis of the formant indicates a characteristic on a voiceprint due to individual differences.

<Variation in Speech Keyword Extraction Determination Processing in Speech Keyword Matching Processing>
In the speech keyword extraction determination in the collation, various methods can be used in the above-described method for extracting the speech keyword in which the similarity that is equal to or higher than the similarity threshold is calculated. For example, in the comparison between the calculated similarity and the threshold of similarity, according to the difference between the two, a plurality of likelihood stages are provided for the degree of matching, and these are sent to the speech keyword search PC terminal 9b as matching results. May be.

  In addition, even when a phoneme group constituting a speech waveform pattern synthesized from an input keyword includes a phoneme that matches or is similar to a phoneme that matches or is not similar, a speech that partially matches or is similar As a non-limiting example, in order to be able to be extracted as a keyword, the likelihood that a speech keyword whose similarity is calculated as 60 to 100 on the scale of similarity as 0 to 100 matches the input search keyword Is extracted as the highest speech keyword, and the speech keyword whose similarity is calculated as 45 to 60 (less than) is the speech with the second highest likelihood of partially matching at least half of the input search keyword. Search with at least less than half of voice keywords extracted as keywords and calculated with a similarity of 30 to 45 (less than) A speech keyword that is extracted as a speech keyword having the third highest likelihood of partially matching a word, and whose similarity is calculated as 0 to 30 (less than) is partially or partially compared with the input search keyword. A mismatched speech keyword with the lowest likelihood of matching may be determined and excluded from extraction targets.

<Hardware Configuration of Spoken Keyword Matching System According to First Embodiment>
FIG. 19 is a block diagram illustrating an example of a hardware configuration of each server device according to the first embodiment. In each server device that is the computer device 110 shown in FIG. 19, the CPU 111 controls the entire system by using the RAM 115 as a work memory for primary storage in accordance with programs stored in the ROM 114 and / or the hard disk drive 116. Furthermore, the CPU 111 executes the speech keyword collation process according to the first embodiment based on a program stored in the hard disk drive 116 in accordance with a user instruction input via the mouse 112a or the keyboard 112. A display such as a CRT or LCD is connected to the display interface 113, and an input standby screen for voice keyword matching processing executed by the CPU 111, processing progress, processing results, search results, and the like are displayed. The removable media drive 117 is mainly used when writing a file from the removable medium to the hard disk drive 116 or writing a file read from the hard disk drive 116 to the removable medium. Removable media include floppy disk (FD), CD-ROM, CD-R, CD-R / W, DVD-ROM, DVD-R, DVD-R / W, DVD-RAM and MO, memory card, CF Cards, smart media, SD cards, memory sticks, etc. can be used.

  A printer such as a laser beam printer or an ink jet printer is connected to the printer interface 118. The network interface 119 is an interface for connecting a computer device to a network.

  The input means for each server device and search keyword input PC terminal 9b according to the first embodiment is not limited to the mouse 112a or the keyboard 112, and any pointing device such as a trackball, a trackpad, a tablet, or the like. Can be used as appropriate. When the portable information terminal is used as an input / output device connected to the server device and the search keyword input PC terminal 9b according to the first embodiment, the input unit may be configured with a button, a mode dial, or the like.

  Further, the hardware configuration of each server according to the first embodiment shown in FIG. 19 is merely an example, and it is needless to say that any other hardware configuration can be used.

  In particular, all or part of the voice keyword extraction processing according to the first embodiment may be realized by the above-described computer terminal device 110 or a portable information terminal device such as a PDA, and the like. Internet or any well-known local network that is interconnected by wireless communication such as Bluetooth (registered trademark) or wired communication lines such as the Internet (TCP / IP), public telephone network (PSTN), integrated service digital network (ISDN) A part or all of the call recording process and the voice keyword matching process may be realized by a network system including an area network (LAN) or a wide area network (WAN).

  As described above, according to the first embodiment, versatility without requiring a voice recognition process, which is suitable for a CRM system that records and stores calls made between a customer's phone and a customer's phone. The speech keyword matching in the speech data is realized with a high and simple configuration and practical search accuracy.

  Further, even when various customer-side utterers and utterance situations are assumed, it is possible to perform speech keyword collation for speech data with a simple configuration and practical search accuracy.

  Further, even when a sentence longer than this is used as a search keyword instead of a short phrase, it is possible to perform speech keyword collation for speech data with a simple configuration and practical search accuracy.

  Furthermore, even when the compressed voice data is targeted, it is possible to perform voice keyword matching on the voice data without impairing the search accuracy.

