JP2002258889A - Dictionary-editable speech recognition device - Google Patents

Abstract

(57) [Summary] [Problem] To improve the recognition rate of an input word. SOLUTION: An edited word is converted into a voice signal, and a degree of coincidence between the converted voice signal and a word in a dictionary is calculated. If the word having the highest degree of matching is different from the edited word, the edited word is determined to be inappropriate, and the result of the determination is output. As a result, the user is prompted to re-edit the word determined to be inappropriate, and the recognition rate can be improved.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus capable of editing a dictionary of words to be recognized.

[0002]

2. Description of the Related Art There is known a dictionary-editable speech recognition apparatus in which a recognition target word is input as a character string in hiragana and registered in a dictionary (for example, Japanese Patent Laid-Open No. 11-311).
No. 991).

[0003]

However, in the conventional dictionary-recognizable speech recognition apparatus, when a user inputs a word whose pronunciation is close to a word already registered in the dictionary, the recognition rate is reduced. Is low.

An object of the present invention is to improve the recognition rate of an input word.

[0005]

According to a first aspect of the present invention, there is provided a voice recognition apparatus which allows a user to edit a dictionary of words to be recognized, wherein the conversion means converts the edited word into a voice signal. Calculating means for calculating the degree of coincidence between the converted speech signal and a word in the dictionary; and determining that the edited word is inappropriate if the word having the highest degree of coincidence is different from the edited word. And output means for outputting the result of the judgment by the judgment means. (2) The dictionary-editable speech recognition device according to claim 2, wherein, if the word having the highest matching degree is an edited word, the difference between the highest matching degree and the next highest matching degree is determined by the determining means. If it is smaller than the threshold value, it is determined that the edited word is inappropriate. (3) The dictionary-editable speech recognition device according to the third aspect is configured such that the conversion unit converts the speech signal into a speech signal using a speech element unique to the user stored in advance. (4) In the voice recognition device capable of editing a dictionary according to claim 4, the noise stored in advance is added to the voice signal by the conversion means. (5) The dictionary-recognizable speech recognition apparatus according to claim 5 is configured to edit a dictionary by inputting a word and its attribute information from an external device. (6) The dictionary-editable speech recognition device according to claim 6, wherein the word input from the external device is combined with the attribute information when the word input from the external device is determined to be inappropriate by the determining means. And generating means for generating a new word.

[0006]

According to the first aspect of the present invention, the user is prompted to re-edit a word determined to be inappropriate, and the recognition rate can be improved. (2) According to the second aspect of the present invention, it is possible to avoid editing an inappropriate word and storing it in the dictionary, and as a result, it is possible to prevent a reduction in the recognition rate for the voice uttered by the user. (3) According to the third aspect of the present invention, a more accurate judgment reflecting the voice of the user is made. (4) According to the fourth aspect of the present invention, a more accurate determination reflecting the noise condition in actual use is made. (5) According to the invention of claim 5, it is possible to reduce the burden of the user editing the dictionary. (6) According to the invention of claim 6, it is possible to reduce the burden of re-editing by the user.

[0007]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS << First Embodiment of the Invention >> A speech recognition unit 1 comprises a signal processing device 11 composed of a CPU 11a and a memory 11b, a dictionary of words to be recognized, and speech from reading words. An external storage device 12 for storing a phoneme pattern for restoring a signal, an AD converter 13 for converting sound recorded by the microphone 2 into a digital signal,
A DA converter 14 that converts a digital audio signal into an analog audio signal, an output amplifier 15 that amplifies the analog audio signal, and the like are provided. The microphone 2, the display 3, the speaker 4, the input device 5, and the like are connected to the voice recognition unit 1. The input device 5 is a device for a user to input a voice recognition start request or an editing operation of a word to be recognized.

FIG. 2 is a flowchart showing a speech recognition process. The signal processing device 11 starts this voice recognition processing when a speech operation is performed by the input device 5. In step 1, the notification sound signal stored in the external storage device 12 is output to the speaker 4 via the DA converter 14 and the output amplifier 15 in order to notify the user of the start of the voice recognition processing, and I do.

[0009] In step 2, the capture of voice is started.
Here, the user speaks a word included in the dictionary. FIG. 4 shows an example of the dictionary stored in the external storage device 12. FIG. 4 shows a personal name dictionary. The signal processing device 11 includes the input device 5
Until the speech switch is operated, the average power of the audio signal recorded by the microphone 2 and converted into the digital signal by the AD converter 13 is calculated. After the utterance switch is operated, when the instantaneous power of the audio signal becomes larger than the average power by a predetermined value or more, it is determined that the user has uttered and the fetching of the voice is started.

