JP3088625B2 - Telephone answering system - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION This invention is related to the telephone response system
More specifically, speech recognition technology using speaker adaptation
And a telephone answering system using the same.

[0002]

2. Description of the Related Art Conventionally, telephone answering systems using telephones have been used. In a conventional telephone answering system, various measures are taken to improve the recognition performance for recognizing a voice and increase the reliability of the system.
For example: (1) limit the vocabulary of recognition symmetry to the minimum required.

(2) The variance of the distribution in a standard pattern used for speech recognition is reduced, and the rate of rejecting unnecessary sounds (irrelevant conversational speech, noise, etc.) is increased.

(3) A speech recognition method of a specific speaker for limiting a person to be recognized is used.

(4) Speech recognition of a speaker adaptation system is used.

And so on.

[0007] In the speech recognition of the speaker adaptation system, a specific vocabulary is uttered by a speaker, and the speech of the uttered speaker is recognized using the vocabulary.

The utterance of this specific vocabulary needs to be performed at least about three times in order to use the pattern data for recognition.

In the conventional speaker adaptation system in speech recognition, since a specific vocabulary is uttered by the speaker, it is a burden on the speaker, which causes annoyance.

[0010] In addition, in a telephone installation location, a large noise peculiar to the installation location is often input, and this noise is intended to be recognized as voice, so that the voice recognition performance may be reduced.

[0011]

SUMMARY OF THE INVENTION An object of the present invention is to make it possible for a telephone answering system to perform speech recognition by a speaker adaptation method without causing a speaker to feel troublesome.

[0012] Also, a single caller first speaks
Perform speaker adaptation in real time using words as target words
A telephone answering system is provided.

[0013]

To achieve the above object, according to an aspect of, the invention are according to claim 1, in a telephone answering system with voice recognition, it sends the message corresponding to the recognized speech using a standard pattern, At the start of the call
The target word is the single word specified first,
A specific word recognition means for recognizing with a standard pattern, the
Identify the specific word recognized by the specific word recognition means.
A word pattern acquisition means for acquiring as a constant word pattern, using a specific word pattern <br/> over emissions obtained by the word pattern acquisition means, a new target rewrites the reference pattern
And the standard pattern changing means for changing a quasi-pattern, call final
Is rewritten by the standard pattern changing means upon completion.
And a telephone response system, characterized in that it includes a standard pattern returning means for returning to the original reference pattern and standard patterns.

According to the second aspect of the present invention, the object word is "high".
The telephone answering system according to claim 1, wherein

[0015]

[0016]

[Function] When a telephone conversation is performed, the same person responds from the appearance of the person who will be talking after the incoming call (hook off) (hereinafter referred to as the telephone responder) to the end of the call (hook on). There is. Therefore, the utterance characteristics of the telephone correspondent are registered in the system in real time at the beginning of one call, and the voice recognition is performed based on the characteristics until the flow during the call ends, thereby real-time speaker adaptation. Is performed, the recognition rate of voice recognition is improved. In this case, adaptation is performed using one uttered voice first as learning data. It is not yet possible to target all the words to be recognized for speaker adaptation, and in this system, the most important word is selected from the words to be recognized and speaker adaptation is performed on this one word. For this reason, the recognition rate of the most important words is improved, and the reliability of the entire system is increased.

Further, since noise peculiar to the telephone installation location is removed by performing the same processing as speech recognition, the degree of recognition is improved and the reliability of the system is increased.

[0018]

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

In the embodiment, a telephone answering system will be described in which an automatic contact system for power operation in a power company is applied.

FIG. 1 is a block diagram showing an apparatus configuration in a case where a telephone answering system according to an embodiment of the present invention is applied to an automatic communication system for power operation. In FIG. 1, reference numerals 1 and 2 denote voice recognition response systems. One housing can correspond to four channels, and there are two units. The voice recognition response systems 1 and 2 are connected to a private branch exchange (PB
X) It is connected to the public line 6 via 5, and is connected to the telephone of the consumer of the electric power. The PBX 5 is also connected to other telephones 4 in the company. Voice recognition response device 1
And 2 are system operation terminals 7 via Ethernet 3
The system operation terminal 7 controls the voice recognition response devices 1 and 2.

FIG. 2 is a block diagram showing the functional configuration of the voice recognition response devices 1 and 2 and the system operation terminal 7. In the system operation terminal, the operating system (O
S) Above, various application systems are operating. A part of this application system includes a speech recognition application development support tool group and a speech recognition control library, which are used to develop and control a speech recognition response device. The voice recognition response device has a voice recognition response driver and a voice recognition card on an operating system, and operates as a voice recognition response server.