Second Embodiment With reference to FIGS. 20 to 22, the second embodiment of the spoken keyword matching system according to the present invention will be described only with respect to differences from the first embodiment.

  In the second embodiment, compared with the first embodiment, the speaker gender, call volume, and speech of the call voice file stored and stored in the compressed call voice file 31 from a plurality of synthesized voice waveform patterns are compared. Attributes such as speed and line type are estimated, and a speech waveform pattern having the estimated attributes is preferentially used for collation.

  Furthermore, in the second embodiment, in the voice keyword matching, the matching process is executed only in the vicinity region of the voice keyword position extracted as a candidate in the first selected voice waveform pattern.

  FIG. 20 shows a non-limiting example of a functional configuration in the speech keyword matching server 5 according to the second embodiment.

  In FIG. 20, the speech keyword matching server 5 according to the second embodiment further includes a speech waveform pattern estimation unit 57 as compared with the speech keyword matching server according to the first embodiment shown in FIG. In addition, a voice keyword matching unit 57 according to the second embodiment is provided instead of the voice keyword matching unit 55 according to the first embodiment.

  The voice waveform pattern estimation unit 57 refers to the compressed call voice file 31 and the call information database 32, and attributes that can be a determinant of the voice waveform pattern shape of the call voice file stored and stored in the compressed call voice file 31, for example, speech The person sex, conversation volume, speech speed, and line type are estimated, and the estimated speaker sex, conversation volume, speech speed, line type, and the like are supplied to the voice keyword fire fighting group 551. Preferably, the speech waveform pattern estimation unit 57 reads the call speech file stored and stored in advance in the compressed call speech file 31 and supplies the estimation result to the speech keyword collation unit 551 prior to inputting the search keyword text. deep.

  FIG. 21 is a flowchart showing a non-limiting example of the speech waveform pattern estimation processing procedure executed by speech waveform pattern estimation unit 57.

  In FIG. 21, the voice waveform pattern estimation unit 57 first reads out the compressed call voice file stored in the compressed call voice file 31 and performs an acoustic analysis of each call voice file to calculate the pitch frequency. Is estimated (step S211).

  Specifically, the pitch frequency (Hz) is grasped as the vibration frequency of the speaker's vocal cords, and the male population and the female population show significantly different distributions due to different vocal cord shapes and sizes. The normal conversation pitch frequency is lower limit 60 Hz, upper limit 260 Hz, average value 110 Hz to 150 Hz, while female normal conversation pitch frequency is lower limit 130 Hz, upper limit 520 Hz, average value 180 Hz to 220 Hz. For this reason, the compressed voice call file is read and the average value of the pitch frequencies is calculated. If the calculated average value of the pitch frequencies is less than 165 Hz, for example, the speaker is estimated to be male, and the other is 165 Hz or higher. If so, it can be estimated that the speaker is a woman.

  Next, the voice waveform pattern estimation unit 57 estimates the call volume and the speech speed by performing acoustic analysis of each call voice file read from the compressed call voice file 31 (step S212).

  Specifically, whether the volume is normal or loud by calculating the average volume (power) of the voiced part in each call voice file and comparing it with a predetermined sound pressure reference value. Alternatively, it can be estimated whether the volume is low.

  Further, by calculating the time length (Sec) of the utterance portion and the number of voiced peaks of each call voice file, and calculating the number of voiced peaks / time length of the utterance portion, the number of phonemes uttered per second can be obtained. Therefore, by comparing this with the reference number of phonemes per second, it is possible to estimate whether the speech speed is normal, high, or low.

  Next, the voice waveform pattern estimation unit 57 refers to the call information database 32, identifies the telephone number of the customer side telephone terminal corresponding to each read compressed call voice file, and the line used is a fixed telephone. The line type, such as whether it is for mobile phone, mobile phone or PHS, is estimated (step S213).

  After step S213 ends, the speech waveform pattern estimation unit 57 supplies the estimated speaker sex, call volume, speech speed, and line type to the speech keyword collation unit 551.

  FIG. 22 is a flowchart showing a non-limiting example of a voice keyword matching processing procedure executed by the voice keyword matching unit 551 according to the second embodiment.

  In FIG. 22, the speech keyword collating unit 551 generates the speech waveform pattern of the middle-aged and elderly layers, which is the estimated speaker gender, call volume, speech speed, and line type supplied from the speech waveform estimation unit 57, and preferably It selects as a 1st audio | voice waveform pattern (step S221). Alternatively, if the age group to which the speaker population belongs most is known in advance, it is preferable to select a speech waveform pattern belonging to this age group.