In step 3, calculation of the degree of coincidence between the fetched voice and the recognition target word in the dictionary stored in the external storage device 12 is started. The degree of coincidence is a value indicating how similar the speech section part is to the individual recognition target words, and the greater the value, the higher the degree of coincidence. During the calculation of the degree of coincidence, the above-described voice capturing processing is performed in parallel. In step 4, when the instantaneous power of the fetched voice becomes equal to or more than a predetermined time and equal to or less than a predetermined value, it is determined that the utterance of the user has ended and the input is ended.

Next, in Step 5, the calculation of the degree of coincidence is waited for, and the word having the highest degree of coincidence is extracted from the recognition target words for which the computation of the degree of coincidence has been performed. Then, the word having the highest matching degree extracted in step 6 is displayed on the display 3. Also,
At the same time, the word with the highest matching degree is output to the operation target device. For example, when the operation target device is a telephone, when the user utters “Kato”, the third person name “Kato” in the dictionary shown in FIG. 4 is obtained, transmitted to the telephone, and transmitted to the telephone number of “Kato”. The call is started.

FIG. 3 is a flowchart showing a dictionary editing input process. When an editing operation is performed by the input device 5, the signal processing device 11 starts the dictionary editing input process.
FIG. 5 shows an example in which "Sato" is additionally edited in the personal name dictionary shown in FIG. In step 11, a speech signal is generated by combining phoneme patterns stored in the external storage device 12 based on the text data indicating the reading of the edited word. This audio signal need not be an audible signal, and need not reflect the pitch frequency or speech rate.

Next, in step 12, the degree of coincidence between the generated speech signal and all the words in the dictionary of the external storage device 12 is calculated. The dictionary includes the edited recognition target words. FIG. 6 shows an example of the calculation result of the degree of coincidence. Since the degree of matching with the audio signal generated from the edited "Sato" was calculated, the degree of matching of "Sato" is usually the highest. In this example, the degree of coincidence of “Kato” is second. In step 13, it is determined whether the edited word "Sato" is appropriate. If the word with the highest match in the dictionary is different from the edited word "Sato", the edited word "Sato" may be mistaken for the word with the highest match already stored in the dictionary Is high and judged inappropriate. On the other hand, if the word with the highest degree of matching in the dictionary matches the edited word "Sato", the edited word "Sato" is not immediately determined to be appropriate, and an accurate determination is made. The following processing is performed.

The word "Sato" in which the word with the highest degree of coincidence has been edited
If so, the highest matching score is compared with the next highest matching score. If the degree of coincidence of the edited word is R0 and the next highest degree of coincidence is R1,

## EQU1 ## If the relationship of R0 (1-k) <R1 is satisfied, there is not much difference in the matching degree between the word having the highest matching degree and the word having the next highest matching degree, and the word having the highest matching degree is obtained. Is an edited word "Sato", it is highly likely that the edited word is the next word with the highest degree of coincidence, so the edited word is determined to be inappropriate. In Equation 1, k is 1
It is a smaller positive value and the optimal value is determined by experiment. In this embodiment, for example, it is 0.01.

In the example shown in FIG. 6, the edited word "Sato"
Is the highest, R0 = 6150, and among the non-edited words, "Kato" is the highest, and R1 = 61.
00. Substituting these values into Equation 1 gives

6150 × (1−0.01) = 6088.5 <6100, and the edited word “Sato” is easily misidentified as “Kato” already stored in the dictionary, and is determined to be inappropriate. Is determined. It should be noted that even when R0 <R1 due to an error in the calculation, Expression 1 is naturally satisfied.

In step 14, the result of the judgment is displayed on the display 3 as a message. If it is determined to be inappropriate, for example, a message that “edited“ Sato ”may be erroneously recognized as“ Kato ”already stored in the dictionary” is displayed.

Thus, it is possible to avoid editing an inappropriate word and storing it in the dictionary, and as a result, it is possible to prevent a reduction in the recognition rate for the voice uttered by the user.

<< Second Embodiment of the Invention >> A second embodiment in which a phoneme pattern generated in advance based on a user's voice is stored in the external storage device 12 will be described. The configuration of the second embodiment is the same as the configuration of the first embodiment shown in FIG.

