JPH03157696A - Voice responding and recognizing system - Google Patents

Abstract

PURPOSE:To improve recognition performance and to minimize the deterioration in the recognition performance by specifying a speaker by the transmission telephone number of the other person, using the standard pattern for the specific speaker only to the required person and changing the pattern to the standard pattern for nonspecific speakers when the result of the recognition is insufficient. CONSTITUTION:A network control section 1 specifies that the other person is the specific speaker when the transmission telephone number of the other person is received in a transmitter's number receiving section 5 of the network control section 1 via a switching network. The speaker is recognized rapidly by using the standard pattern for the specific speaker built in a voice recognizing section 2 and the voices of the other person are announced to a main control section 4 by which the voice response, such as required guidance, is executed and the recognition performance is enhanced. The voice recognizing section 2 changes the standard pattern to the standard pattern for the nonspecific speakers when the result of the recognition is insufficient. The recognition is thus executed and the deterioration in the recognition performance is thus minimized.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a recognition method for a voice response recognition device that is connected to a network such as the 15DN network and provides services, and in particular, by transmitting voice guidance. The present invention relates to a voice response recognition method that recognizes the user's voice uttered in response to the voice response and provides a service based on the recognition result.

[Conventional Technology J] Demand has been increasing in recent years because it is possible to significantly save labor, become more economical, and improve efficiency by replacing voice-based information provision services that were previously performed manually with voice responses by machines. Recognize voice.

Voice response recognition services that respond depending on the content include automatic guidance phone number notification service, seat reservation service,
It is widely used for securities companies' stock price notification services, bank balance inquiry services, etc.

When speech recognition methods are classified based on how they deal with individual differences, there are two types: specific speaker speech recognition and non-specific speaker speech recognition. In speaker-specific speech recognition, a standard pattern is used every time the speaker changes! ! I need to re-record it. To do this, training audio is required prior to recognition processing, so
The speaker must be pre-recorded for learning purposes. In contrast, in speaker-independent recognition, recognition processing can be immediately started for all speakers. Therefore, compared to specific speaker recognition, speaker-independent recognition requires a more advanced recognition technique that takes into account the variations in speech due to differences in speakers. A voice response recognition device that uses speaker-independent voice recognition technology is
It has already been commercialized and is used for services such as bank balance inquiries and transfer notifications, and stock price inquiries at securities companies.

In the first method of word speech recognition, the time series itself of the short-term spectrum of word speech is registered as a standard pattern. Then, the degree of similarity is directly calculated by matching the time series of the short-term spectrum of the input speech and the standard pattern. Also, in the second method.

Register the short-time spectrum of phonemes as a standard pattern,
Each word is registered in the word dictionary as a phoneme symbol sequence. In the recognition process, the time series of the short-time spectrum of the input speech is once recognized using a phoneme standard pattern, and the word similarity is calculated between the phoneme recognition result and the phoneme symbol sequence in the word dictionary.

For information on voice recognition methods, see, for example, "Electronic Information and Communication Handbook Vol. 1" March 30, 1988, Co., Ltd.
Published by Ohmsha, pp. 1191-1207.

[Problem to be solved by the invention]

By the way, in speech recognition for non-specific speakers, recognition performance is very good for some speakers, but recognition performance deteriorates significantly for certain speakers, so it cannot be evaluated based on average recognition performance alone. For example, there is a sheave boat phenomenon. Therefore, solving the deterioration of recognition performance for a specific speaker is an issue for systems using speech recognition.

However, conventional devices that have been put into practical use use speech recognition devices for unspecified speakers, and even if the recognition performance is not good for a specific speaker, standard patterns are added individually. It is not possible.

On the other hand, it is also possible to prepare individual standard patterns for all users, but this method requires an enormous amount of memory and is not practical in a large-scale system.

