JP3104659B2 - Speech input device and machine-readable recording medium recording program - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice input device, and more particularly to a voice input device having a function of reducing recognition errors by adapting to the contents of a user's utterance.

[0002]

2. Description of the Related Art In the case of inputting a document by voice, a method of improving a recognition performance by creating a statistical language model relating to the order of occurrence of words from a large-scale text database and using the same in voice recognition has been used. Known from. However, the statistical language model created as described above is a general-purpose one. For example, the user is different, or the recognition task (speech content) is different from the text database used to create the statistical language model. The case was less effective.

[0003] In order to solve such a problem, a general-purpose statistical language model is created by using a learning language model created from a text database similar to a recognition task as disclosed in Japanese Patent Laid-Open No. Hei 4-291399. Techniques for adapting and performing speech recognition using the adapted statistical language model have been proposed in the past.

FIG. 7 is a block diagram of this kind of the prior art.

The standard pattern memory 115 stores a plurality of standard patterns learned in advance. The general-purpose statistical language model 116 is created based on a general-purpose text database and represents the order in which words occur. The learning language model 117 is created from a text database similar to the recognition task. The adaptive statistical language model 118 corresponds to the learning language model 11.
7, a general-purpose statistical language model 116 is adapted. As an adaptation method, for example, there is a deletion interpolation method. Specifically, general-purpose statistical language model 1
Assuming that 16 is P, the learning language model 117 is Q, and the adaptive statistical language model 118 is R, the adaptive statistical language model R is represented by R = λ × P + (1−λ) × Q. Here, 0 ≦ λ ≦ 1. in this way,
By mixing P and Q, the adaptive statistical language model 1
18 are obtained.

[0006] The voice input from the voice signal input terminal 111 is converted into a digital signal in the feature extraction unit 112, further subjected to LPC cepstrum analysis, and then converted into feature parameters for each frame.

The recognizing unit 113 obtains the similarity (likelihood) between the standard pattern of the candidates and the parameters of the input speech for a plurality of candidates selected using the adaptive statistical language model 118. Further, the adaptive statistical language model 11
8 and the linear sum of the likelihood of the standard pattern and the parameters of the input speech as the total likelihood, and the candidate with the highest total likelihood as the recognition result is output as the recognition result output unit 114.
Output to

[0008]

The conventional technique using the above-mentioned adaptive statistical language model can realize higher recognition performance than the technique using only a general-purpose statistical language model. There is a problem that a text database similar to the above must be prepared in advance, and an adaptive statistical language model must be created based on the text database.

Accordingly, an object of the present invention is to improve recognition performance by adapting to a recognition task without using text data similar to the recognition task in advance by utilizing the correction of the recognition result by the user. It is an object of the present invention to provide a voice input device capable of performing the above.

[0010]

In order to achieve the above object, the speech input device of the present invention selects an optimal recognition result candidate from a plurality of recognition result candidates obtained for an input speech and recognizes the recognition result for the input speech. An adaptive score storage unit for storing adaptive scores of a plurality of words that can be recognition result candidates; and selecting the recognition result from among a plurality of recognition result candidates for input speech. means for selecting said stored in score storage unit each recognition result candidate recognition result adaptive score even considering the, along with modifying the recognition result selected by said means in accordance with an instruction of the user, among the recognition result candidates From the modified recognition result
From the value of the adaptive score of the recognition result candidate
A user correction unit for reducing the fixed value . Or on
Selected by said means instead of the user correction section
The recognition result is corrected according to the user's instructions,
For recognition result candidates positioned higher than the
And if there is an adaptation score other than the initial value, reduce a certain value
A user correction unit is provided.

In this configuration, when a recognition result is selected from a plurality of recognition result candidates, the recognition result is selected in consideration of the adaptive score of each recognition result candidate stored in the adaptive score storage unit. . If there is an error in the recognition result, the user instructs the user correction unit to correct the recognition result. With this, the user correction unit corrects the recognition result according to the user's instruction , and, among the recognition result candidates,
Recognition result candidates that are higher than the
After subtracting a certain value from the value of the adaptive score, or
For the recognition result candidates located higher than the recognition result,
If there is an adaptation score other than the period value, a certain value is reduced. Therefore, when a word having an erroneous recognition result is input thereafter, the probability of obtaining a correct recognition result increases.

In order to achieve the above object, a voice input device of the present invention analyzes a input voice and outputs a result of the analysis, and a standard pattern storage unit storing standard patterns for a plurality of words. And a voice recognition unit that obtains a distance between the analysis result of the parameter analysis unit and each of the standard patterns stored in the standard pattern storage unit and obtains a plurality of recognition result candidates based on the obtained distances. And an adaptive score storage unit that stores adaptive scores of a plurality of words that can be recognition result candidates, and an adaptive score storage unit that stores, from among a plurality of recognition result candidates obtained by the speech recognition unit, A candidate selection unit that selects a recognition result based on an adaptation score of each recognition result candidate and a distance between a standard pattern of each recognition result candidate and a candidate selection unit; Together to correct it has been recognition result selected in accordance with an instruction of the user I, of recognition result candidates,
Recognition result candidates that are higher than the
A user correction unit for subtracting a fixed value from the value of the adaptive score . Alternatively, instead of the user correction unit,
User's instruction on the recognition result selected by the candidate selection unit
And higher than the corrected recognition result
For the recognition result candidates located in
A user correction unit that reduces a certain value if
You.