The dictionary editing and inputting process of the second embodiment is the same as that of the first embodiment shown in FIG.
The processing is the same except for the processing described above, and only the differences will be described with illustration omitted. In the second embodiment, step 11
Then, based on the text data indicating the reading of the edited word, a speech signal is generated by combining the phoneme patterns stored in the external storage device 12.
Is a phoneme pattern generated in advance based on the voice of the user.

According to the second embodiment, it is possible to accurately judge whether or not the recognition rate is inappropriate even if the reduction in the recognition rate depends on the voice quality of the user. Can be prevented.

<< Third Embodiment of the Invention >> A third embodiment in which data for outputting an audible audio signal and noise data during actual use are stored in the external storage device 12 will be described. I do. The configuration of the third embodiment is the same as the configuration of the first embodiment shown in FIG.

The dictionary editing and inputting process of the third embodiment is the same as that of the first embodiment shown in FIG.
The processing is the same except for the processing described above, and only the differences will be described with illustration omitted. In the third embodiment, in step 11 of FIG. 3, an audible audio signal is generated by combining phoneme patterns stored in the external storage device 12 based on text data indicating the reading of the edited word. Generate. A method of generating an audible audio signal is described in, for example,
No. 81175 discloses this. This technique is
A phoneme is created while presenting a melody to the user and singing a song, and when text data to be converted into a voice signal is input, phoneme pieces are combined according to the text data and converted into a voice signal. It is not always necessary to present the melody to the user.

In the third embodiment, in step 11 of FIG. 3, noise data is added to the generated audio signal, and the result is output as a new audio signal. The noise signal is obtained by acquiring the noise included in the non-utterance part and storing it in the external storage device 12 when the user has used the voice recognition device in the past. If the noise environment is dynamic, multiple types of noise are stored. In this case, a plurality of audio signals are obtained for one editing, and steps S12, S1,
The process in 3 is performed a plurality of times.

According to the third embodiment, it is possible to accurately determine whether or not the recognition rate is inappropriate even if the reduction in the recognition rate depends on noise. As a result, the reduction in the recognition rate is prevented. be able to.

<< Fourth Embodiment of the Invention >> A description will be given of a fourth embodiment in which a word to be recognized is transferred from a mobile phone and a dictionary editing operation is performed. FIG. 7 shows the configuration of the fourth embodiment. Note that the same reference numerals are given to the same devices as those in the configuration of the first embodiment shown in FIG. In the fourth embodiment, the mobile phone 6 is connected to the signal processing device 11 of the voice recognition unit 1.

FIG. 8 is a flowchart showing a dictionary editing input process according to the fourth embodiment. The signal processing device 11
When the recognition target word is transferred from the mobile phone 6, the dictionary editing input processing is started. Here, an example is shown in which the word "Sato" is transferred from the mobile phone 6 to the personal name dictionary shown in FIG. 4, and the dictionary is additionally edited as shown in FIG. In step 21, a speech signal is generated by combining the phoneme patterns stored in the external storage device 12 based on the text data indicating the reading of the additionally edited word. This audio signal need not be an audible signal, and need not reflect the pitch frequency or speech rate.

Next, in step 22, the degree of coincidence between the generated speech signal and all the words in the dictionary of the external storage device 12 is calculated. The dictionary includes the edited recognition target words. FIG. 6 shows an example of the calculation result of the degree of coincidence. Since the degree of matching with the audio signal generated from the edited "Sato" was calculated, the degree of matching of "Sato" is usually the highest. In this example, the degree of coincidence of “Kato” is second. In step 23, it is determined whether the edited word "Sato" is appropriate. If the word with the highest match in the dictionary is different from the edited word "Sato", the edited word "Sato" may be mistaken for the word with the highest match already stored in the dictionary Is high and judged inappropriate. On the other hand, if the word with the highest degree of matching in the dictionary matches the edited word "Sato", the edited word "Sato" is not immediately determined to be appropriate, and an accurate determination is made. The following processing is performed.

The word "Sato" in which the word with the highest degree of matching has been edited
If so, the highest matching score is compared with the next highest matching score. If the degree of coincidence of the edited word is R0 and the next highest degree of coincidence is R1,

## EQU3 ## If the relationship of R0 (1-k) <R1 is satisfied, there is not much difference between the word having the highest matching degree and the word having the next highest matching degree, and the word having the highest matching degree is obtained. Is an edited word "Sato", it is highly likely that the edited word is the next word with the highest degree of coincidence, so the edited word is determined to be inappropriate. In Equation 1, k is 1
It is a smaller positive value and the optimal value is determined by experiment. In this embodiment, for example, it is 0.01.

