JPH11352986A - Recognition error moderating method of device utilizing voice recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enhance the voice recognition of a voice dial PHS portable telephone and make it easy to use. SOLUTION: In this recognition error moderating method, when an opposite name 'Meguro' is inputted by voice (S2), similarity to a standard pattern is calculated (S3). When similarity does not exceed a regulated value (S4), resounding is urged by a display or synthesized voice with a different content every time in the form of conversation as 'Pardon. I was thinking about another thing. What?', and the process is returned to step S2 (S6). When the similarity exceeds the regulated value, 'Meguro' 03-1234-5678 is displayed, and automatic transmission to this telephone number is performed. A standard pattern is speaker-adapted by the use of the input voice of the same user.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for recognizing an input voice such as a voice dial and performing a process based on the result of the recognition to reduce erroneous recognition.

[0002]

2. Description of the Related Art In recent years, a voice recognition function has attracted a great deal of attention from the viewpoint of supporting behavior of a user (user). A number of products that use this speech recognition function have been produced, and many conference presentations have been made regarding them. For example, Company A has released a speech recognition word processor software named VOICE TYPE, and Company B has released a car navigation system having a voice recognition function as a product. In addition, the present inventors have provided a wristwatch-type PHS telephone with a voice recognition function capable of dialing by voice to the operating staff of the Nagano Olympic Games held in February 1998 in newspapers, magazines or television. It has been widely introduced in broadcasts and the like.

[0003] As is well known, the voice recognition function allows a user to input various commands (names) and names to the target system by voice, that is, a user who uses both hands for other purposes. Was regarded as a very preferable input means for the user. For example, a driver who is a user of a car navigation system needs to concentrate on driving a car, and his hands and feet are hesitant to operate steering wheels and brakes. For this reason, it has been attempted to use a voice recognition function as a means for switching screens of a car navigation system or extracting information on a distance to a destination. In addition, a voice recognition function is utilized in mobile phones and the like as a means of alleviating the inconvenience of small dial button operability. For example, although a mobile phone of Company C needs to be pre-recorded in real voice, it is possible to display the telephone number of the other party by dialing out the name and dial the number.

[0004] Despite its convenience, products and the like with a voice recognition function typified by the above-mentioned cases have not yet gained more users. The reason is that car navigation operation by voice while driving a car causes driving noise and engine sound of the car, and dial operation by voice on the street causes street noise and surrounding conversation,
This is because it is difficult to realize a speech recognition rate of 100%. As a means for systematically reducing a speech recognition rate other than 100%, a method of confirming the result of speech recognition to a user is used. For example, the following specific example is shown.

[0005] User: activates voice recognition function (S1)
Speech "Aoki-san" (S2) The speech recognition function calculates the similarity between the utterance and a previously stored standard speech (S3). This value is assumed to be, for example, 60 (a typical numerical value with 100 as an upper limit). It is determined whether or not the calculated similarity exceeds a predetermined value (threshold) (S4). If it exceeds the threshold, the terminal: displays the recognition result on the liquid crystal display screen [Aokisan 0
(3-5234-5678) "(S5) User: Press dial dial button Terminal: Send dial tone to telephone line Generally, the flow is similar, and the similarity of recognition becomes a predetermined value in step S4. If not exceeded, the terminal sends a fixed phrase "Please speak again" on the liquid crystal screen or the like, or utters the synthesized speech, and returns to step S2 (S6). If the recognition result display incorrectly recognized "Aokisan" as "Aoyamasan" in step S5, it was necessary to repeat the utterance. These steps
It was almost the same when the driver operated a car navigation system. In summary, speech recognition features
As a result of not being able to achieve a recognition rate of 100%, the confirmation operation of the user (driver) is unavoidable, which leads to an evaluation of being troublesome, and has not been widely accepted in the market.

[0006] In addition, toys using a voice recognition function are known. For example, a few years ago, a small dog toy having a kind of voice recognition function was released. This means that when a person is detected by the infrared sensor, it makes a cry,
It was a toy making a spoiled voice. Detecting the human body with an infrared sensor is not limited to toys, but is a widely used technique. The feature of the toy in this case is that it is a means of recognizing human voices at low cost, and detects the frequency components of consonants contained in the sound of "shi" emitted by humans with a kind of wavelength filter. . Not surprisingly, the response was uniform and did not reach high popularity.

