JP2003029776A - Voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognition device which keeps a recognition performance close to the performance of a specific speaker system and is enhanced in convenience by automatically conducting the training. SOLUTION: When recognition words are registered, text information of the recognition words is inputted into a voice synthesis processing section 15, synthesized voice signals that are to be outputted are converted into synthesized sound acoustic data by a synthesized voice sound converting section 19. The data are registered in a word acoustic data storage section 21 in place of the uttering of a speaker in a conventional training. The coincident process for acoustic data in a word recognition section 26 in the voice recognition process, that is conducted after the registration, is conducted for the synthesized sound acoustic data. Thus, the voice recognition device for a specific speaker is realized at a low cost and the convenience which is similar to the convenience of an unspecified speaker system voice recognition device that does not required training can be provided.

Description

Detailed Description of the Invention

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device, and in a voice recognition technique for a specific speaker, a voice recognition device characterized by performing voice recognition without a training procedure by the speaker. .

2. Description of the Related Art In recent years, in information processing devices such as telephones, fax machines, and car navigation systems, so-called voice recognition technology has been applied to commercialize devices that can be operated by voice input. The speech recognition technology has a speaker-independent method (speaker
independent) and a specific speaker method (speak
er dependent). The unspecified speaker system extracts linguistic features contained in speech and applies pattern recognition technology typified by neural networks to estimate the utterance content of the speaker. However, the speech uttered by the speaker has a voice quality peculiar to each individual, and complicated processing is required to secure a stable recognition rate for an unspecified speaker. As a result, the cost of the product increases. On the other hand, the specific speaker method obtains a good speech recognition rate in an inexpensive system by limiting the speakers who can handle it. In this method, it is necessary to register (train) the voice quality of the speaker himself when the device is first used.
It takes time and effort. In the speech recognition process, the basic operation is to identify the word corresponding to the word uttered by the speaker from the words stored in advance in the database in the form of a database and return the result to the speaker. Becomes A general operation of a conventional specific speaker type voice recognition device will be described below with reference to the drawings. 9 is a block diagram of a conventional specific speaker type speech recognition device, FIG. 10 is a detailed diagram of a speech recognition processing unit in FIG. 9, and FIG. 11 is a detailed diagram of a word acoustic data storage unit in FIG. The word uttered by the speaker is converted into an electric signal by the microphone 1, converted into a voice signal in a format suitable for the subsequent processing by the signal processing unit 2, and sent to the voice recognition processing unit 4. The sound processing unit 6 in the voice recognition processing unit 4 extracts the acoustic feature amount from the sound signal, and the word identification unit 8 holds the one that most matches the input sound data in the word sound data storage unit 7. Find out from the sound data that is present. As a result, the word identifier associated with the matched acoustic data is used as the identification information by the signal processing unit 2.
Then, the signal processing unit 2 can recognize the word uttered by the speaker and perform appropriate processing control. The above is the flow of the basic recognition processing common to the speaker-unspecified and speaker-specific voice recognition methods, but the basic difference between both methods is the method of generating word acoustic data in the word acoustic data storage unit 7. It is in. As described above, in the speaker identification method, word sound data is generated by training. Therefore, in the initial state of the device, the word acoustic data is in an undefined state, and this training is indispensable before the voice recognition processing. Training is a process in which a speaker utters all words to be recognized and registers them in the word acoustic data storage unit. In training,
A specific recognition target word uttered by the speaker is input by the microphone 1 and converted into a voice signal by the signal processing unit 2. At this time, a word identifier for distinguishing each recognition target word is added. In the voice recognition processing unit 4, this voice signal is converted into acoustic data by the acoustic processing unit 6 and supplied to the word acoustic data storage unit 7 together with the word identifier. The word sound data storage unit 7 stores the sound data and the word identifier in association with each other. In this way, the speech recognition becomes possible only by repeating the same training for all the speech recognition target words (see FIG. 7). On the other hand, in the speaker-independent method, it is not necessary to utter the speaker to create the word sound data in the word sound data storage unit, so it can be set in advance before the speaker's voice recognition operation. And there is no burden on the speaker. However, there is a drawback that it is difficult to realize with a small-scale speech recognition device because it requires complicated calculation to generate word acoustic data in order for the speaker to add a new recognized word. In this respect, in the speaker identification method, the generation of the acoustic data is easily realized by the acoustic processing unit 8 which is common to the voice recognition processing, so that the speaker can easily add a new recognition word even with a small-scale recognition device. There is an advantage that can be realized. The voice synthesis processing unit 3 in the figure is a processing unit for converting text into voice. Although it is not directly related to the voice recognition process, it is generally used together in a device having a voice recognition function.
Since the feature is that this voice synthesis function is positively used, it is also shown for the sake of explanation. The text to be output as a sound is sent as text information from the signal processing unit 2 to the voice synthesis processing unit 3, and the resultant synthesized voice signal is returned. This synthesized voice signal is output from the speaker 12 to be transmitted to the speaker as voice.