  The voice keyword matching unit 551 uses the selected first voice waveform pattern, and the voice keyword matching unit 551 refers to the compressed call voice file 31 and selects a compressed call voice file that matches the input search condition. The read voice data is compared with the first voice waveform pattern in the range from the beginning to the end of the read compressed call voice file (step S222).

  If there is a voice keyword whose similarity equal to or greater than the first similarity threshold is present in the compressed call voice file to be searched by this collation, the process proceeds to step S224, and on the other hand, in the compressed call voice file to be searched. If there is no voice keyword whose similarity equal to or greater than the first similarity threshold is present, it is determined whether there is another voice waveform pattern to be used for matching the compressed call voice file. (Step S231Y), another speech waveform pattern is selected and the process returns to Step S222 (Step S232). If there is no other (Step S231N), the process is terminated.

  Returning to step S224, the speech keyword matching unit 551 converts only the age group of the first speech waveform pattern so as to be compatible with other age groups, and selects it as the second speech waveform pattern (step S224).

  Next, the starting point position and the ending position on the compressed call voice file of the voice keyword extracted as a candidate in step S223 by collation in step S222 are specified, and a predetermined time after the specified starting point position and after the ending position. The width is added and specified as a neighborhood area of the candidate speech keyword (step S225). As a non-limiting example, if the length of a voice keyword extracted as a candidate is 0.1 seconds, the voice is 1.1 seconds including 0.5 seconds ahead of the start position and 0.5 seconds from the end position. The waveform area is specified as a verification target area (step S225).

  In the second embodiment, by using the selected second (and third and subsequent) voice waveform patterns, the entire compressed call voice file is used as a comparison target. The speech keyword collation unit 551 collates speech keywords by using only the identified region to be collated as a collation target (step S226).

  If there is a voice keyword whose similarity is greater than or equal to the second similarity threshold greater than the first similarity threshold in the compressed call voice file to be searched by this collation, the second similarity threshold is set. The voice keyword for which the above similarity is calculated is extracted as a voice keyword to be obtained as a search result (step S228), and the process proceeds to step S229. On the other hand, the compressed call voice file to be searched is larger than the first similarity threshold. If there is no voice keyword for which a similarity greater than or equal to the second similarity threshold is calculated, the process proceeds to step S229.

  In step S229, when there is a speech waveform pattern of another age group to be applied as a speech waveform pattern (step S229Y), the process returns to step S224, while another age group to be further applied as a speech waveform pattern. If the voice waveform pattern does not exist (step S229N), the voice keyword that is equal to or higher than the second similarity threshold and the highest similarity is calculated, its position information, and the applied voice waveform Pattern attributes (speaker sex, call volume, speech speed, line type, age group, etc.) and the calculated similarity are extracted (step S230).

  The voice keyword extracted in step S230 and its position information, the attribute of the applied voice waveform pattern (speaker gender, call volume, speech speed, line type, age group, etc.) and the calculated similarity are the voice keyword It can be appropriately displayed as a search result on the display device of the search PC terminal 9b.

  According to the second embodiment, the attribute of the voice waveform pattern to be preferentially applied is estimated in advance, and then the estimated voice waveform pattern is first applied to the entire search target compressed call voice file to obtain the first similarity. A speech keyword for which a similarity equal to or greater than the threshold is calculated is extracted as a candidate, and a speech waveform pattern close to the estimated speech waveform pattern is applied to only the neighborhood region of the speech keyword extracted as the candidate. Then, collation is executed, and the speech keyword and its position information for which the second similarity threshold value greater than the first similarity threshold value is calculated are extracted as search results. For this reason, compared with the first embodiment, more efficient voice keyword matching processing is realized, voice keyword search results can be obtained more quickly, and the CPU load for executing matching is reduced.

  The scope of the present invention is not limited to the illustrated and described exemplary embodiments, and includes all embodiments that provide the same effects as those intended by the present invention, and does not depart from the spirit of the present invention. Various improvements or changes can be made within the scope. For example, the telephone number analysis processing, voiceprint analysis processing, and emotion analysis processing disclosed in the present embodiment may be implemented alone in the harmful customer detection system according to the present embodiment, or may be implemented in any combination. Also good.

  For example, in the present embodiment, the threshold value of the similarity is in addition to, or in place of, the number of input keyword characters and the compression rate of call voice data, the type of the telephone terminal in which the call is recorded (for example, a fixed phone, a mobile phone, an IP It may be changed according to telephone, PHS, etc. In general, the sound quality of a call made on a mobile phone is inferior to that of a call made on a landline phone. For example, in the case of a call on a mobile phone, the similarity threshold is lowered (increase ambiguity). Thus, search omissions can be reduced even for calls on mobile phones.