In the example shown in FIG. 6, the edited word "Sato"
Is the highest, R0 = 6150, and among the non-edited words, "Kato" is the highest, and R1 = 61.
00. Substituting these values into Equation 3 gives

6150 × (1−0.01) = 6088.5 <6100, and the edited word “Sato” is easily misunderstood as “Kato” already stored in the dictionary, and is determined to be inappropriate. Is determined. It should be noted that even when R0 <R1 due to an error in the calculation, Expression 1 is naturally satisfied.

If the word "Sato" edited in step 23 is appropriate, the processing is terminated.
Proceed to. In step 24, the edited recognition target word is re-edited. In the address book of the mobile phone 6, attribute data of a person name as a word to be recognized is stored. For example, gender, relatives, company names, etc. are described in text. By adding this before or after the word to be recognized, a new word to be recognized is generated. This is shown in FIG. To the transferred “Sato”, “●● company” to which the company belongs is added. If gender is used in addition to this, words such as "●● kun" can be added to the end of words. If there is no text to add,
If there are no more combinations to be added by the loop processing, the process fails and the process proceeds to step 25. On the other hand, when there is another combination to be added, the process returns to step 21 and
Evaluate again.

In step 25, the judgment result is displayed on the display 3 as a message. If incorrect, "Edited" ●● "may be mistaken for" xx "in the dictionary" or "No valid additional words in address book" message is displayed . If the dictionary has been re-edited successfully, the display 3 displays ""
The message "Registered company Sato" is displayed.

According to the fourth embodiment, a dictionary and a dictionary are edited by acquiring words and their attribute information from a portable telephone, so that the burden on the user to edit the dictionary can be reduced. Further, when it is determined that a word input from the mobile phone is inappropriate, a new word is generated by combining the word input from the mobile phone and its attribute information, and the dictionary is edited based on the new word. Therefore, the burden on the user for reediting can be reduced.

In the configuration of the above embodiment, the voice recognition unit 1 constitutes the conversion means, the arithmetic means, the judgment means, the output means and the generation means, and the portable telephone 6 constitutes the external equipment. Note that the external device is not limited to a mobile phone.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration of a first embodiment.

FIG. 2 is a flowchart illustrating an utterance switch operation input process according to the first embodiment;

FIG. 3 is a flowchart illustrating a dictionary editing input process according to the first embodiment;

FIG. 4 is a diagram showing an example of a personal name dictionary.

FIG. 5 is a diagram showing an example of editing a dictionary.

FIG. 6 is a diagram illustrating an example of a calculation result of a degree of coincidence between an edited word and a word to be recognized in a dictionary.

FIG. 7 is a diagram illustrating a configuration of a fourth embodiment.

FIG. 8 is a flowchart illustrating a dictionary editing input process according to the fourth embodiment.

FIG. 9 is a diagram showing an example of a dictionary re-edited based on personal name attribute data.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 voice recognition unit 2 microphone 3 display 4 speaker 5 input device 11 signal processing device 11 a CPU 11 b memory 12 external storage device 13 AD converter 14 DA converter 15 output amplifier

Claims (6)

[Claims] 1. A speech recognition apparatus which allows a user to edit a dictionary of words to be recognized, comprising: conversion means for converting an edited word into a speech signal; Calculating means for calculating the degree of coincidence with, if the word having the highest degree of coincidence is different from the edited word,
A dictionary-editable speech recognition device, comprising: a determination unit that determines that the edited word is inappropriate; and an output unit that outputs a determination result of the determination unit. 2. The dictionary-editable speech recognition device according to claim 1, wherein the determining unit determines that the word having the highest matching degree is the edited word and the word having the highest matching degree is next to the word having the highest matching degree. A dictionary-editable speech recognition device, characterized in that if the difference from a high degree of matching is smaller than a threshold value, the edited word is determined to be inappropriate. 3. The voice recognition apparatus according to claim 1, wherein said conversion means converts the voice signal into a voice signal by using a user-stored phoneme fragment stored in advance. Dictionary-recognizable voice recognition device. 4. A speech recognition apparatus capable of editing a dictionary according to claim 1, wherein said conversion means adds noise stored in advance to a speech signal. . 5. The dictionary-recognizable speech recognition apparatus according to claim 1, wherein the dictionary is edited by inputting a word and its attribute information from an external device. . 6. The speech recognition device capable of editing a dictionary according to claim 5, wherein the word input from the external device is used when the word input from the external device is judged to be inappropriate by the judgment means. A dictionary-editable speech recognition device, comprising: a generation unit configured to generate a new word by combining the word and attribute information thereof.