As described above, the specific speaker-type speech recognition apparatus requires training work, and in a system with many words to be recognized, the time required for training the speaker is long and The burden on the person has reduced the convenience of the system. On the other hand, in a speaker-unspecified system, it is difficult to add a new recognition word in a small-scale device typified by an embedded device, which limits the expandability and applicability of the system.
If the calculation ability of the device is increased to the extent that the recognition words can be added, there is a problem that the cost increases and it cannot be applied to a small-scale system.

In order to solve the above-described conventional problems, the present invention is based on a voice recognition device for a specific speaker that can be realized at low cost, and is generally used in combination with voice recognition. A voice recognition device characterized by using a voice synthesis function and using a voice signal generated therefrom for training instead of utterance by a speaker. By automatically performing the training that burdens the speaker inside the device, it is possible to provide the same convenience as that of an unspecified speaker-type speech recognition device without training.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The invention according to claim 1 of the present invention includes a microphone as a voice input device, a speaker as a voice output device, an input device such as a keyboard, and a display device for displaying a recognition result. , The microphone, the speaker,
A signal processing unit connected to the input device and the output device for performing processing control of the entire voice recognition device; a voice synthesis processing unit for receiving text information from the signal processing unit and outputting a synthesized voice signal; In a voice recognition device equipped with a voice recognition processing unit that compares a voice signal input from a processing unit with a plurality of acoustic data held therein and outputs the matching result to the signal processing as a result of voice recognition. By providing a synthetic voice conversion unit that converts the synthetic voice signal from the voice synthesis processing unit into synthetic voice acoustic data that is acoustic data, the synthetic voice acoustic data is generated by the training of the speaker. Instead of, by holding it in the voice recognition processing unit and using it for voice recognition,
It is characterized by not requiring training that burdens the speaker, and can provide the same convenience as that of an unspecified speaker-type speech recognition device without training. The invention according to claim 2 of the present invention is, in the speech recognition processing part of claim 1, acoustic data for selecting synthetic sound acoustic data output from a synthetic speech conversion part or acoustic data output from the acoustic processing part. By including the selection unit, the synthesized voice sound data is stored in the word sound data storage unit in the initial stage of the device, and when a certain recognition word is recognized, the sound data corresponding to the speech sound of the corresponding word is stored. The feature is that by replacing the synthetic voice and acoustic data, it is possible to recognize the acoustic data with the actual pronunciation from the first time onwards, and it is possible to prevent the voice recognition rate from decreasing even when the synthetic speech and the actual pronunciation are different. Have. According to a third aspect of the present invention, in the word acoustic data storage section of the second aspect,
Both of the synthetic sound acoustic data output from the synthetic speech conversion unit and the acoustic data of the utterance voice of the speaker output from the acoustic processing unit are held, and the case corresponds to any one of them. It is characterized by outputting the identification information of words, and has an effect that the voice recognition rate can be further improved. According to a fourth aspect of the present invention, in the word sound data storage section of the first aspect, there is a configuration in which a plurality of pronunciation methods for one word are input and stored from the synthetic speech conversion section. It is a feature, no matter which pronunciation the speaker makes,
This has the effect of correctly recognizing the corresponding word.
According to a fifth aspect of the present invention, in the word acoustic data storage section of the fourth aspect, the speaker utters a corresponding word,
It is characterized in that the matched acoustic data is left and the other acoustic data is deleted, and has an effect that unnecessary identification processing can be omitted from the next time and the recognition processing can be performed at higher speed. According to a sixth aspect of the present invention, in the word acoustic data storage section of the fifth aspect, a mechanism is added to hold the frequency at which the occurrence of the speaker coincides with each of a plurality of synthetic voice acoustic data for a certain word. . It is characterized by deleting only the acoustic data of the matching frequency below the threshold level when the corresponding word is recognized, and unnecessary recognition processing will be omitted from the next time, and recognition processing will be performed at high speed. It has the effect that According to a seventh aspect of the present invention, in the first aspect of the invention, when the synthesized sound data for each identification target word is held in the speech recognition processing unit, a synthesized sound is output to the speaker from the speaker. It is characterized by playing back, checking whether it is the synthesized sound intended by the speaker, and performing the training procedure by the speaker only when it is not intended, reducing the frequency of training by the speaker It has the effect of being able to. Hereinafter, embodiments of the present invention will be described with reference to the drawings. (Embodiment 1) FIG. 1 shows a basic configuration example of a voice recognition device to which the present invention is applied. FIG. 2 is an internal configuration diagram of the voice recognition processing unit 16 in FIG. 1, the basic configuration is the same as that of the conventional general specific speaker system speech recognition device of FIG. 9, but is characterized in that a synthetic speech conversion unit 19 is provided. In FIG. 1, reference numeral 13 is a microphone which is a voice input device and converts a voice uttered by a speaker into an electric signal. Reference numeral 18 is a speaker, and 23 is a display unit. 17