  Further, the scope of the present invention is not limited to the combination of features of the invention defined by claim 1 but can be defined by any desired combination of specific features among all the disclosed features. .

PBX 1
Voice acquisition server 2
Call recording server 3
Control server 4
Voice keyword matching server 5
Customer phone terminal 7
PSTN 8
Operator telephone terminal 9a
Voice keyword search PC terminal 9b
Private lines 11a, 11b, 11c
Compressed call audio file 31
Call information database 32
Customer information database 33

Claims (18)

An audio database for storing audio data in a reproducible manner;
A search keyword input unit for inputting a search keyword text and a search condition for collating voice keywords to be reproduced in the voice database;
A speech waveform synthesizer that synthesizes a standard speech waveform pattern from the input search keyword text and synthesizes a plurality of derived speech waveform patterns having different speech speeds and / or volumes from the standard speech waveform pattern;
A similarity threshold calculation unit that variably sets a threshold of similarity in matching between the search keyword text and the voice keyword by increasing or decreasing a reference threshold;
With reference to the speech keyword database, the speech data matching the search condition is read out, and the standard speech waveform pattern and the plurality of derived speech waveform patterns are sequentially compared with the speech waveform pattern of the speech data read out. A speech keyword matching unit that calculates the similarity and obtains the position of the speech keyword for which the similarity equal to or greater than the similarity threshold is calculated;
And an output unit that transmits the position of the obtained voice keyword to the terminal device, thereby enabling reproduction of the voice data from the position of the voice keyword obtained on the terminal device. Keyword matching server device. The voice keyword matching unit, when either inbound utterance or outbound utterance is selected and input as the search condition, refers to call control information corresponding to the voice data and is selectively input The speech keyword matching server device according to claim 1, wherein a speech waveform pattern of only one of the utterance and the outbound utterance is limited as a matching target. The speech keyword matching unit performs a first comparison process between a derived speech waveform pattern having a higher speech speed and / or lower volume than the standard speech waveform pattern and the speech waveform pattern of the read speech data, If no voice keyword candidate is obtained in the first comparison, a second comparison process is performed between the standard waveform pattern and the waveform pattern of the read voice data, and the second comparison If no voice keyword candidate is obtained in the comparison process, the derived speech waveform pattern having a lower speech speed and / or louder volume than the standard speech waveform pattern and the speech waveform pattern of the speech data read out. The voice keyword matching server apparatus according to claim 1 or 2, wherein a third comparison process is performed. The similarity threshold calculation unit decreases the reference threshold when the number of characters in the search keyword text is large, and increases the threshold from the reference threshold when the number of characters in the search keyword text is small. The speech keyword matching server device according to claim 1, wherein the speech keyword matching server device is set. The similarity threshold calculation unit is configured to decrease from the reference threshold when the compression rate of the audio file in the audio database is high, and to increase from the reference threshold when the compression rate of the audio file is low. The speech keyword matching server device according to claim 1, wherein a threshold value for similarity is set. The speech waveform synthesis unit synthesizes a plurality of derived speech waveform patterns having different voice qualities characterized by sex and / or age group from the input search keyword text. The voice keyword matching server device described. A speech keyword matching method executed by a speech keyword matching server device comprising a speech database, a search keyword input unit, a speech waveform synthesis unit, a similarity threshold calculation unit, a speech keyword collation unit, and an output unit. ,
A step of inputting search keyword text and search conditions for collating voice keywords to be reproduced in a voice database storing voice data so as to be reproducible by the search keyword input unit;
Synthesizing a standard speech waveform pattern from the input search keyword text by the speech waveform synthesis unit, and synthesizing a plurality of derived speech waveform patterns whose speech speed and / or volume are different from the standard speech waveform pattern; ,
A step of variably setting a similarity threshold in matching between the search keyword text and the voice keyword by increasing or decreasing a reference threshold by the similarity calculation unit;
The voice keyword collating unit refers to the voice keyword database, reads voice data that matches the search condition, and reads the standard voice waveform pattern and the plurality of derived voice waveform patterns of the voice data that has been read. Calculating the similarity by sequentially comparing with a speech waveform pattern, and obtaining a position of a speech keyword from which a similarity equal to or greater than a threshold of the similarity is calculated;