Another option is to install a voice response recognition device that has standard patterns for both unspecified speakers and specific speakers, but with the current analog network, it is possible to prepare standard patterns for specific speakers. However, it was difficult to realize this because it required some method of identifying the speaker. In addition, as a method for identifying the speaker, it is possible to provide a certain input opportunity after a call arrives at the voice response recognition device and have the speaker input signal (for example, subscriber number, etc.). means(
There are also problems such as the need to provide a signal transmitting/receiving device (signal transmitting/receiving device, etc.), and the sequence becomes complicated.

Note that standard patterns for non-specific speakers are usually designed to have high recognition performance on average for many people, so there is no particular problem. Improvements in recognition performance are desired only in these areas.

An object of the present invention is to solve such conventional problems and to provide a voice response that can improve recognition performance even for specific speakers among unspecified speakers for whom recognition performance is significantly degraded. The objective is to provide a recognition method.

[Means for Solving the Problems 1] In order to achieve the above object, the voice response recognition method of the present invention includes a standard pattern for 86 unspecified speakers and a standard pattern for a plurality of specific speakers in the voice response recognition device. Register and
If the caller's phone number is not notified, the standard pattern for unspecified speakers is selected and recognition processing is performed, and if the caller's phone number is notified, the standard pattern for unspecified speakers or the standard pattern for unspecified speakers is selected. If one of the standard patterns for a specific speaker corresponding to the telephone number is selected and recognized, and as a result of performing the recognition process using the standard pattern for the specific speaker, the predetermined performance is not satisfied. The feature is that it automatically changes to an unspecified pattern and continues the recognition process and subsequent services.

[Function] In the present invention, as one of the methods for identifying the other party,
The recognition performance is improved by using the caller's telephone number notified from the network and using a standard pattern for a specific speaker for the user identified by the caller's telephone number. That is, first, the caller is identified by receiving the caller's phone number notified from the network, and it is determined whether or not there is a need for a standard pattern for specified users for the identified user, and if necessary, , retrieve the specific speaker pattern by searching the telephone number-standard pattern correspondence table,
Recognition is performed using it. At this time, if the recognition performance is good, recognition is performed using a specific speaker pattern, and if the recognition performance is poor, recognition is performed using an unspecified speaker pattern.

[Example J Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 2 is a block diagram of a device to which the voice response recognition method of the present invention is applied.

The voice response recognition device includes a network control unit 1 that controls a network such as 15DN, and a voice recognition unit 2 that recognizes the speaker's voice.
and a voice response unit 3 that responds to an incoming call with voice;
It is composed of a main control section 4 that controls these circuits.

The network control unit l is equipped with a call origination/reception function for making and receiving calls, and a caller number receiving unit 5 for receiving the caller's telephone number, and is connected to the main control unit 4 via a communication line. We are exchanging and receiving information. When the voice recognition unit 2 recognizes the user's voice, it reports the result to the main control unit 4. Further, the voice response section 3 responds to the user with voice guidance, and outputs various different voice guidance under instructions from the main control section 4. The main control unit 4 not only controls each unit but also performs business processing such as service progress management and stored data search.

FIG. 3 is a diagram showing an example of the configuration of the speech recognition section in FIG. 2.

In FIG. 3, reference numeral 6 denotes a voice analysis unit that analyzes the input voice, 7 a standard pattern memory that stores standard patterns, 8 a matching operation unit that compares and matches the standard pattern and the input voice, and 9 is a result determination unit that determines whether or not there is a match as a result of comparison and verification, and compares it with a dictionary to determine which word it is, and lO is a dictionary memory that stores a word dictionary.

First, the voice analysis section 6 extracts voice characteristic parameters from the input voice. In the standard pattern memory 7, standard sounds are stored in advance in the form of characteristic parameters. The matching calculation unit 8 performs a matching calculation between the input voice and the standard pattern, calculates the distance between the input data and the standard pattern, and outputs the result to the result determination unit 9. The result determination unit 9 stores the dictionary memory 10 based on the distance calculation result.
, and determines what kind of words the input voice is, and notifies the main control unit 4 of the same.

FIG. 4 is a diagram showing an example of the configuration of the standard pattern stored in the standard pattern memory in FIG. 3.

Various methods can be considered for creating the standard pattern. Here, it is assumed that standard patterns for specific speakers and non-specific speakers have been created.