Is an input device for a speaker to input a key to confirm the recognition result and to control the entire device. Reference numeral 14 is a signal processing unit, 15 is a voice synthesis processing unit, 16 is a voice recognition processing unit, 1
Reference numeral 9 is a synthetic voice conversion unit. In the voice recognition processing unit 16 shown in FIG. 2, 20 is an acoustic processing unit for extracting acoustic feature quantities from the input voice signal, and 21 is the names of all the destinations in the telephone directory as recognition words. It is a word sound data storage unit that holds word sound data. 22
Is a word identification unit that searches the word sound data storage unit 21 for the best match with the input sound data. The word spoken by the speaker is converted into an electric signal by the microphone 13 and input to the signal processing unit 14. The signal processing unit 14 converts the input voice signal into a voice signal in a format suitable for the processing in the voice recognition processing unit 16. In the speech recognition processing unit 16 shown in FIG.
The acoustic feature amount is extracted from the voice signal output by the output device 4 and is output to the word identifying unit 22 as acoustic data. The word identification unit 22 searches the acoustic data stored in the word acoustic data storage unit 21 for the best match with the input acoustic data. As a result, the signal processing unit 1 uses the word identifier associated with the matched acoustic data as identification information.
Returned to 4. In FIG. 1, the signal processing unit 14 recognizes a word uttered by a speaker based on the identification information which is the result of the voice recognition, and based on that, performs appropriate processing control of the device, or the speaker via the display device 23. The recognition result is fed back to. Here, the voice recognition device of the present invention is applied to a telephone, and for voice telephone directory search processing in the telephone, registration processing of acoustic data in the word acoustic data storage unit by voice synthesis performed instead of training is performed. explain. In the telephone, the telephone directory is a kind of database that holds the name of the person with whom the speaker communicates and the personal information such as the telephone number and mail address of the person inside the telephone. By registering the information of the other party in this telephone directory, the speaker can easily make a call without entering the telephone number each time. A telephone with a voice recognition device is often used as a so-called voice dialing function, in which a speaker speaks the name of the other party to automatically retrieve the other party's phone number from the phone book and make a call. Many. In order to realize the voice dialing function, it is necessary to hold the names of all the destinations in the telephone directory as recognition words and store the respective word sound data in the word sound data storage unit 21. At this time, the name of the other party in text format or the entry number in the telephone directory is held as the word identifier. In the conventional voice recognition device for a specific speaker, since the speaker needs to utter the acoustic data, it is necessary to utter the names of all the other parties in the telephone directory. In the present invention, the process corresponding thereto is implicitly executed by the speaker using the speech synthesis processing unit 15 that generates a synthetic speech signal and the synthetic speech conversion unit 19 that converts the synthetic speech into acoustic data. As a specific example, consider the case where 100 names are newly registered in the telephone directory as follows. The data of this telephone directory is stored in the internal memory of the signal processing unit 14, but as described above, it is necessary to register the acoustic data in the word acoustic data storage unit 21 for voice dialing. There is. In the past, the speaker called the destination name "ad
When "am" and its telephone number "111-2222" are input by the key of the input device 23, these are processed by the signal processing unit 1.
Although registered in the entry 1 of the telephone directory in 4, the microphone 13 causes “adam” to be generated by the sound processing unit 20 to generate the sound data required as the input of the word sound data storage unit 21.
Is actually uttered. At the same time, the entry number “1” corresponding to this is associated with the acoustic data as a word identifier and registered in the word acoustic data storage unit 21. In addition to the time required for the speaker's utterance itself, the registration of the acoustic data accompanied by this utterance requires the time for utterance timing with the registration process, and it takes several seconds to several tens of seconds per case. Since the procedure had to be repeated 100 times, the load on the speaker was heavy. In the present invention, as in the conventional case, the signal processing unit 14 is provided immediately after key input of the destination name and telephone number.
The destination name “adam” is sent to the voice synthesis processing unit 15 as text information under the control of. Synthetic voice processing unit 15
Generates a synthetic voice signal "Adam" which is a standard vocalization sound corresponding to "adam". The synthetic speech signal is generated by the synthetic speech conversion unit 19 as synthetic sound acoustic data which is the feature amount data. The synthesized voice sound data is stored in the word sound data storage unit 21 in the voice recognition processing unit 16 together with the word identifier “1” from the signal processing unit 14. In this way, since a speaker does not utter, a series of processing for acoustic data registration is automatically performed within a few seconds.
The user repeats the key input of the destination name and the telephone number 100 times to complete the registration process of the acoustic data in the word acoustic data storage unit 21. In this case, unlike the acoustic data created by converting the voice from the microphone, the synthetic sound acoustic data is not affected by the noise around the microphone.
Although the user inputs the recipient's name and telephone number for the number of registrations, the user does not need to speak into the microphone for training, and there is no time required to speak with the registration process. Is less burdensome. Table 1 shows a configuration example of the telephone directory of the telephone of the present invention.

