JPH06152728A - Answer message switching speech equipment - Google Patents

Abstract

PURPOSE:To provide a speech equipment such as a telephone equipment and an interphone equipment, etc., by which it is unnecessary to limit an object voice to the utterance contents learned in advance, at the time of recognizing a speaker. CONSTITUTION:At the time of recognizing a speaker by using a neural network, in a speaker recognizing part of the answer message switching speech equipment, the sequence of a vector for showing a rough form of a spectirum is inputted for a short period, the sequence of a network output is integrated by the sum, the product, the majority, etc., of the result of recognition by each output and one result of recognition is obtained, and the number of learning data of the network is curtailed by clustering.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a response message switching communication device capable of returning a special message addressed to a person who has made a call. Also,
The present invention relates to a response message switching intercom device capable of returning a special message addressed to a user of an intercom installed at a front door or the like.

[0002]

2. Description of the Related Art Conventional Japanese Patent Publication No. 3-59619 "Answering machine with artificial intelligence function", Japanese Unexamined Patent Publication No. 61-39756 "Answering telephone system", Japanese Unexamined Patent Publication No. 64-86742 "Message storage and reproducing apparatus", Japanese Unexamined Patent Publication No. Hei 2 -
294146 "Recording / playback device" etc. have been proposed. Although these prior arts have been shown to use speech recognition (or speaker recognition) technology, there is no description of a specific recognition method, or even if it is described, it is only as shown below. Also, most of them are based on the premise that the caller speaks a specific word (keyword, caller's name, etc.).

A feature amount (spectrum, pitch, cepstrum, etc.) included in the input voice is extracted, and a distance between the extracted data and dictionary data previously extracted by a similar method is calculated (DP matching, etc.). Then, the voice of the input voice is determined from the calculation result of this, and the subsequent processing is controlled according to the determination result.

Further, conventionally, Japanese Patent Laid-Open No. Sho 61-39756 "answering machine system" has been proposed. This prior art has been shown to utilize a voice recognition technique, but the description of a specific recognition method is as follows. Also, it is premised that the intercom user speaks a keyword.

A feature amount (pitch, combination of vowel and consonant, etc.) included in the input voice is extracted, and a distance between the extracted data and dictionary data previously extracted by a similar method is calculated, The voice of the input voice is determined from the result of the calculation, and the subsequent processing is controlled according to the result of the determination.

[0006]

In the above-mentioned prior art, there is little description about a specific recognition method of voice recognition (or speaker recognition), and there is room for improvement in recognition rate in the described recognition method. It seems that there is. In addition, when it is assumed that the caller utters a specific word, it is not easy to use, and if other words are input, malfunction may occur. Also, since it is premised that the intercom user speaks a keyword, it is not easy to use, and if other words are input, malfunction may occur.

It is an object of the present invention to provide a communication device such as a telephone device or an intercom device, which does not need to limit a target voice to a pre-learned utterance content when recognizing a speaker. .

[0008]

The present invention as set forth in claim 1, is a response that recognizes who the calling party is and can respond to a person who is registered in advance with a special message addressed to that person. In a message switching telephone device, an automatic response unit that automatically responds to an incoming call, a speaker recognition unit that recognizes a speaker from the voice of a caller who responds to the automatic response unit, and a recognition of the speaker recognition unit According to the result, it has a response control unit for controlling the basic communication circuit and a message storage unit for storing a message. When the speaker recognition unit recognizes the speaker by using the neural network, it represents the outline of the short-time spectrum. A series of vectors is input, and the series of network outputs are combined by summing, multiplying, and majority voting of the recognition results of the individual outputs to obtain one recognition result. It is obtained so as to reduce the Staring.

The present invention as set forth in claim 2 is a response message switching telephone apparatus which recognizes who the calling party is and can respond to a person who is registered in advance with a special message addressed to the person. An automatic response unit that automatically responds to an incoming call, a speaker recognition unit that recognizes a speaker from the voice of the caller who responds to the automatic response unit, and a call basic circuit based on the recognition result of the speaker recognition unit. When a speaker is recognized using a neural network in the speaker recognition unit, which has a response control unit for controlling and a message storage unit for storing a message,
By inputting a series of vectors that represent the outline of the short-time spectrum and selecting the output vector from the network output series using the output vector selection threshold, the sum, product, majority vote, etc. of the recognition results for each output One recognition result is obtained as a whole, and the number of learning data of the network is reduced by clustering.