FIG. 3 is a flow chart for explaining a message flow in the above-described automatic communication system for power operation. First, the automatic contact system for power operation automatically calls a customer. At that time, "This is the SNW telephone communication system for the Tokyo Electric Power Company Ginza Branch. Is it Nagashima Shoji. Please answer" high "or" no. " (301) and the message are transmitted by voice from the telephone recognition response system. As a result, the next message is sent as shown in FIG. 3 in response to the "high" or "no" response to the caller.

If there is a response of "high" (302), in this embodiment, "we will contact you about SNW line 1 stop power transmission. If you want to request the operation as it is," high ", someone If you are in charge of the person in charge, answer "No". (303) and a message is sent.

When there is "No" (316), the message "317 will be connected directly to the employees of our company, please wait." Automatically connect the telephone to the member (318).

Now, after the message of 303, "high"
If the answer is (304), the next message is sent. In the case of the answer of "No" (319), a message "320 (OK), please do not hang up the phone and change to the person in charge."
Return to 3. The loop does not exit until a "high" (304) response is received within a predetermined number of times or within a predetermined time from the message of 303. Well, this "high" (304)
"Responder" who responds to future calls
It is.

The automatic operation system for power operation sends a message as shown in FIG. 3 and performs automatic communication while changing the message according to the response (3.
05-315, 322-329). If the automatic contact system does not respond, the telephone is automatically connected to the waiting agent (318).

In the automatic communication system for power operation, a "high" response is important. For example, in FIG. 3, the communication items are confirmed in response to “high” in 310 and 312. Failure to properly recognize this "high" response significantly reduces the reliability of the automatic contact system. Therefore, it is necessary to set the first "high" (304) of the "telephone correspondent" of this call as a target word for speaker adaptation.

Adaptation using the "high" of 304 will be described below.

First, it is recognized that it is "high". FIG. 4 illustrates a standard pattern distribution indicating the probability of extracting the utterance feature of the input speech “high”, FIG. 4 illustrating speech recognition, and FIG. 5 illustrates a stochastic automaton of speech recognition “high” ( This will be described with reference to FIG. 6, which is a hidden Markov model (HMM).

The comparison between each input voice and the standard pattern will be described with reference to FIG. As shown in FIG. 4A, an input voice is input. Then, in the speech recognition section, the input speech is decomposed into each frame Xi. Each frame Xi is compared with a standard pattern distribution SH of phonemes of "H" in the system (see FIG. 4B). The standard pattern distribution S is shown as a function of an acoustic space composed of phoneme features and a probability distribution corresponding to the acoustic space.

FIG. 5 shows how the probability is obtained for each phoneme. Decomposing "high" into phonemes,
There are five types: "H", "HA", "A", "AI", and "I". Therefore, the symbol string having the highest probability is recognized as “high”. In order to determine the generation probability of this symbol string, a standard pattern distribution is prepared for each phoneme. This is illustrated in FIG. Using this standard pattern distribution, the probability of each phoneme of each frame of the input voice is obtained. Thus, the voice recognition is performed by determining the probability of “high”. The probability of “high” is shown in the diagram of the stochastic automaton in FIG. 6 (Hidden Markov Model HM
M). In this figure, S _{1 to} S ₅
Is a state indicating detection of each phoneme segment. State S
₁ has a high probability of detecting the "H", or return to the state S ₁ is detected after the loop, advances to state S _2. Hereinafter, the state S _{2 is} similarly set.
Has a high probability of detecting the "HA", to return to the state S ₂ is detected after the loop, advances to state S _3. States S ₃ , S ₄ , S
_The same applies to ₅ . Then, when finally goes to state S _6, the probability automaton is completed, the probability of "high" is determined with respect to the input voice.

In this way, when "high" is recognized,
The standard pattern of each phoneme is changed so that the probability of each "high" phoneme of the recognized "phone correspondent" is increased. This will be described with reference to FIGS.

Here, Xi is each frame of the voice recognized as "high". Each of these frames is stored in the system. Now, the frames that are recognized as "H" from the stored voice frames are represented by X ₁ to X _1.
j (j <i). Extracting a frame to be registered as a feature of the "telephone responder 'from the frame of the X ₁ to X _j. FIG. 7 illustrates this. First, compared with the standard pattern distribution SH and frame X ₁ to X _j of "H", the most probability to select a higher frame (FIG. 7 (1) -
(J)). In this case frame X ₂ showed the highest probability, to adopt it. Then, the part extracted as this feature is extended so that the highest probability is output by the variance value based on the specific speaker recognition (see FIG. 8A),
It is registered as a new standard pattern (see FIG. 8B). During this call, voice recognition is performed using this pattern. By performing this for all the phoneme segments necessary for recognizing “high”, during this call, voice recognition can be performed according to the “high” feature of the “phone responder”. it can. When the call ends, that is, when the "phone responder" hangs up, the pattern is returned to the original standard pattern. with this,
When contacting the next customer, speech recognition can be performed using the original standard pattern.