[Table 1] After the registration is completed, when the user wants to call "john", for example, the name of the other party "John" is input from the microphone 13. This voice is sent to the voice recognition processing unit 16 as a voice signal via the signal processing unit 14. Sound processing unit 20
Converts this "John" into acoustic data, and the word identification unit 22
Send to. The word identification unit 22 compares the acoustic data with 100 pieces of acoustic data stored in the word acoustic data storage unit, and returns the word identifier “3” of the matching entry as identification information to the signal processing unit 14. The signal processing unit 14 can search the internal telephone directory data and obtain the telephone number “123-3344” of the entry number “john” from the identification information “3”. Normally, the call is made after confirming the speaker for this telephone number. As described above, according to the present embodiment, even though it is a low-cost specific-speaker-type speech recognition device, the training that is burdensome to the speaker is automatically performed inside the device, so that apparent training As a result, it is possible to obtain the voice recognition device having the same convenience as the voice recognition device of the unspecified speaker system. (Embodiment 2) As described in Embodiment 1, by setting all the acoustic data in the word acoustic data storage unit as synthetic speech acoustic data, the speaker can be released from training. However, this method makes it difficult for the speaker's word to be recognized when the word spoken by the speaker is very different from the standard pronunciation generated by the speech synthesis conversion unit due to the influence of dialects and the like. It is possible that Embodiment 2
In order to solve this problem, the basic configuration is the same as that of the first embodiment, but by providing an acoustic data selection unit 27 in the speech recognition processing unit as shown in FIG. It is possible to select the acoustic data of the recognition word to be stored from the synthetic speech acoustic data or the acoustic data from the acoustic processing unit 24. The operation will be described below by taking voice dialing on a telephone similar to that of the first embodiment as an example.
In the present invention, when the acoustic data is registered in the word acoustic data storage unit 25 associated with 100 telephone directory registrations, the acoustic data selection unit 27 selects the synthetic sound acoustic data.
Therefore, when the registration is completed, the word acoustic data storage unit 2
The registration contents of No. 5 are the same as those of the form 1. However, consider a case where the synthesized voice signal for "Henry" of entry number 2 is "Henry" but the actual reading is "Henry". The ratio (so-called voice recognition rate) in which the word identification unit 26 correctly returns the entry number “2” as identification information to the signal processing unit for “Henri” issued by the speaker in voice dialing is stored in the word acoustic data storage unit 25. It will be smaller than when correct reading "Anri" is registered. In the present invention, in order to prevent a decrease in the voice recognition rate when there is a difference between such a synthesized voice and an actual vocal sound, based on the identification information from the word identification unit 26,
A mechanism is provided for replacing the corresponding sound data in the word sound data storage unit 25 with the sound data output by the sound processing unit 24. In this case, when the speaker utters "Henri" and the entry number "2" is returned from the word identifying unit 26 to the signal processing unit as identification information, the "Henri" held in the acoustic processing unit 24 is held. The acoustic data corresponding to “” is selected by the acoustic data selection 27, and at the same time, the word identifier “2” is input to the word acoustic data storage unit 25 from the signal processing unit.
From these data, the entry 2 of the word acoustic data storage section