In this configuration, when the user inputs a word by voice, the parameter analysis unit analyzes the input voice, and the voice recognition unit stores the analysis result of the parameter analysis unit and the standard pattern storage unit. A plurality of recognition result candidates are obtained based on a distance from each standard pattern, and a candidate selection unit is stored in an adaptive score storage unit from among a plurality of recognition result candidates obtained by a speech recognition unit. A recognition result is selected based on the adaptive score of each recognition result candidate and the distance between the standard pattern of each recognition result candidate.

If there is an error in the recognition result selected by the candidate selection unit, the user instructs the user correction unit to correct the error. Accordingly, the user correcting unit corrects the recognition result selected by the candidate selecting unit in accordance with the user's instruction , and, among the recognition result candidates, the corrected recognition result.
Is the value of the adaptation score of the recognition result candidate positioned higher than
From the recognition result after correction, or
Of the recognition result candidate located at
If there is a core, subtract a certain value.

Further, in order to further improve the recognition performance, the voice input device of the present invention analyzes the input voice and outputs a result of the analysis, and a standard pattern storing a standard pattern for a plurality of words. A voice for obtaining a distance between the analysis result of the parameter analysis unit and each standard pattern stored in the standard pattern storage unit, and obtaining a plurality of recognition result candidates based on the obtained distances; A recognition unit, an adaptive score storage unit for storing adaptive scores of a plurality of words that can be recognition result candidates, a statistical language model indicating an order of occurrence of words, and a plurality of recognition result candidates obtained by the speech recognition unit. , The distance between the adaptive score of each recognition result candidate stored in the adaptive score storage unit and the standard pattern of each recognition result candidate, Based on the content of the language model, a language processing unit for selecting a recognition result, a recognition result selected by該言word processor as well as modified according to a user instruction, of the recognition result candidates, the recognition result after correction Above
Constant from the value of the adaptive score of the recognition result candidate located at
And a user correction unit for reducing Alternatively,
Selected by the language processing unit instead of the user correction unit.
Correct the recognition result according to the user's instruction,
Recognition result candidates positioned higher than the corrected recognition results
On the other hand, if there is an adaptation score other than the initial value, a certain value is reduced.
User correction unit.

In this configuration, the language processing unit selects, from among the plurality of recognition result candidates obtained by the speech recognition unit, the adaptive score of each of the recognition result candidates stored in the adaptive score storage unit; A recognition result is selected based on the distance between each recognition result candidate and the standard pattern and the contents of the statistical language model.

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

Embodiments of the present invention will now be described in detail with reference to the drawings.

First, the functional outline of the voice input device of the present invention will be briefly described. The voice input device of the present invention obtains a plurality of recognition result candidates for an input voice, and when selecting a recognition result from among them, not only the distance between the input voice and the recognition result candidate, but also The value of the adaptive score given to each recognition result candidate based on the correction operation performed by the user is also considered.

FIG. 1 is a block diagram showing an example of the configuration of a voice input device according to one embodiment of the present invention. A parameter analysis unit 1, a standard pattern storage unit 2, a voice recognition unit 3, a candidate selection unit 4, an adaptive score storage unit 5, a display unit 6, and a user correction unit 7.

The standard pattern storage unit 2 stores a plurality of standard patterns (for example, expressed as a parameter vector sequence) of speech analyzed in advance. The adaptive score storage unit 5 stores adaptive scores of a plurality of words that can be recognition result candidates.

The parameter analyzer 1 has a function of analyzing an input voice and outputting an analysis result (for example, a parameter vector sequence). Such a parameter analyzer 1
For example, it can be configured using a filter bank, a Fourier transform, a linear prediction coefficient type analyzer, or the like.

The voice recognition unit 3 calculates the distance between the analysis result of the parameter analysis unit 1 and each of the standard patterns stored in the standard pattern storage unit 2, and based on the obtained distances, a plurality of recognition results. It has a function of obtaining candidates and a function of outputting each recognition result candidate together with its distance. Note that certainty can be used instead of the distance.

The candidate selecting section 4 selects, from among the recognition result candidates output from the speech recognition section 3, the adaptive score of each of the recognition result candidates stored in the adaptive score storage section 5 and the adaptive score of each of the recognition result candidates. It has a function of selecting a recognition result based on the distance from the standard pattern.

The display unit 6 displays the recognition result selected by the candidate selection unit 4 and the recognition result candidates.