[0007] Further, although slightly different from recognition using voice, a computer has appeared 10 or more years ago as if it had intelligence. This means that when a person enters a sentence such as "Today was a difficult time" from the keyboard, the answer "What do you want to do?" Is displayed on the screen, and the computer (or the other side of the line) Has been talked about as recognizing words (in this case, sentences entered from the keyboard). This device is a computer with built-in software that creates untouchable sentences based on keywords in sentences entered by the other party,
It was obvious that constructive responses were difficult. This, like other cases, was forgotten at times.

[0008] In addition, a so-called electronic pet, which is hardly an intellect, but takes care of its owner with programmed movements in a small game machine that fits in a pocket, existed before this invention. Electronic pets are a type of game, and the owner performs “care: feeding, cleaning, disciplining, etc.”
They grew up and lost their health. It is a well-known fact that more than several dozen types of these have been released, but as the game's “finish” method has become known, it has become tired.

[0009]

SUMMARY OF THE INVENTION The object of the present invention is distinguished from the cases enumerated so far by providing a speech recognition function, which is one of the recognition techniques, and is inevitable by the speech recognition function. It is another object of the present invention to provide a method and a system having a function of covering "misrecognition" without discomfort.

[0010]

According to the present invention, there is provided a voice recognition function, and a display or guidance (voice) prompting the user to re-produce a speech that is presumed to be erroneously recognized or is difficult to recognize. (Synthesis utterance) so that the user does not open the system to be dealt with, and the guidance is not continuously given the same contents in order to prevent the system from opening. It is.

In addition, the user's voice is input many times, so that the voice recognition function can reach a sufficient recognition rate in a relatively short time by so-called speaker adaptation. become.

[0012]

FIG. 1 shows an embodiment of the present invention. Here, the main body 1 is a communication terminal such as a telephone, a game machine, a so-called remote controller, an electric appliance at home, or a combination thereof, or a combination thereof. In this case, the main unit 1
Is described as a PHS telephone which is a kind of a mobile telephone, but it is apparent that the present invention can be applied to the various systems, terminals and products mentioned above. By the way, PHS telephone 1
Is a system basic function unit 2 that realizes a basic function of a telephone, an RF unit 3 that realizes a wireless function, an interface unit 4 that provides a user with a communication interface function, and a voice recognition that enables an operation based on a user's voice. Unit 5 and the like. Further, the interface unit 4 controls a microphone 41, a speaker 42, a screen display unit 43 such as a liquid crystal for displaying dialed numbers, a key operation unit 44 (not shown), and a key operation unit 44, which are indispensable for transmitting and receiving voice. It comprises an interface control unit 45 and the like. In addition, if necessary, the voice guidance unit 4
6 (not shown), or a memory 47 (not shown) for storing a telephone number and guidance voice are added. The part characterizing the present invention shown in the center of FIG.

Here, a schematic configuration of a PHS telephone system, which is one of application examples of the present invention, will be briefly described with reference to FIG. Since the shape in FIG. 2 is expressed with priority on intelligibility, there is no problem in adopting any other configuration or shape. In FIG. 2, there is an ISDN 71 as a so-called network.
So-called base stations 721 and 722 connected to DN71
And a service interface unit 73 for managing these and providing a telephone service, and a terminal 741 owned by the user.
And a terminal 742 (not shown) constitute a PHS telephone system, and provide a microwave band wireless communication service (hereinafter referred to as a PHS service) described below. Of course, not only PHS mutual communication services, but also other networks 75 (not shown), for example, PSTN or
It can be connected to or connected to the Internet or the like, and can also communicate with general home phones.