Is replaced with acoustic data corresponding to "Henry". Since this process is implicitly executed for the speaker, the burden on the speaker does not occur. In this way, even if the actual voice is slightly different from the synthetic voice, the acoustic data for the identified word is constantly updated to the acoustic data corresponding to the utterance sound of the actual speaker. It is effective in preventing the permanent reduction of the voice recognition rate based on. (Third Embodiment) In the second embodiment, when the recognition is performed, the acoustic data of the corresponding word is replaced with the uttered acoustic data to prevent a decrease in the speech recognition rate when the synthesized speech is different from the actual pronunciation. It was intended. However, the voice registered as the replacement is not always appropriate. For example, when a plurality of speakers with different pronunciations share a voice recognition device, or when surrounding noise is mixed even in the same speaker's utterance, the acoustic data of a particular utterance is once the acoustic data of the relevant word. If it is registered in the word sound data storage unit as, it is conceivable that the voice recognition rate will decrease even if standard utterance is performed. In order to solve this problem in the third embodiment, as shown in FIG. 4, the word sound data storage unit is used to synthesize synthetic sound sound data 29 for each recognition target word.
The vocal sound data 30 can be held. This makes it possible to prevent a decrease in the voice recognition rate by matching with any one of the acoustic data even if the utterance is close to the standard synthetic voice or is not the standard utterance peculiar to the speaker. The operation will be described using a precedent of a telephone directory. Consider a case where the synthesized voice signal for "Henry" of entry number 2 is "Henry", but the actual reading is "Henry". First, in the word sound data storage unit, the entry number "2" is stored in the word identifier 28 when the telephone directory is registered, and "Henry" is stored in the synthesized sound sound data 29 related thereto. After this, recognition is successful due to the actual occurrence "Henri" of the speaker, and when the entry "2" is returned to the signal control unit as the identification information, "Henri" becomes the generated acoustic data 30 related to the entry "2". Is stored.
This series of processes is the same except that the rewriting of acoustic data does not occur in the second embodiment. In this way, two acoustic data for the word "Henry" in the telephone directory are held in the word acoustic data storage unit. In this way, in the next voice dialing, even if the speaker utters "Henri" or another speaker utters "Henry", the voice recognition processing is performed without lowering the voice recognition rate. The effect is that you can do it. The uttered sound data 30 is updated to the sound data corresponding to the utterance of the speaker each time the recognition of the entry “2” is successful, as in the second embodiment. (Embodiment 4) In Embodiment 1, only one piece of acoustic data can be registered for one recognized word. However, if it is known in advance that there are multiple pronunciations, such as "Henry" in the phonebook registration example,
The voice recognition rate can be improved by registering all of them when registering the phone book. The fourth embodiment has the same basic configuration as that of the first embodiment, but is provided with a word sound data storage unit capable of registering a plurality of synthetic sound sound data with respect to one recognized word as shown in FIG. As a feature, when the utterance of the speaker matches with any one of the synthetic sound and acoustic data, the corresponding word can be recognized. For example, "He
In the example of “nry”, the entry number “2” is added to the word identifier 31.
Is stored, and the synthetic sound acoustic data A32 related thereto stores acoustic data corresponding to “Henry”, and the synthetic sound acoustic data B33 stores synthetic sound acoustic data corresponding to “Henri”. These two synthetic sound and acoustic data are automatically generated in the synthetic speech conversion unit using general reading knowledge. Thus, even when the speaker utters "Henry" or "Henri", "Henry" is correctly recognized without lowering the recognition rate. (Fifth Embodiment) In the fourth embodiment, it is possible to improve the reduction of the recognition rate for different utterances by allowing a plurality of synthetic voice-acoustic data for one recognized word. However, there is a problem in that the speech recognition processing time is increased because the amount of acoustic data compared by the word identification unit increases during recognition. In the fifth embodiment, with respect to the synthetic sound acoustic data of each word in the word acoustic data storage unit, the synthetic sound acoustic data that best matches the utterance of the speaker among the recognized words is left, and the other synthetic sound acoustic data is deleted. It is characterized in that the unnecessary recognition process is omitted from the next recognition process and the recognition process is performed at a higher speed by including the mechanism. (Sixth Embodiment) In the fifth embodiment, all of