As shown in (a), two types of standard patterns for unspecified speakers are prepared: one for men and one for women. In this case, since it is initially impossible to determine whether the user is male or female, both are used for recognition. In addition, as shown in (b), as a standard pattern for specific speakers, n
Each will be prepared.

FIG. 1 is a processing flowchart of a voice response recognition method showing an embodiment of the present invention.

When there is an incoming call, it is checked whether the calling telephone number was notified at the time of the incoming call (step 101). If not, a standard pattern for an unspecified speaker is selected and processed (step 107). Additionally, if notified,
First, the main control unit 4 determines whether there is a need for a standard pattern for the telephone number and a specific speaker (step 102), and if there is no need, performs recognition processing using an unspecified speaker pattern (step 107). If necessary, a standard pattern corresponding to the telephone number is obtained from the table (step 103), and recognition processing is performed using the standard pattern for the specific speaker (step 104). As a result of the recognition processing (step 105), if the recognition performance is good, that specific speaker pattern is selected and subsequent recognition is performed using that standard pattern (step 106), and if the recognition performance is poor, the specific speaker pattern is selected. A specific speaker pattern is selected and subsequent recognition processing is performed (step 107). Note that the necessity of a standard button for a specific speaker is determined by collating the telephone number itself with the main control unit 4. Alternatively, a table describing telephone numbers and whether or not they are necessary may be used.

In this manner, the present invention improves overall speech recognition performance by selecting a standard pattern using the caller's telephone number notified from the network. However, even if the caller's phone number notification function is provided as a network function, the user can choose whether or not to send the phone number, so consideration should be given to what to do if the caller's phone number is not sent. There is. Additionally, the caller can generally be identified by using the calling phone number, but
For example, it may be used by family members, so it may not be exactly the same.

In consideration of such problems, the voice response recognition device of the present invention recognizes the user's voice and automatically responds while proceeding with the service. A standard pattern for users is selected and recognition processing is performed, and the user's calling phone number is notified.
Only when the system determines that a standard pattern for a specific speaker is necessary, a standard pattern for a specific speaker corresponding to the telephone number is selected and recognition processing is performed.

In addition, even if a user's calling phone number is received, if good recognition performance has already been obtained for that user using the standard pattern for speaker-independent use, We will select a standard pattern and perform recognition processing. Furthermore, in case the pattern selected in advance based on the calling phone number does not satisfy the predetermined performance, in case a person different from the expected user utters the same phone number,
The pattern is automatically changed to one for unspecified speakers and the service continues.

FIG. 5 is a diagram showing an example of a correspondence table between telephone numbers and standard patterns used in the present invention.

Here, it is assumed that n standard patterns are prepared corresponding to telephone numbers that require standard patterns for specific speakers. This table may be one that simply describes the starting addresses of standard patterns. Although different from the process shown in FIG. 1, it is also possible to perform the process using one table in addition to determining whether or not a standard pattern for a specific speaker is necessary.

FIG. 6 is a diagram showing an example of a table used in another embodiment of the present invention.

In FIG. 6, the telephone number portion shows the telephone numbers of all the subjects, and correspondingly the storage address numbers of the standard pattern to be selected are shown. Furthermore, if a standard pattern for unspecified speakers is used, this is stated. In this case, the pattern number O indicates unspecified use. Also, here, the storage address of the standard pattern corresponding to telephone number 1 is address 1000, and the storage address of the standard pattern corresponding to telephone number 4 is address 2.
Address 000...what's going on?

Using the table in Figure 6, it is possible to judge whether a standard pattern for a specific speaker is required for all telephone numbers using O or a number other than 0, and All you have to do is look at the standard pattern storage address for , access the memory at that address, and read the standard pattern.

FIG. 7 is a diagram showing an example of changing the standard pattern used in the present invention.

Here, a specific example is shown in which the standard pattern is changed by checking the recognition performance.

