JP2011257529A - Method, device and program for holding-related utterance extraction - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technology capable of extracting an utterance related to holding more appropriately.SOLUTION: A voice feature amount calculation unit 2 extracts a voice feature amount of a voice signal. A voice recognition unit 3 uses the voice feature amount, an acoustic model and a language model to execute voice recognition for the voice signal, to detect an utterance included in the voice signal and to generate information on the detected utterance. A holding interval detection unit 4 uses the information on the utterance to detect a holding interval between adjacent utterances which is equal to or longer than a predetermined time. An extraction unit 5 extracts more utterances from a group of utterances adjacent to the holding interval when the holding interval is longer.

Description

  The present invention relates to a technique for extracting utterances related to holding.

  There is a method in which a phone call is converted into text by a voice recognition technology and an important word in the phone call is extracted by a text processing technology (for example, see Non-Patent Document 1).

  Conventionally, an important word is extracted by applying the method described in Non-Patent Document 1 to the entire call on hold, and the cause of the hold is searched.

Takenobu Tokunaga (Author), Junichi Sakurai (Editor), "Language and Calculation (5) Information Retrieval and Language Processing", The University of Tokyo Press, November 1999

  However, there is a problem that the important word extracted from the entire call on hold is a word related to a topic that has nothing to do with hold.

  An object of the present invention is to more appropriately extract utterances related to holding.

  In order to solve the above-described problem, a voice feature amount of a voice signal is extracted. Voice recognition is performed on the voice signal using a voice feature, an acoustic model, and a language model, an utterance included in the voice signal is detected, and information about the detected utterance is generated. Using the information about the utterance, a section where the interval between adjacent utterances is equal to or longer than a predetermined time is detected as a reserved section. From the set of utterances adjacent to the reserved section, a larger number of utterances are extracted as the reserved section is longer.

  It is possible to more appropriately extract utterances related to holding.

The functional block diagram of the example of a pending | holding related speech extraction apparatus. The flowchart which shows the example of the hold related utterance extraction method. The flowchart which shows the example of step S4. The figure for demonstrating the example of a detection of a pending | holding area. The figure for demonstrating the example of a holding | maintenance related speech extraction.

  An embodiment of the present invention will be described below with reference to the drawings.

  As shown in FIG. 1, the hold-related utterance extraction device includes, for example, a voice signal acquisition unit 1, a voice feature amount calculation unit 2, a voice recognition unit 3, a hold section detection unit 4, and an extraction unit 5. This hold-related utterance extraction device executes each step of the hold-related utterance extraction method illustrated in FIG.

  The voice acquisition unit 1 performs A / D conversion on the input analog voice signal to generate a digital voice signal (step S1). The digital voice signal is sent to the voice feature quantity extraction unit 2. The analog audio signals input to the audio acquisition unit 1 are a plurality of analog audio signals respectively corresponding to a plurality of channels. In this example, it is assumed that the number of channels is 2, one of which is channel A for operator's voice and the other is channel B for customer's voice.

  The voice feature quantity extraction unit 2 extracts the voice feature quantity of the digital voice signal (step S2). Information about the extracted voice feature amount is sent to the voice recognition unit 3. The voice feature amount is, for example, MFCC (Mel-Frequency Cepstrum Coefficient) or ΔMFCC which is the amount of change in MFCC, and may be anything that can be used by the voice recognition unit 3 described later. An existing technique may be used to extract the voice feature amount.

  The speech recognition unit 3 performs speech recognition on the speech signal using the speech feature value, the acoustic model, and the language model, detects an utterance included in the speech signal, and generates information about the detected utterance ( Step S3). Information about the detected utterance is sent to the hold section detection unit 4 and the extraction unit 5. For voice recognition, existing technology may be used. Since it is sufficient to be able to recognize an incoming call phrase and a turn-off phrase, which will be described later, a relatively light processing speech recognition technique may be used.

  The information about the utterance includes, for example, the start time Sci and end time Eci of each utterance Uci (i = 1, 2,...) Of the customer, and the start time Soi of each utterance Uoi (i = 1, 2,...) Of the operator. , End time Eoi, Mci word notation Wci1, Wci2,..., WciMci constituting each customer utterance Uci (i = 1, 2,...), Part of speech information Pci1, Pci2,..., PciMci, Information on Moi words constituting each utterance Uoi (i = 1, 2,...) Of the operator, Woi1, Woi2,.

  The holding section detection unit 4 uses information about an utterance, and the interval between adjacent utterances is a predetermined time or more, and at least one of the adjacent utterances includes a typical phrase used at the time of holding. A section is detected (step S4). Information about the detected pending section is sent to the extraction unit 5.