the plurality of synthetic sound-acoustic data of the recognized word are deleted except for the best match. However, it is possible that one or more pronunciations occur at the same degree for a certain recognized word, and in this case, the speech recognition rate other than the remaining pronunciations becomes low. In the sixth embodiment, as means for solving this, in the word sound data storage unit of FIG. 6, the frequency information matched when the corresponding word is recognized for each of the plurality of synthetic sound sound data of each word is obtained. Add a recording mechanism. Only when this frequency information falls below a certain threshold level, the synthetic sound acoustic data is deleted.
In this way, by gradually deleting the synthetic voice acoustic data with low utterance frequency each time the voice is recognized, the synthetic voice acoustic data with the possibility of utterance is left, thereby preventing a decrease in the voice recognition rate and eliminating unnecessary. This has an effect of preventing an increase in voice recognition time due to acoustic data. The operation will be described using the example of "Henry" in the telephone directory. In the word sound data storage unit of FIG. 6, the entry number “2” is stored in the word identifier 35 at the time of registering in the telephone directory, and the synthetic sound sound data A related thereto is stored.
There are three possibilities of 36: acoustic data corresponding to "Henry", synthetic sound acoustic data B38: synthetic sound acoustic data corresponding to "Henri", synthetic sound acoustic data C40: synthetic sound acoustic data corresponding to "Henry". It is assumed that a certain pronunciation is stored. At the same time, at the time of this registration, the initial value 10 is the frequency information A3 as the frequency information of each synthetic sound acoustic data.
7, frequency information B39 and frequency information C41. First, in the voice recognition, when the speaker utters "Henry" and "Henry" is recognized, the numerical value of the frequency information A corresponding to "Henry" is incremented by 1, but the maximum value does not exceed the initial value 10. Therefore, the value remains 10, and the frequency information other than “Henry”, that is, the numerical values of the frequency information B and the frequency information C are decremented by 1, and become 9 respectively. When the recognition of “Henry” continues seven times with the pronunciation of “Henry”, the frequency information A, B, and C becomes 10, 2, and 2, respectively. Next, when recognition by "Henri" occurs, the frequency information A, B, C becomes 9, 3, 1 respectively. Next, when recognition by "Henry" occurs, frequency information A, B,
C becomes 10, 2 and 0, respectively. At this time, the frequency information corresponding to "Henri" falls below the threshold level 1, and the voice sound acoustic data C40 is deleted. After that, the comparison process for "Henri" is not performed. In this way, only the synthetic sound acoustic data having a low utterance frequency is deleted as the speech recognition progresses. (Embodiment 7) In Embodiment 1, synthetic sound acoustic data is registered as acoustic data of a recognized word, but if the actual reading of the word by the speaker and the synthetic speech are completely different, the corresponding word is recognized. There is a problem that it becomes difficult. In order to solve this problem, the configuration of the invention is similar to that of the second embodiment, but when the synthesized speech signal to be registered is generated by the speech synthesis processing unit, the signal is input to the synthesized speech conversion unit and at the same time the signal processing unit is input. In addition, the signal processing unit is provided with a mechanism for obtaining confirmation input from the speaker through the input device after feeding back to the speaker using the speaker. As a result, the speaker can give O when the synthesized speech is close to the intended reading.
K is input for confirmation, and if the synthetic speech is completely different from the intended reading, NG is input for confirmation. If the confirmation input is NG, the signal processing unit inputs the speaker's utterance for the word through the microphone, and stores the voice signal in the acoustic processing unit and the acoustic data selection unit in the word acoustic data registration unit. Control. As described above, by performing confirmation processing for the speaker when the recognition word is registered as acoustic data, processing by speaker training is minimized, and unintended synthetic speech is registered and the speech recognition rate is reduced. Can be prevented.