The present invention according to claim 3 provides a response message switching intercom device which recognizes who the intercom user is and can respond to a person who is registered in advance with a special message addressed to the person. An intercom (slave unit) installed at a front door, an automatic response unit that automatically responds to the intercom user, and a speaker recognition unit that recognizes a speaker from the voice of the intercom user who responds to the automatic response unit. And a response control unit that controls the intercom basic circuit according to the recognition result of the speaker recognition unit, and a message storage unit that stores a message. When the speaker recognition unit uses the neural network to recognize the speaker. ,
A series of vectors that represent the outline of a short-time spectrum is input, and the series of network outputs are combined by summing, recognizing, and majority voting of the recognition results of individual outputs to obtain a single recognition result. The number of learning data is reduced by clustering.

According to a fourth aspect of the present invention, in a response message switching intercom device, which recognizes who the intercom user is and responds to a person who is registered in advance with a special message addressed to the person, the front door is provided. An intercom (slave unit) installed in the front, an automatic response unit that automatically responds to the intercom user, and a speaker recognition unit that recognizes a speaker from the voice of the intercom user who responds to the automatic response unit. , A response control unit that controls the intercom basic circuit according to the recognition result of the speaker recognition unit, and a message storage unit that stores a message, and when the speaker recognition unit uses the neural network to recognize the speaker,
By inputting a series of vectors that represent the outline of the short-time spectrum and selecting the output vector from the network output series using the output vector selection threshold, the sum, product, majority vote, etc. of the recognition results for each output One recognition result is obtained as a whole, and the number of learning data of the network is reduced by clustering.

The present invention according to claim 5 uses the telephone when recording / learning the voice of the registrant by combining the response message switching intercom apparatus according to claim 3 or 4 with the telephone. It is made possible to record audio.

[0013]

The operation of the speaker recognition section of the present invention according to claim 1 will be described. First, the neural network is trained by using the outline of the short-time spectrum obtained from the training voice. At this time, the number of learning data is reduced by performing clustering for each speaker. At the time of recognition, an outline of the same short-time spectrum as described above is obtained from an arbitrary utterance, the sequence is input to the network, and a sequence of network output is obtained. The output vector of the obtained network suggests the speaker for each input for a short time, and by making a comprehensive judgment such as sum, product, majority decision, etc. in the entire series, one recognition result is obtained. To get

In the speaker recognition method according to the first aspect, by using the speaker recognition technology that does not limit the utterance content, it is not necessary for the caller to utter a specific word, and the answer message switching telephone device is provided. The usability has improved. In addition, it became possible to respond to any words entered.

Since the representative vector of a plurality of data is used as learning data by clustering, the number of learning data can be reduced while maintaining the learning effect. As a result, the learning time of the neural network can be greatly reduced.

The operation of the speaker recognition section of the present invention according to claim 2 will be described. First, the neural network is trained by using the outline of the short-time spectrum obtained from the training voice. At this time, the number of learning data is reduced by performing clustering for each speaker. At the time of recognition, an outline of the same short-time spectrum as described above is obtained from an arbitrary utterance, the sequence is input to the network, and a sequence of network output is obtained. The output vector of the obtained network suggests a speaker for each input for a short time, but a threshold for output vector selection is provided, and only a reliable output vector is selected among them. One recognition result is obtained by making a comprehensive judgment such as sum, product, majority vote, etc. for all of these. A specific method of selecting an output vector is, for example, a method of selecting a vector that exceeds a preset threshold value for selecting an output vector even if one value is included in the output vector.

According to the speaker recognition method of the second aspect, by selecting a highly reliable output vector from among the output vectors, the comprehensive judgment can be made more reliable and the recognition rate can be improved.

The operation of the speaker recognition unit of the present invention according to claim 3 will be described. First, the neural network is trained by using the outline of the short-time spectrum obtained from the training voice. At this time, the number of learning data is reduced by performing clustering for each speaker. At the time of recognition, an outline of the same short-time spectrum as described above is obtained from an arbitrary utterance, the sequence is input to the network, and a sequence of network output is obtained. The output vector of the obtained network suggests the speaker for each input for a short time, and by making a comprehensive judgment such as sum, product, majority decision, etc. in the entire series, one recognition result is obtained. To get

In the speaker recognition method according to the third aspect, by using the speaker recognition technology that does not limit the utterance content, it is not necessary for the intercom user to utter a keyword, and the intercom response message switching interphone device is easy to use. Has improved. In addition, it became possible to respond to any words entered.