Also, in this power operation automatic communication system, it is necessary to call the person in charge of the electric equipment.
Of course, the telephone of the person in charge of the electrical equipment is often installed near the electrical equipment. In this case, there is a strong possibility that noise peculiar to the electric equipment is input from the telephone handset. When this noise is large, it may be treated as a voice section. Such a possible sound includes a sudden sound such as an operation sound of a circuit breaker, and this sound may be treated as a plosive sound. In such a case, misrecognition is likely to occur. In the power operation system of the embodiment,
Such a sound peculiar to power equipment is treated as voice, and then removed as noise.

This will be described with reference to FIGS. Now, in this case, a sound or the like peculiar to the power equipment to be treated as a voice is collected and registered in the system as a standard pattern as shown in FIG. In FIG.
The collected data is prepared (S901) and subjected to frequency analysis and the like to create a matching pattern (S90).
2). Abnormal data is removed from the matching pattern (S903), and a representative pattern is created (S905).
The clustering of a plurality of representative patterns is performed to determine a standard pattern (S906).

Next, with reference to the flow shown in FIG. 10, a description will be given of how to remove a sound or the like peculiar to power equipment to be handled as a predetermined sound by voice recognition.

In FIG. 10, steps up to S1006 are ordinary speech recognition procedures, and a description thereof will be omitted. In step S1006, the recognition result is output. In this embodiment, the recognition result may be output not only by voice but also by operation sound peculiar to electric power. In this case, the sound is recognized as noise, and in the subsequent processing, this sound is treated as not being recognized. In this way, the speech processing includes noise elimination, so that the distinction between speech and noise is made reliably and easily.

This processing can be applied to not only noise peculiar to power equipment during power operation but also noise peculiar to a telephone installation location. For example, there is noise from a train or the like at a railway station.

[0039]

As described above , according to the present invention , a call can be made.
Single word identified first by responder at start
Specific words that are recognized using standard patterns
Recognition means and a specific word recognized by the specific word recognition means.
Word pattern acquisition to acquire words as specific word patterns
Obtaining means and a specific word obtained by the word pattern obtaining means
Using a pattern to rewrite the standard pattern
Standard pattern changing means for changing to the standard pattern
The mark rewritten by the standard pattern changing means
Standard pattern restoration that restores the standard pattern to the original standard pattern
Means after receiving an incoming call (hook-off).
The first uttered voice during a call is used as training data.
Can just perform speaker adaptation in real time
And high speech recognition rate for specific words (most important words)
Can Mel, improve the overall reliability of the telephone answering system
Can be done.

Also, with the end of the call (hook on)
At last, the operation of the telephone answering system was finished and rewritten.
The standard pattern can be restored to the original standard pattern.
The new standard pattern for new customers.
Voice recognition.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a device configuration of an automatic communication system according to an embodiment of the present invention.

FIG. 2 is a block diagram illustrating a functional configuration of a voice recognition response device and a system operation terminal according to an embodiment of the present invention.

FIG. 3 is a flowchart illustrating a message flow of the automatic power management communication system according to the embodiment of the present invention.

FIG. 4 is a diagram showing an input voice and a standard pattern.

FIG. 5 is a diagram illustrating speech recognition.

FIG. 6 is a diagram illustrating a stochastic automaton used for speech recognition.

FIG. 7 is a diagram for explaining selection of a frame indicating characteristics of a telephone correspondent;

FIG. 8 is a diagram for describing rewriting of a standard pattern.

FIG. 9 is a flowchart illustrating creation of a standard pattern of noise.

FIG. 10 is a flowchart illustrating how noise is removed by speech recognition.

[Explanation of symbols]

 1, 2 voice recognition response system 3 Ethernet 4 other telephones 5 private branch exchange (PBX) 6 public line 7 system operation terminal

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ identification code FI G10L 21/02 G10L 3/02 301D H04M 3/493 (58) Investigated field (Int.Cl. ⁷ , DB name) H04M 3 / 42-3/58 G10L 3/00

Claims (2)

(57) [Claims] 1. Speech recognition using a standard pattern, and recognition
In a telephone answering system that sends out a message corresponding to a given voice, the responder first speaks at the beginning of the call
Using the identified single word as the target word, the standard pattern
A specific word recognition means for recognizing with, the specific word recognition
The specific word recognized by the means to a specific word pattern
A word pattern acquisition means for acquiring as down, using a specific word pattern obtained by the word pattern <br/> over emissions obtaining means,
Rewrite the standard pattern and change to a new standard pattern
Standard pattern changing means for changing the standard
Standard pattern rewritten by pattern changing means
Bei a standard pattern restoration means the return to the original reference pattern
Telephone response system, characterized in that was example. 2. The telephone answering system according to claim 1, wherein the target word is “high”.