The present invention is based on a voice recognition device for a specific speaker that can be realized at low cost, and utilizes a voice synthesis function that is often used together with voice recognition. It is possible to provide a speech recognition apparatus that does not seem to be a training burden on the speaker by automatically performing it inside the device instead of speaking by the speaker.

[Brief description of drawings]

FIG. 1 is a block diagram showing a basic configuration of a telephone equipped with a voice recognition device of the present invention.

FIG. 2 is a block diagram showing a configuration of a voice recognition processing unit according to the first embodiment of the present invention.

FIG. 3 is a block diagram showing a configuration of a voice recognition processing unit according to the second embodiment of the present invention.

FIG. 4 is a block diagram showing an example of a word acoustic data storage unit necessary for the voice recognition device of the present invention.

FIG. 5 is a block diagram showing another example of a word sound data storage unit necessary for the voice recognition device of the present invention.

FIG. 6 is a block diagram showing another example of a word sound data storage unit necessary for the voice recognition device of the present invention.

FIG. 7 is a flowchart showing a conventional registration process procedure of word acoustic data accompanied by training.

FIG. 8 is a flowchart showing a procedure for registering word acoustic data without training according to the present patent.

FIG. 9 is a block diagram showing the configuration of a general specific speaker method speech recognition device.

FIG. 10 is a block diagram showing a configuration of a voice recognition processing unit of the voice recognition device.

FIG. 11 is a block diagram showing a configuration of a word acoustic data storage unit of the voice recognition device.