[0014] The PHS service uses a radio wave of 1.9 GHz band and transmits a radio signal of 10 mW or less (hereinafter referred to as RF) from a terminal.
The output is 50 from the public base station side (721, 722, etc.).
This is a personal use communication service that transmits and receives voice or digital data with an output of 0 mW or less (the output value is controlled according to the intended coverage of the base station). P
In the HS telephone system, a transmission / reception time slot (625 μs / slot) is allocated in a unit time of 5 ms called a TDMA / TDD frame, and three base station voice channels are provided for one base station. . A channel for controlling this voice channel: a control channel is provided between one base station and three terminals. Such a low-power PHS telephone can be made small, but if it is made small, the key operation unit must be small, and the size of human fingers and hands is limited. was there. One way to overcome such a restriction of miniaturization is to provide a dial transmission function by a voice recognition technology realized by the inventors. In this method, a human voice (number or name) is replaced with a dialed number using a microphone and a voice recognition function instead of a key operation unit normally provided in a telephone, and is transmitted to a telephone line. It is a well-known fact that the provision of this function has made it possible to reduce the size of a PHS telephone to the size of a wristwatch. Forty-four wristwatch-type PHS telephones were operated at the Nagano Olympics venue held in February 1998. Used.

Here, a specific configuration of the speech recognition function actually used to verify the effect of the present invention will be described. In FIG. 3, an audio input unit 11 is a unit for receiving audio, and is, for example, an audio microphone or an analog signal input terminal for receiving acoustic waveform data. The waveform converter 12 is means for converting voice data obtained from the voice input unit 11 into digital numerical values for analysis. The waveform converter 12 converts, for example, an analog voice waveform into digital data. The voice feature extraction unit 13 is a part for extracting a feature amount for voice section detection and voice recognition from the voice waveform data obtained by the waveform conversion unit 12, and a DP described in this description as an example.
In the matching method (so-called specific speaker speech recognition method),
For example, an analysis method well known in acoustic recognition technology such as a short-time logarithmic power analysis and a cepstrum analysis is used. In addition, as a matter of course, an experiment was conducted on an analysis procedure for speech recognition based on a hidden Markov model (a so-called speaker-independent speaker recognition method) and the use of an equivalent method, and a good result was obtained. In recognition for Japanese,
The unspecified speaker speech recognition method based on the Hidden Markov Model yielded good results, and in cases other than Japanese, the specific speaker speech recognition method of DP matching yielded good results.

The procedure described above is based on the case where the speech waveform is 300 to 3
Similar results were obtained from input from a telephone handset limited to a 400 Hz bandwidth. Voice section detector 1
5 determines a part to store the audio feature amount obtained from the audio feature extraction unit 13. The input pattern storage unit 16 is a storage unit that fetches a voice feature amount with emphasis on a vowel in a section from the voice start point to the voice end determined by the voice section detection unit 15 and sets the input voice feature quantity as an unknown input pattern. The standard pattern storage unit 17 stores a standard pattern for collating with the unknown input pattern stored in the input pattern storage unit 16.
It goes without saying that the standard pattern storage unit 17 has different storage contents depending on the applied speech recognition method. For example, in the case of DP matching, which is frequently used for a specific speaker recognition method that can also support foreign languages, the feature patterns of voices registered in advance are those obtained by converting the real voice of the recognition target person into a feature pattern. is there. In the case of speech recognition using a hidden Markov model that is frequently used for recognition of an unspecified speaker, the speech is divided into units called phonemes, and usually includes 43 elements,
Alternatively, 26 elements or the like are stored as so-called vector information. In this case, unlike the above-mentioned DP matching, it does not have a valid meaning by itself, so the system needs to be equipped with a hidden Markov model network that models the recognition target. This is specified as the recognition target model storage unit 24. Of course, these are notational expedients, and it is also possible to collectively write 17 and 24 in the same block (hereinafter, a part storing a standard pattern to be compared with an input pattern in both or a part thereof). The standard pattern storage unit 700). To put it simply, the standard pattern storage unit 700 is a storage unit that stores and analyzes a plurality of voice standard patterns for recognition with label names, which are analyzed and stored in the same procedure as the input pattern storage unit. is there.

The likelihood calculating section 22 stores the input pattern storing section 1
6 is compared with a plurality of standard patterns generated or output by the standard pattern storage unit 700. The similarity is defined as a distance value (a distance value on a feature amount defined by a mathematical expression such as a Mahalanobis distance) between the input voice pattern and the standard pattern. Alternatively, it is defined as a probability of generating an actually observed input pattern assuming that a standard pattern is uttered (however, a logarithmic value of this probability is generally used). In the former case, it is determined that the one with the smallest distance value has the highest likelihood in the latter case. It is an extremely practical configuration that the threshold value of the likelihood calculation unit 22 can be set from the outside. This is clearly shown in FIG. By changing the setting of the likelihood, it becomes possible to adjust the sensitivity of the system to noise, or to adopt a system configuration in which a plurality of candidates having extremely similarities are simultaneously selected. In addition, a system (not shown) for recording the similarity when the output is in accordance with the intention of the speaker and a system configuration connecting the same to the likelihood threshold setting unit 25 make this possible. The system can be provided with a learning function. A learning function or a speaker adaptation technique that makes it easier to recognize the voice of a target speaker is known as a general function improvement method for speech recognition.