That is, in the present invention, since the user is specified by the telephone number, there may be cases where the expected user does not necessarily speak. Furthermore, in the case of a standard pattern for a specific speaker, it may be difficult to respond to sudden changes in voice quality such as when a person has a cold. Therefore, when a standard pattern for a specific speaker corresponding to a telephone number is selected and used, the recognition performance is checked, and if the results are not good, the standard pattern is changed again to the standard pattern for an unspecified speaker and recognition processing is performed. Proceed.

For example, as shown in Figure 7, first the output guidance says "
"Please give me your PIN number" is sent to the caller, and when the input voice from the caller is lrl 234A, this is recognized using a standard pattern for a specific speaker, and the results are not very good.So, The output guidance is ``R2285,'' which is the PIN number resulting from the recognition. In response, the caller returned a negative answer of ``No,'' so the output guidance was ``Please enter your PIN again.'' When the sender sends the PIN number 1r1234j again, recognition processing is performed using a standard pattern for unspecified speakers. In this way, by checking the recognition performance and changing the standard pattern for specific speakers to the standard pattern for non-specific speakers as appropriate, recognition deterioration due to inappropriate standard pattern selection can be minimized to the necessary minimum. Can be done. Note that although a command is issued to instruct the recognition device to perform recognition, only a few bytes of a standard pattern number are added after the command, so it is possible to change the standard pattern in real time.

In the past, services were provided using only standard patterns for unspecified speakers, resulting in a phenomenon in which recognition performance was significantly degraded for specific people. Recognition performance is improved by identifying speakers to some extent by numbers and using standard patterns for specific speakers only for those who need them. In addition, the recognition performance is checked even after selecting a standard pattern for a specific speaker, in case the expected speaker does not appear for the calling phone number, or the voice quality changes due to a cold, etc. However, if the performance is not sufficient, it is possible to change back to the standard pattern for unspecified speakers.

〔Effect of the invention〕

As explained above, according to the present invention, the speaker is identified by the calling party's telephone number, and recognition processing is performed using a standard pattern for the specific speaker only for those who need it, thereby improving recognition performance. In addition to improving recognition performance, if recognition performance is insufficient, by changing to a standard pattern for unspecified speakers,
Deterioration in recognition performance can be minimized.

[Brief explanation of the drawing]

FIG. 1 is a processing flowchart of a voice response recognition method showing an embodiment of the present invention, FIG. 2 is a block diagram of a voice response recognition device using the recognition method of the present invention, and FIG. 3 is a voice recognition method in FIG. FIG. 4 is a diagram showing a configuration example of a standard pattern used in the present invention, FIG. 5 is a diagram showing an example of a telephone number standard pattern correspondence table used in the present invention, and FIG. 6 is a diagram showing an example of the standard pattern correspondence table used in the present invention. FIG. 7 is an explanatory diagram of a correspondence table showing another embodiment of the present invention, and is an explanatory diagram of a case where a standard pattern is changed by a recognition check in the present invention. 1: Network control unit, 2: Voice recognition unit, 3: Voice response unit, 4:
Main control unit, 5: Caller number receiving unit, 6: Voice analysis unit, 7
: Standard pattern memory, 8: Verification calculation section, 9: Result judgment section, lO: Dictionary. Figure: Incoming call size L: 1 m; YO Figure: Example of telephone number standard slam correspondence table Figure 6

Claims (1)

[Claims] (1) Receive the caller's phone number notified from the network, compare and recognize the caller's voice with a standard pattern, and respond to the caller with voice guidance based on the recognition results to provide service. In the voice response recognition device to be advanced, a standard pattern for unspecified speakers and a standard pattern for multiple specific speakers are registered in the voice response recognition device, and if the caller's telephone number is not notified, , selects the standard pattern for unspecified speakers and performs recognition processing, and when the caller's phone number is notified, the standard pattern for unspecified speakers or the specific speaker corresponding to the phone number is selected. Select one of the standard patterns for the specific speaker and perform the recognition process, and if the result of the recognition process using the standard pattern for the specific speaker does not meet the predetermined performance, it will automatically change to the unspecified pattern above. A voice response recognition method characterized in that the recognition process is repeated and the subsequent service is continued.