[Explanation of symbols]

13 microphone 14 Signal processing unit 15 Speech synthesis processing unit 16 Speech recognition processing unit 17 Input device 18 speakers 19 Synthetic speech converter 20 Sound processing unit 21 Word acoustic data storage 22 Word identifier 23 Display 27 Acoustic data selection section 28 word identifiers 29 Synthetic sound acoustic data 30 vocal sound data

Continued front page    (72) Inventor Masakazu Tateyama             1006 Kadoma, Kadoma-shi, Osaka Matsushita Electric             Sangyo Co., Ltd. (72) Inventor Hiroto Nishida             1006 Kadoma, Kadoma-shi, Osaka Matsushita Electric             Sangyo Co., Ltd. (72) Inventor Yoshiaki Kuroki             1006 Kadoma, Kadoma-shi, Osaka Matsushita Electric             Sangyo Co., Ltd. (72) Inventor Yasuyuki Nishioka             1006 Kadoma, Kadoma-shi, Osaka Matsushita Electric             Sangyo Co., Ltd. (72) Inventor Tatsuhiro Goto             1006 Kadoma, Kadoma-shi, Osaka Matsushita Electric             Sangyo Co., Ltd. F-term (reference) 5D015 AA03 HH04 KK04                 5D045 AB30

Claims (7)

[Claims] 1. A microphone as a voice input device, a speaker as a voice output device, and an input device such as a keyboard,
A display device for displaying a recognition result, the microphone, the speaker, the input device and the output device are connected,
A signal processing unit for performing processing control of the entire voice recognition device;
A voice synthesis processing unit that inputs text information from the signal processing unit and outputs a synthesized voice signal, and a voice signal input from the signal processing unit is compared with a plurality of acoustic data held internally, and the results match. In a voice recognition device provided with a voice recognition processing unit for outputting the result to the signal processing as a result of voice recognition, a synthetic voice conversion for converting a synthetic voice signal from the voice synthesis processing unit into synthetic voice acoustic data which is acoustic data. By providing a section, synthetic speech sound data is stored in the speech recognition processing unit instead of the sound data that was generated by training by the speaker in the past and used for speech recognition, so training that is a burden on the speaker is required. A voice recognition device characterized by not performing. 2. The voice recognition processing unit according to claim 1, wherein the voice recognition processing unit has an acoustic data selection unit for selecting either the synthetic sound acoustic data output from the synthetic speech conversion unit or the acoustic data output from the acoustic processing unit. Synthesized voice acoustic data is stored in the word acoustic data storage unit in the initial stage of the device, and when a certain recognition word is recognized, the synthetic speech acoustic data is replaced with the acoustic data corresponding to the speech of the corresponding word. A voice recognition device characterized in that it is possible to recognize acoustic data with actual pronunciation from the first time onward, and prevents a decrease in the voice recognition rate even when the synthetic voice and the actual pronunciation are different. 3. The word sound data storage unit according to claim 2, wherein both of the synthesized sound sound data output from the synthesized sound conversion unit and the sound data of the talker's speech output from the sound processing unit are stored. A voice recognition device, characterized in that the voice recognition rate is improved by holding and outputting the identification information of the corresponding word when it matches with either of them. 4. The word sound data storage unit according to claim 1, having a configuration in which a plurality of pronunciation methods for one word are input and stored from the synthetic speech conversion unit, and a speaker makes any pronunciation. A voice recognition device characterized by being able to correctly recognize the corresponding word even in case of failure. 5. The word acoustic data storage unit according to claim 4, wherein a speaker utters a corresponding word, leaves the corresponding acoustic data and deletes other acoustic data, thereby performing unnecessary identification processing from the next time. A voice recognition device, which is omitted and performs recognition processing at a higher speed. 6. The word sound data storage unit according to claim 5, wherein a mechanism for holding the frequency at which the occurrence of a speaker matches each of a plurality of synthetic sound sound data for a certain word is added, and the corresponding word is recognized. A speech recognition device characterized by deleting only acoustic data having a matching frequency equal to or lower than a threshold level when being broken. 7. The invention according to claim 1, wherein when the synthesized sound data for each word to be identified is held in the speech recognition processing unit, a synthesized sound is reproduced from the speaker for the speaker, and the intention of the speaker. A voice recognition device characterized by performing a training procedure by a speaker only when it is not intended and confirming whether or not it is a synthetic sound.