Transmitting the position of the obtained voice keyword to the terminal device by the output unit, thereby enabling reproduction of voice data from the position of the voice keyword obtained on the terminal device. Voice keyword matching method. In the step of obtaining the position of the voice keyword, when either the inbound utterance or the outbound utterance is selected and input as the search condition, the selection input is performed by referring to the call control information corresponding to the voice data. The speech keyword matching method according to claim 7, wherein a speech waveform pattern of only one of inbound utterances and outbound utterances is limited as a subject to be collated. In the step of obtaining the position of the speech keyword, a first comparison process is performed between a derived speech waveform pattern having a higher speech speed and / or lower volume than the standard speech waveform pattern and a speech waveform pattern of the read speech data. When a speech keyword candidate is not obtained in the first comparison, a second comparison process is performed between the standard waveform pattern and the waveform pattern of the read voice data. When a speech keyword candidate is not obtained in the second comparison process, a derived speech waveform pattern having a lower speech speed and / or louder volume than the standard speech waveform pattern, and a speech waveform pattern of the speech data read out The voice keyword matching method according to claim 7, wherein a third comparison process is performed between the voice keywords. In the step of setting the similarity threshold, when the number of characters of the search keyword text is large, the similarity is decreased from the reference threshold, and when the number of characters of the search keyword text is small, the similarity is increased. The speech keyword matching method according to claim 7, wherein a threshold value is set. In the step of setting the similarity threshold, when the compression rate of the audio file in the audio database is high, the threshold is decreased from the reference threshold, and when the compression rate of the audio file is low, it is increased from the reference threshold. The speech keyword matching method according to claim 7, wherein a threshold value for the similarity is set. 12. The step of synthesizing the speech waveform includes synthesizing a plurality of derived speech waveform patterns having different voice qualities characterized by sex and / or age group from the input search keyword text. One of the voice keyword matching methods described. A spoken keyword matching program for causing a computer to execute a spoken keyword matching process, the program comprising:
Voice data storage processing for storing voice data in a voice database so as to be reproducible
A search keyword input process for inputting a search keyword text and a search condition for collating voice keywords to be reproduced in the voice database;
A speech waveform synthesis process for synthesizing a standard speech waveform pattern from the input search keyword text and synthesizing a plurality of derived speech waveform patterns having different speech speeds and / or volumes from the standard speech waveform pattern;
A similarity threshold calculation process for variably setting a threshold of similarity in matching between the search keyword text and the voice keyword by increasing or decreasing a reference threshold;
With reference to the speech keyword database, speech data matching the search condition is read out, and the standard speech waveform pattern and the plurality of derived speech waveform patterns are sequentially compared with the speech waveform pattern of the speech data that has been read out. Voice keyword matching processing for calculating the similarity and obtaining the position of the voice keyword for which the similarity equal to or greater than the similarity threshold is calculated;
Transmitting the position of the obtained voice keyword to the terminal device, thereby executing processing including output processing that enables reproduction of voice data from the position of the voice keyword obtained on the terminal device A speech keyword matching program characterized by In the voice keyword matching process, when either an inbound utterance or an outbound utterance is selected and input as the search condition, the inbound utterance selected and input with reference to call control information corresponding to the voice data is selected. The speech keyword collation program according to claim 13, wherein the speech waveform pattern of only one of the utterance and the outbound utterance is limited as a collation target. The voice keyword matching process performs a first comparison process between a derived voice waveform pattern having a higher speech speed and / or lower volume than the standard voice waveform pattern and a voice waveform pattern of the read voice data, If no voice keyword candidate is obtained in the first comparison, a second comparison process is performed between the standard waveform pattern and the waveform pattern of the read voice data, and the second comparison If no voice keyword candidate is obtained in the comparison process, the derived speech waveform pattern having a lower speech speed and / or louder volume than the standard speech waveform pattern and the speech waveform pattern of the speech data read out. The speech keyword matching program according to claim 13 or 14, wherein a third comparison process is performed. The similarity threshold calculation process is configured to decrease the reference threshold when the number of characters of the search keyword text is large, and to increase from the reference threshold when the number of characters of the search keyword text is small. The speech keyword matching program according to claim 13, wherein: The similarity threshold calculation process is configured to decrease from the reference threshold when the compression rate of the audio file in the audio database is high, and to increase from the reference threshold when the compression rate of the audio file is low. The speech keyword matching program according to any one of claims 13 to 16, wherein a threshold value of similarity is set. 18. The voice waveform synthesis process synthesizes a plurality of derived voice waveform patterns having different voice qualities characterized by sex and / or age group from the input search keyword text. The spoken keyword matching program described.