Since the representative vector of a plurality of data is used as learning data by clustering, it is possible to reduce the number of learning data while maintaining the learning effect. As a result, the learning time of the neural network can be greatly reduced.

The operation of the speaker recognition section of the present invention according to claim 4 will be described. First, the neural network is trained by using the outline of the short-time spectrum obtained from the training voice. At this time, the number of learning data is reduced by performing clustering for each speaker. At the time of recognition, an outline of the same short-time spectrum as described above is obtained from an arbitrary utterance, the sequence is input to the network, and a sequence of network output is obtained. The output vector of the obtained network suggests a speaker for each input for a short time, but a threshold for output vector selection is provided, and only a reliable output vector is selected among them. One recognition result is obtained by making a comprehensive judgment such as sum, product, majority vote, etc. for all of these. A specific method of selecting an output vector is, for example, a method of selecting a vector that exceeds a preset threshold value for selecting an output vector even if one value is included in the output vector.

According to the speaker recognition method of the fourth aspect, by selecting a highly reliable output vector from among the output vectors, comprehensive judgment can be made more reliable and the recognition rate can be improved.

However, the description of the "neural network" in the present invention is as follows (1) to (4).

(1) The neural network has the structure shown in FIG.
It can be roughly classified into two types of mutual coupling networks shown in (B). The present invention may be configured by using either of both networks, but the hierarchical network is more useful because a simple learning algorithm as described later has been established.

(2) Network Structure As shown in FIG. 6, the hierarchical network has a hierarchical structure including an input layer, an intermediate layer, and an output layer. Each layer is composed of one or more units. The coupling is only forward coupling such as input layer → middle layer → output layer, and there is no coupling within each layer.

(3) Structure of Unit The unit is a model of a brain neuron as shown in FIG. 7, and its structure is simple. Receive input from other units,
The sum is taken and converted by a certain rule (conversion function), and the result is output. A variable weight that represents the strength of the connection is attached to each of the connections with other units.

(4) Learning (Back Propagation) Learning the network is to bring the actual output closer to the target value (desired output), and generally, the conversion function and weight of each unit shown in FIG. Is learned by changing.

As a learning algorithm, for example, Rumelhart, DE, McClelland, JL and the PDP Re
search Group, PARALLEL DISTRIBUTED PROCESSING, the
Backpropagation described in MIT Press, 1986. can be used.

[0029]

1 is a schematic diagram showing an example of a response message switching telephone device, FIG. 2 is a schematic diagram showing an output value of a speaker, FIG. 3 is a schematic diagram showing an example of a response message switching intercom device, and FIG. 5 is a schematic diagram showing a speaker output value, FIG. 5 is a schematic diagram showing a neural network, FIG. 6 is a schematic diagram showing a hierarchical neural network, and FIG. 7 is a schematic diagram showing a structure of a unit.

(First Embodiment) (Refer to FIGS. 1 and 2) FIG. 1 shows a response message switching telephone apparatus, and its use procedure is as follows (A) to (C).

(A) When learning the voice of the registrant, it is as shown in Table 1.

[0032]

[Table 1]

(B) When registering a message Pick up the handset Press the registration mode button Synthetic sound "Who is it?" Specify Mr. Yamada with the PB button (You can also use voice input instead of the PB button) Synthetic sound "Message "Voice, Mr. Yamada, at the usual place at 8 o'clock" Put the handset and return to normal mode

(C) When a call is received, it is as shown in Table 2.

[0035]

[Table 2]

The details of the speaker recognition method will be described below. For 5 registrants and 25 non-registrants, the text for learning was sampled at a sampling frequency of 10 kHz, frame length of 25.6 msec,
Fourier analysis with a frame period of 12.8 msec, 100 to 5000 Hz
A series of 68ch (1/12 Oct.) power vectors is obtained in the band.

Hierarchical clustering is performed from these power vectors to obtain about 200 representative vectors for each speaker.

These representative vectors are used as inputs of the neural network (input layer 68 units, input pattern is obtained by the number of speakers × the number of representative vectors after clustering), and the corresponding output unit is activated only in the case of a registrant. To learn enough.

A sequence of power vectors is obtained in the same manner as for any utterance.

This is input to the network learned by, and the sequence of output vectors {x ¹ , x ² ..., X ⁿ } x ^t = (x ^t ₁ , ..., x ^t ₅ ) n: the number of frames is obtained.