The likelihood comparing section 23 receives the result of the likelihood calculation and determines which standard pattern the input voice is most similar to. At this time, when there are a plurality of standard patterns having similar similarities, all of them are output to the result totaling unit 19 when all the standard patterns are similar, and when there is only one similar standard pattern. The result totaling unit 19 arranges the recognition results in order of likelihood for each standard pattern and sends the result to the output unit 20. At this time, whether or not the likelihood value is also sent depends on the usage of this system. In addition, the result totaling unit 19 passes processing to the user specializing unit 6 when a sufficient threshold value cannot be obtained for the pattern determined to have the highest similarity,
The characteristic procedure of the present invention, that is, the display and the utterance (by the synthetic voice) that make the speaker feel smooth are exhausted. This will be specifically described below.

In short, the user specializing section 6 captures the user's habits and characteristics through the interface control section 45 of the main body 1 and assists the user in inputting more accurately the operation of the user. is there. In addition, this is a so-called speaker adaptation mechanism in which the user's utterance habit and tone are recorded as data, and the data in the standard pattern storage unit 700 in FIG. further,
If the original function is not performed due to ambient noise, etc.,
Rather than giving instructions such as "please utter again" in the form of a crest, it is intended to alleviate the user's discomfort by giving a response that considers anthropomorphism or petness. Hereinafter, effects and configurations of the present invention will be described based on the case of a PHS telephone with a voice recognition function.

FIG. 4 shows an operation of uttering “Meguro” using a voice recognition function built in the PHS telephone and calling a dialed number registered as “Meguro” (referred to as voice repertoire dial or voice dial, etc.). 2) shows the procedure flow. The part to be emphasized in this case is the case where the speech recognition function is erroneously recognized. As is well known, misrecognition occurs when the start or end of a user's utterance cannot be accurately identified, or when another person's voice or surrounding noise is superimposed during the utterance, and other than the original voice waveform. Occurs when information is taken in. In such a case, it is known that the similarity (also referred to as likelihood) between the captured speech waveform and the recognition candidate is smaller than the value that should originally be possessed. In such a case, the system that realizes the conventional voice recognition function is to raise a candidate having the highest value among the small similarities as a recognition result or to return a certain response such as "cannot be recognized". .

For example, "Meguro (me / gu / ro)"
When the sound of “gu” in the middle is mixed with the surrounding noise, the information “me / ** / ro” is obtained. In this case, although the two place names “Meguro” and “Meguro” are smaller than the value that should be originally expected, they have similar degrees of similarity as compared with the case where “meguro” is uttered. As a result, depending on the tone of the utterance, "mejiro" was returned as a result with a considerable probability, resulting in erroneous recognition. According to the present invention, in such a case,
Focusing on taking a value in which the similarity between the first candidate and the second candidate is extremely close, there is a great feature in returning another response. In the operation procedure of FIG. 4, first, the voice recognition function is activated (S1), "Meguro" is uttered (S2), and the first utterance is received to calculate the similarity value (S3). It is determined whether or not it has exceeded (S4),
It is determined whether or not the difference between the second candidate and the first candidate exceeds a predetermined value. When these values do not satisfy the predetermined similarity or when the difference between the first and second candidates is small, the first candidate is not returned as a recognition result.
One of the features is that the PHS phone returns a response as if it were willing, rather than returning a specific dry and dry response that is common in conventional systems. It is. In this example, PHS
In this case, a natural voice is urged in the form of "I'm sorry, I was thinking about something else." In this conversation, various patterns are recorded in the user specializing section 6, and they are displayed (in the case of a screen) or uttered (in the case of a speaker output) as necessary, and then the same response is made. By not doing so, the personification can be enhanced. In other words, a different word (sentence) prompts a resound.