  The reserved section detection unit 4 includes a silent section extraction unit 41 and a fixed expression extraction unit 42. First, the silent section extraction unit 41 detects a silent section in which the interval between adjacent utterances is equal to or longer than a predetermined time, using information about the utterance. Then, the fixed expression extraction unit 42 determines whether a typical phrase used at the time of holding is included in at least one of the utterances adjacent to the detected silent section. If included, the silent section extraction unit 41 sets this silent section as a reserved section.

  Whether or not a phrase is included in an utterance is determined by whether or not M words or more constituting the phrase are included in the utterance. Here, the threshold value M is calculated as M = 求 め る N × k┐, where N is the total number of words constituting the phrase. However, k is an arbitrary constant of 0 or more and 1 or less, and ┌ / ┐ represents the smallest integer of ≧. If it is desired to extract the reserved section more accurately, k may be set to a large value, and if it is desired to reduce extraction omission, k may be set to a small value. Thus, by extracting the utterance based on the ratio including the words constituting the phrase, detection can be performed more flexibly than when all the words are matched. Whether or not a word is included in an utterance is determined by whether or not a word having the same notation and part of speech information as the notation and part of speech information of the word is included in the utterance. Alternatively, the part of speech information may be ignored and the determination may be made based on whether or not a word having the same notation as that word is included in the utterance.

  As typical phrases used at the time of holding, for example, “please wait a little”, “sorry for you”, etc. can be considered. “Please wait a bit” is composed of multiple words like “a little: verbs”, “o: coronal verbs”, “waits: verbs”, etc. The notation and part of speech information for each word is “notation: part of speech” Information ". Using at least one of these notations and part-of-speech information, it is determined whether or not the word is included in the utterance. Notation and part-of-speech information can be obtained by existing morphological analysis techniques.

  More specifically, the reserved section detection unit 4 performs the process of FIG. Hereinafter, the process of FIG. 3 will be described. In the process of FIG. 3, in this example, the pending section is extracted in consideration of only the operator's utterance Uoi (i = 1, 2,..., No). This is due to the fact that there are customers who do not interact and that the hold melody flows from the customer's phone.

  The silent section extraction unit 41 initializes i and h with i = 2 and h = 1 (step S41).

  The silent section extraction unit 41 determines whether i> No (step S42). No is the total number of utterances by the operator.

  If i> No. The silent section extraction unit 41 has an interval Eoi-So (i-1) between the operator's i-th utterance Uoi and the operator's i-1th utterance Uo (i-1) greater than a predetermined time Th. It is determined whether or not Eoi-So (i-1)> Th (step S43).

  If Eoi-So (i-1)> Th, the typical expression extraction unit 42 is typically used for holding at least one of the i-th utterance Uoi and the i-1th utterance Uo (i-1). It is determined whether or not a phrase is included (step S44).

  If included, the silent section extraction unit 41 sets the section between the i-th utterance Uoi and the (i-1) th utterance Uo (i-1) as the h-th reserved section (step S45). For example, the holding section is specified by the number of utterances between the start of the call and the utterance of the operator and the utterance of the customer. If the order from the start of the call of the (i-1) th utterance Uo (i-1) is Hsh, and the order from the start of the call of the ith utterance Uoi is Heh, the hth hold section is specified as Hsh to Heh. Is done.

  After step S45, the silent section extraction unit 41 increments h by 1 with h = h + 1 (step S46).

  After step S46, if it is determined in step S43 that Eoi-So (i-1)> Th is not satisfied, or if it is determined in step S44 that a typical phrase is not included, the silent section extraction unit 41 Sets i = i + 1 and increments i by 1 (step S47). After step S47, the process proceeds to step S42.

  If it is determined in step S42 that i> No, the silent section extraction unit 41 determines whether h is 1 (step S48). That is, it is determined whether a pending section is detected. When h is 1, it means that no reserved section has been detected.

  If h is not 1, the process of step S4 is finished and the process proceeds to step S5. If h is 1, the subsequent process of step S5 is not performed.

  In the example of FIG. 4, the section between the utterance Uo2 and the utterance Uo3 is equal to or greater than Th. Therefore, it is determined whether or not typical phrases used at the time of holding are included in the utterances Uo2 and Uo3 adjacent to this section. If the phrase is included, this section is set as the holding section 1, 3 that is the order from the start of the call of the utterance Uo2 is set to Hs1, and 5 that is the order from the start of the call of the call Uo3 is set to He1. The

  The extraction unit 5 extracts a larger number of utterances as the utterance section is longer from the set of utterances adjacent to the reserved section (step S5). The extracted utterance is subject to analysis as a hold-related utterance.

  In the example of FIG. 5, when the holding section is long, a total of nine utterances are extracted from the set of utterances adjacent to the holding section, and when the holding section is short, a total of five utterances are extracted from the set of utterances adjacent to the holding section. Extracting utterances.