The following three methods are used for the above vector series to determine whether the input is a registered person or a non-registered person. (1) Registered if the maximum value of Σ _t x ^t _s (s = 1 to 5) exceeds the preset threshold for speaker determination, otherwise unregistered (2) Π _t If the maximum value of x ^t _s (s = 1 to 5) exceeds the preset threshold for speaker determination, it is a registrant; otherwise, it is a non-registrant (3) max {x ^t ₁ , ... , X ^t ₅ } = x ^t _S (s = 1 to 5) has the maximum value,
Registered person if it exceeds the threshold for speaker determination set in advance, otherwise non-registered person

The following methods may be used instead of the above three methods. (1) Only the maximum value of Σ _t x ^t _s (s = 1 to 5) exceeds a preset first speaker determination threshold value, and other values have a preset second speaker. If it is below the threshold for judgment, it is the registered person, otherwise it is the non-registered person. (2) Only the maximum value of Π _t x ^t _s (s = 1 to 5) is set as the preset first speaker judgment. If it exceeds the threshold for communication and the other values are below the preset threshold for second speaker determination, it is a registrant; otherwise, it is a non-registrant (3) max {x ^t ₁ , …, X ^t ₅ } = x ^t _s (s = 1 to 5)
Exceeds a preset first speaker determination threshold value, and
If the other values are below the preset threshold for second speaker determination, it is a registered person, otherwise, it is a non-registered person.

As a result of the above, when it is determined that the user is a registrant, it is determined who the speaker is based on which unit has the maximum output value.

As an example of an arbitrary utterance, in response to the short sentence for learning, "I'm sorry I'm going to Tokyo tomorrow, but I'll be away." When I did a recognition experiment,
We were able to fully recognize the 5 registrants used for learning and 26 non-registrants not used for learning.

(Second Embodiment) (Refer to FIGS. 3 and 4) FIG. 3 shows a response message switching intercom apparatus, and its use procedure is as follows (A) to (C).

(A) When learning the voice of the registrant (when the voice is recorded using the telephone), it is as shown in Table 3.

[0047]

[Table 3]

(B) When registering a message, take the handset and press the registration mode button. Synthetic sound "Who is it?" PB button designates a 〇 × liquor store (voice input is also possible instead of the PB button) Synthetic sound " Please leave a message. ”Say“ Beer, 1 case please ”Put the handset Back to normal mode

(C) When there is a visitor, it is as shown in Table 4.

[0050]

[Table 4]

The details of the speaker recognition method will be described below. For 5 registrants and 25 non-registrants, the text for learning was sampled at a sampling frequency of 10 kHz, frame length of 25.6 msec,
Fourier analysis with a frame period of 12.8 msec, 100 to 5000 Hz
A series of 68ch (1/12 Oct.) power vectors is obtained in the band.

Hierarchical clustering is performed from these power vectors to obtain about 200 representative vectors for each speaker.

These representative vectors are used as the input of the neural network (input layer 68 units, the input pattern is obtained by the number of speakers × the number of representative vectors after clustering), and the corresponding output unit is activated only in the case of the registrant. To learn enough.

A sequence of power vectors is obtained in the same manner as for any utterance.

This is input to the network learned by, and the sequence of output vectors {x ¹ , x ² ..., X ⁿ } x ^t = (x ^t ₁ , ..., x ^t ₅ ) n: the number of frames is obtained.

The following three methods are used for the above vector series to determine whether the input is a registered person or a non-registered person. (1) Registered if the maximum value of Σ _t x ^t _s (s = 1 to 5) exceeds the preset threshold for speaker determination, otherwise unregistered (2) Π _t If the maximum value of x ^t _s (s = 1 to 5) exceeds the preset threshold for speaker determination, it is a registrant; otherwise, it is a non-registrant (3) max {x ^t ₁ , ... , X ^t ₅ } = x ^t _s (s = 1 to 5) is maximum,
Registered person if it exceeds the threshold for speaker determination set in advance, otherwise non-registered person

The following methods may be used instead of the above three methods. (1) Only the maximum value of Σ _t x ^t _s (s = 1 to 5) exceeds the preset first speaker determination threshold value, and the other values have the preset second story. If it is below the threshold for person determination, it is the registered person, otherwise it is the non-registered person. (2) Only the maximum value of Π _t x ^t _s (s = 1 to 5) is set to the preset first speaker. If the threshold value for judgment is exceeded and the other values are lower than the preset threshold value for second speaker judgment, the registered person; otherwise, the non-registered person (3) max {x ^t ₁ ,…, X ^t ₅ } = x ^t _s (s = 1 to 5)
Exceeds a preset first speaker determination threshold value, and
If the other values are below the preset second speaker determination threshold value, it is a registered person, otherwise, it is a non-registered person.