In addition to such a response, it is a preliminary experiment to incorporate a service specific to the PHS telephone into the conversation to further anthropomorphize or enhance the pet property and increase the degree of attachment to the terminal as seen by the user. Was confirmed. This is shown in FIG.
The latter half of the flow. In this example, if the PHS telephone is located in Yokohama by the PHS location information service, for example, when the condition is set in conjunction with the location information service of the PHS telephone and Yokohama is set as a condition, the condition is set. Then, [Meguro-san, "Tell me that Yokohama is hot today,""Justconnect"] is displayed on the display (S8), and the user knows that "Meguro" has been correctly recognized. Then, a fixed phrase appears after Yokohama, and it can be recognized that voice dialing is performed with "Ja, Tsunagune". In this way, if you set up a number of condition settings,
Different conditions are established for each use, and sentences and guidance appear according to the conditions. Then, the user is interested in what will appear, and the user will be attached to the use (including reoccurrence). As a result, re-speaking by the same user and voice input for use are performed for many days, adaptation of the standard pattern to the voice of the user progresses, and the voice recognition rate of the user increases. In addition, when the speech recognition rate is increased as described above, a device using this kind of speech recognition becomes easier and easier to use.

As described above, based on the location information service,
In the above example, the position of Yokohama is set as a condition, and if this is satisfied, the fixed response is created as described above. However, other various conditions are set, and if it is satisfied, a response sentence corresponding to the condition is created. You can also For example, based on information obtained by linking with a weather forecast or proprietary information provided by a PHS telephone service provider, a structure is adopted in which the location and climate of the terminal are put into a conversation, thereby reducing the inorganicity of the terminal. be able to. These are just a few examples. Some of the phone numbers that users often call will display a message such as "Is this an ordinary person?" It is one Example of this invention. As an example of such a condition, by using personal information registered by the user, for example, information such as a name, a date of birth, and a place of residence, a response sentence (including a guide sentence) corresponding to the user using the main body 1 is used. ) Is displayed or uttered to increase the attachment to the main body (processing device) 1. For example, when making the first call on the birthday of the user, on the condition of the date of birth, say or display "Happy Birthday XX" and then say "Let's call XX, leave it to me." Have a highlighted display or utterance. Conversations and displays that bring these anthropomorphisms and pets can be instantly tiring with fixed phrases, but using the data transfer function of the PHS phone,
The HS telephone service provider can send various responses or conversations to the user specializing unit 6 in FIG. As a matter of course, it is conceivable that this data transfer is performed by a PHS telephone service provider in order to avoid the burden of the user's call charge. Also, of course,
It is desirable to use a so-called non-ringing incoming call so that the user does not answer a call necessary for data transfer as a call from an acquaintance.

The feature of the present invention described in the above case is that when it is difficult to recognize the utterance of the user or when there is a high possibility of erroneous recognition, a person-like response is returned to return the voice of the user. Clearly, performance dissatisfaction can be mitigated.
In addition, as a result of such a re-utterance being made by the user, features of the user's utterances are gathered, and the recognition rate is improved by subjecting the standard pattern to speaker adaptation processing. This is the result of having the system learn user utterances.

The user specializing unit 6 characterizing the present invention
Is a so-called arithmetic processing function to achieve its purpose,
It is necessary to have a memory function. The arithmetic processing function and the memory function are generally included in the system basic function unit or the voice recognition unit of the PHS telephone, and by using these, it is possible to reduce the cost of the entire system. Utilize a digital signal processor (DSP: not shown) for performing a voice recognition function or a CPU (not shown) for controlling a telephone function, or use a processing capability remaining after the functions of the two have been performed. Performed no problem and performed a predetermined function.
On the other hand, having an arithmetic processing function and a memory dedicated to the user specialization unit instead of diversion of these components is effective in avoiding interference and memory conflict between the user specialization unit 6 and the system basic function unit. It was suitable for the form of providing more advanced responses and conversations attempted in the application examples of the present invention.