  When the holding section is long, the cause of the holding is considered to be complicated, and the analysis target is set widely. On the contrary, when the holding section is short, the cause of the holding is considered to be simple, and the analysis target is set to be narrow. In this manner, the hold-related utterance can be more appropriately extracted by expanding / contracting the analysis target according to the length of the hold section.

  For example, ┌K × t┐ utterances are extracted from the set of utterances adjacent to the hold interval, where K is a predetermined constant, t is the length of the hold interval, and ┌ · ┐ is the smallest integer. For example, K = arbitrary number of utterances / average holding time. The arbitrary number of utterances is the number of utterances to be analyzed when holding is held for an average holding time, for example, five.

  Also, a set of utterances adjacent to the holding section and including a time length within K ′ × t time adjacent to the holding section is extracted, where K ′ is a predetermined constant and t is the length of the holding section. May be. For example, K ′ = arbitrary time length / average holding time.

  The set of utterances adjacent to the holding section is any of a set of utterances immediately before the holding section, a set of utterances immediately after the holding section, and a set of utterances immediately before and after the holding section. When the set of utterances adjacent to the hold section is a set of utterances immediately before and after the hold section, the number of hold related utterances extracted immediately before the hold section and the hold relation extracted from immediately after the hold section The number of utterances may be the same or different.

  Instead of performing the processing of the standard expression extracting unit 42, the holding section detecting unit 4 may set a section in which the interval between adjacent utterances is a predetermined time or more as a holding section.

  The hold-related utterance extraction apparatus and method can be realized by a computer. In this case, the processing content of each part of this apparatus is described by a program. Then, by executing this program on a computer, each unit in this apparatus realizes each step in this method on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. In this embodiment, these apparatuses are configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

  The present invention is not limited to the above-described embodiment, and can be modified as appropriate without departing from the spirit of the present invention.

DESCRIPTION OF SYMBOLS 1 Voice acquisition part 2 Voice feature-value extraction part 3 Voice recognition part 4 Reservation area detection part 41 Silent area extraction part 42 Standard expression extraction part 5 Extraction part

Claims (9)

An audio feature extraction step for extracting an audio feature of the audio signal;
A speech recognition step of performing speech recognition on the speech signal using the speech feature, an acoustic model, and a language model, detecting speech included in the speech signal, and generating information about the detected speech;
Using the information about the utterance, a holding section detecting step for detecting a holding section in which the interval between adjacent utterances is a predetermined time or more;
An extraction step of extracting a larger number of utterances as the holding section is longer from a set of utterances adjacent to the holding section;
Hold related utterance extraction method. In the holding related utterance extraction method according to claim 1,
The holding section detecting step uses information about the utterance, and an interval between adjacent utterances is a predetermined time or more, and at least one of the adjacent utterances includes a typical phrase used at the time of holding. A step of detecting a holding interval;
An on-hold related utterance extraction method. In the holding related utterance extraction method according to claim 1 or 2,
The above extracted utterances are ┌K × t┐ utterances, where K is a predetermined constant, t is the length of the holding section, and ┌ · ┐ is the smallest integer.
An on-hold related utterance extraction method. In the holding related utterance extraction method according to claim 1 or 2,
The extracted utterance is a utterance included in the time length within K ′ × t time adjacent to the reserved section, where K ′ is a predetermined constant and t is the length of the reserved section.
An on-hold related utterance extraction method. In the holding related utterance extraction method according to any one of claims 1 to 4,
The set of utterances adjacent to the reserved section is a set of utterances immediately before the reserved section.
An on-hold related utterance extraction method. In the hold related utterance extraction device according to any one of claims 1 to 4,
The set of utterances adjacent to the holding section is a set of utterances immediately after the holding section.
An on-hold related utterance extraction method. A voice feature amount extraction unit that extracts a voice feature amount of a voice signal;
A speech recognition unit that performs speech recognition on the speech signal using the speech feature, an acoustic model, and a language model, detects an utterance included in the speech signal, and generates information about the detected utterance;
Using the information about the utterance, a holding section detecting unit that detects a holding section in which the interval between adjacent utterances is equal to or longer than a predetermined time; and
An extraction unit that extracts a larger number of utterances as the holding section is longer from a set of utterances adjacent to the holding section;
Hold-related utterance extraction device. In the hold related utterance extraction device according to claim 7,
The holding section detection unit uses the information about the utterance, the interval between adjacent utterances is equal to or longer than a predetermined time, and at least one of the adjacent utterances includes a typical phrase used at the time of holding. Detect pending intervals,
A hold-related utterance extraction device characterized by the above.   A holding related utterance extraction program for causing a computer to execute each step of the holding related utterance extracting method according to claim 1.

Images