As a result of the above, when it is determined that the user is a registrant, it is determined who the speaker is based on which unit has the maximum output value.

As an example of voluntary utterance, in response to the short sentence for learning, "I'm sorry I'm going to Tokyo tomorrow, but I'll be away." When I did a recognition experiment,
We were able to fully recognize the 5 registrants used for learning and 26 non-registrants not used for learning.

Furthermore, the present invention can be applied to various devices that require switching of reactions depending on users when a plurality of people use one device.

[0061]

As described above, according to the present invention, in a communication device such as a telephone device or an intercom device, when a speaker is recognized, it is not necessary to limit a target voice to a pre-learned utterance content. Can be obtained.

[Brief description of drawings]

FIG. 1 is a schematic diagram showing an example of a response message switching telephone device.

FIG. 2 is a schematic diagram showing output values of a speaker.

FIG. 3 is a schematic diagram showing an example of a response message switching intercom apparatus.

FIG. 4 is a schematic diagram showing output values of a speaker.

FIG. 5 is a schematic diagram showing a neural network.

FIG. 6 is a schematic diagram showing a hierarchical neural network.

FIG. 7 is a schematic diagram showing a structure of a unit.

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Office reference number FI technical display location H04M 9/00 D 8523-5K

Claims (5)

[Claims] 1. A response message switching telephone device capable of recognizing who a calling party is and responding to a person who is registered in advance with a special message addressed to that person, by automatically responding to an incoming call. A response section, a speaker recognition section for recognizing a speaker from the voice of the caller who replies to the automatic response section, a response control section for controlling the basic communication circuit based on the recognition result of the speaker recognition section, and a message When a speaker is recognized by the above-mentioned speaker recognition unit using a neural network, a series of vectors representing the outline of a short-time spectrum is input and the network output series is It obtains one recognition result by summing, accumulating, and majority decision of recognition results by the output of, and is characterized by reducing the number of learning data of the network by clustering. Response message switching communication device. 2. An answering message switching telephone device capable of recognizing who a calling party is and responding to a person who is registered in advance with a special message addressed to that person, by automatically answering an incoming call. A response section, a speaker recognition section for recognizing a speaker from the voice of the caller who replies to the automatic response section, a response control section for controlling the basic communication circuit based on the recognition result of the speaker recognition section, and a message When a speaker is recognized using a neural network in the speaker recognition unit, a series of vectors representing the outline of a short-time spectrum is input, and an output vector is output from the series of network outputs. For the output vector selected using the threshold for selection, one recognition result is obtained by summing the recognition results by the individual outputs, products, majority voting, etc. A response message switching communication device, which is characterized by reducing the number of training learning data by clustering. 3. An intercom device installed at a front door or the like in a response message switching intercom device capable of recognizing who the intercom user is and responding with a special message addressed to the person registered in advance. Slave unit), an automatic response unit that automatically responds to the intercom user, a speaker recognition unit that recognizes a speaker from the voice of the intercom user who responds to the automatic response unit, and a speaker recognition unit of the speaker recognition unit. A response control unit that controls the intercom basic circuit according to the recognition result,
A message storage unit for storing a message is provided, and in the speaker recognition unit, when a speaker is recognized using a neural network, a series of vectors representing an outline of a short-time spectrum is input, and a series of network output is An answer message switching intercom device, which obtains one recognition result by summing, accumulating, majority voting, etc. of recognition results by individual outputs, and reduces the number of learning data of a network by clustering. 4. An intercom installed in a front door or the like in a response message switching intercom device capable of recognizing an intercom user and responding to a person who is registered in advance with a special message addressed to the person. Device), an automatic response unit that automatically responds to the intercom user, a speaker recognition unit that recognizes a speaker from the voice of the intercom user that responds to the automatic response unit, and a recognition of the speaker recognition unit. A response control unit that controls the intercom basic circuit according to the result,
A message storage unit for storing a message is provided, and when the speaker recognition unit uses the neural network to recognize a speaker, a series of vectors representing the outline of the short-time spectrum is input and output from the network output series. For the output vector selected using the vector selection threshold, one recognition result is obtained by summing the recognition results by each output, product, majority vote, etc. Clustering the number of learning data of the network. Response message switching intercom device characterized by reduction by 5. A response characterized in that, by combining the response message switching intercom apparatus according to claim 3 or 4 with a telephone, voice recording can be performed using the telephone when recording / learning the voice of the registrant. Message switching intercom device.