Here, an example of application to a device other than a telephone will be described. Since it is incomparable to send the previous conversation sentence or response sentence by communication, it is inevitable that expressions that can be answered will be interposed. However, a great effect can be brought about by the presence of the user specializing unit 6. The recognition rate of the conventional remote control with voice recognition function is at a practical level (generally 90
% Is said to have been reached). One is the lack of perfection of speech recognition programs, which is a well-known fact that it has improved significantly in recent years. Another is the fact that the user was not accustomed to making utterances suitable for the speech recognition function. This has been caused by the irregular timing of the utterance and the habit of the user to add an unnecessary word such as “Eh, ah”. Forcibly recognizing utterances that are difficult to recognize,
The rate of misrecognition increased, and eventually the user had the impression that it was "useless". On the other hand, the present invention is characterized in that a humorous re-utterance procedure that expresses anthropomorphism or petness is used for utterances that are likely to be erroneously recognized. For example, there is a case where ambient noise is high and the similarity of voice calculated after utterance does not reach a predetermined value. In such a case, the subject said, "I'm sorry, the surroundings were too loud to hear.
Please tell me again. "
The effect was that the aversion to the repetition of the remote control was reduced.

[0027] When the resistance to the re-speaking is reduced in this manner, the re-speaking from the user can be obtained, and more data specific to the user's voice can be accumulated. As a result, This has the effect of increasing the recognition rate of the user's utterance. To do this, when a recognition result is accepted by the user, it is compared with the utterance content for recognition that was repeated immediately before. In the process, the user analyzes voice recognition problems, such as whether the user easily utters unnecessary words, the timing of the utterance is shifted, the ambient noise is high, etc. By making a change (also called learning) and transferring data or a program from the user specializing unit 6 to the voice recognizing unit 5 in the configuration of FIG. 1, a voice recognizing function suitable for user utterance can be achieved. It is.

[0028]

By using the configuration of the present invention, a user can use a communication terminal such as a telephone, a game machine, a so-called remote controller, an electric appliance at home, or a combination of these, or a combination of these. It is possible to increase the recognition rate of the voice recognition function while enjoying a processing device having a voice recognition function such as an object. In addition, by promoting anthropomorphization and petization specialized for user's utterance, attachment to a processing device such as a target terminal or product is increased, and an effect of extending a use period is produced.

In particular, in telephones and the like having a communication function, personality and petness can be further enhanced by appropriately rewriting the built-in data, displaying a response message or uttering according to the season or place, and thus attaching to the user. It is possible to improve. As a result, this can also be an effective countermeasure for PHS phones, etc., which will be canceled immediately.

[Brief description of the drawings]

FIG. 1 is a schematic diagram showing a configuration of the present invention.

FIG. 2 shows a PHS with a voice recognition function to which the present invention is applied.
FIG. 3 is a schematic diagram of a functional configuration of a telephone.

FIG. 3 is a diagram showing one embodiment of a voice recognition function which is one of the components of the present invention.

FIG. 4 is a flowchart showing an example of a procedure for using the voice recognition function of the present invention.

FIG. 5 is a flowchart showing an example of a procedure for using a conventional voice recognition function.

 ────────────────────────────────────────────────── ─── Continued on the front page (72) Inventor Yutaka Nishino 3-19-2 Nishi-Shinjuku, Shinjuku-ku, Tokyo Nippon Telegraph and Telephone Corporation

Claims (4)

[Claims] An input voice is recognized by a voice recognition function.
In a device that performs processing based on the recognition result, the similarity is compared with a threshold value in the recognition processing for the input voice, and if the result of the comparison determines that the recognition is erroneous recognition or difficult, the apparatus compares with the previous time. An erroneous recognition mitigation method for a speech recognition utilizing apparatus, characterized by prompting reoccurrence by display of different contents or voice. 2. An input voice is recognized by a voice recognition function.
An apparatus for performing a response process based on the recognition result, wherein a response based on the recognition result is displayed or uttered according to a user according to a user. . 3. A method according to claim 1, wherein the input speech is recognized by a speech recognition function and a response process is performed based on the result of the recognition. Determining whether or not a condition is satisfied; reading out the fixed phrase according to the condition if the condition is satisfied, generating a response sentence, and displaying or uttering the response sentence; How to mitigate false recognition. 4. A method according to claim 1, wherein said apparatus is provided with a data transmission function, and said data required for said display or utterance is externally transferred. A method for reducing false recognition of a recognition and utilization device.