The user correction unit 7 has a function of displaying another recognition result candidate on the display unit 6 when the recognition result displayed on the display unit 6 is incorrect and the correction is instructed by the user. 6, a function of setting the recognition result candidate selected by the user among the recognition result candidates displayed as the recognition result after the correction, and an adaptive score so that the corrected recognition result is easily selected thereafter (to be advantageous). It has a function of updating the contents of the storage unit 5.

FIG. 2 is a flowchart of a processing example of the voice input device shown in FIG. 1, and the operation will be described below with reference to the drawings.

When the user utters a word to be input (step S1), the parameter analysis unit 1 analyzes the input voice and outputs an analysis result, and the voice recognition unit 3 analyzes the input voice and the standard pattern storage unit 2. Is compared with the standard pattern of each word, and the distance between the input voice and the standard pattern of each word is obtained, and a plurality of recognition result candidates are output together with the distance (step S2).

When the plurality of recognition result candidates and their distances are output from the voice recognition unit 3, the candidate selection unit 4 obtains an adaptive score for each of the recognition result candidates from the adaptive score storage unit 5 (step S3). . After that, the candidate selecting unit 4 calculates a linear sum of the inverse of the sign of the distance and the value of the adaptive score for each recognition result candidate, sorts the recognition result candidates in descending order, and selects the first recognition result candidate ( The recognition result candidate having the largest linear sum) is displayed on the display unit 6 as the recognition result (step S4).

FIG. 3 shows a specific example of the reordering process of the recognition result candidates performed in step S4. In this example, the user utters the word “voice”, and the recognition result candidates “hot spring”, “voice”, “head office” and distances “30”, “36”, and “44” are output from the voice recognition unit 3. Then, the recognition result candidates “hot spring”, “voice”,
"0", "1" as the adaptation score of "Headquarters"
The processing when “0” and “16” are stored is shown.

The candidate selection unit 4 determines whether each of the recognition result candidates “hot spring”, “voice”, and “head office” has the opposite sign of the distance, such as “−30”, “−36”, and “−44”. The linear sum of the score values “0”, “10”, and “16” is obtained, and the recognition result candidates are rearranged in descending order of the linear sum. In this example, the linear sums of the recognition result candidates “hot spring”, “voice”, and “head office” are “−30” and “−”, respectively.
26 "," -28 ", so" voice "," head office ",
The recognition result candidates are sorted in the order of “hot spring”.

If the recognition result displayed on the display unit 6 is correct (No in step S5), the user utters the next word to be input (step S1), and if the recognition result contains an error, A correction instruction is input to the user correction unit 7 from a keyboard, a mouse (not shown), or the like.

When a correction instruction is input to the user correction section 7 (Yes in step S5), the processing in step S6 described below is performed.

First, the user correction unit 7 requests the candidate selection unit 4 for a recognition result candidate. In response to this request, the candidate selection unit 4 passes to the user correction unit 7 a plurality of recognition result candidates passed from the speech recognition unit 3 excluding the recognition result candidate that is the current recognition result.

Thus, the user correcting section 7 displays all the recognition result candidates passed from the candidate selecting section 4 on the display section 6. When the recognition result candidates are displayed on the display unit 6, the user selects a correct recognition result candidate from the display using the mouse or the like.

When a correct recognition result candidate is selected by the user, the user correction section 7 performs a correction process using the selected recognition result candidate as a recognition result. Further, the user correction unit 7 stores the selected recognition result candidate (correction result) in the adaptive score storage unit 5 by a method shown in (A) below, for example, so that the candidate is easily selected by the candidate selection unit 4 thereafter. Update the content.

(A) If the adaptive score resulting from the correction is an initial value, the value is set to a constant value A. If the adaptation score resulting from the correction is other than the initial value, a fixed value B is added to the value. If the recognition result candidate higher than the correction result has an adaptation score other than the initial value, a certain value C is reduced. (here,
A, B, and C preferably have a relationship of A> B, A> C)

FIG. 4 shows a specific example of the adaptive score updating process. In this example, the word "voice" is uttered, and "Headquarters" is obtained as the recognition result. Also, "hot spring" is ranked second as a recognition result candidate, and "voice" is ranked third.
Is obtained, and the process is performed when “voice” is selected from the displayed recognition candidates. In this example, the initial value of the adaptive score is set to 0.

Since the adaptation score (10 in the example in the figure) other than the initial value has already been set in the correction result “voice”,
A fixed value (4 in the example in the figure) is added to the adaptation score for “voice” to obtain a new adaptation score (14 in the example in the figure). In addition, the recognition result candidates higher than the correction result “voice” and having the adaptation score other than the initial value are “head office”
Therefore, a certain value (5 in the example in the figure) is subtracted from the adaptation score of the recognition result candidate “head office” to obtain a new adaptation score (11 in the example in the figure). As described above, by updating the adaptive score, when “speech” is input again, the candidate selection unit 4 can easily select the recognition result candidate “speech”. In the example of FIG. 4, the initial value of the adaptive score is set to 0. However, the initial value of the adaptive score is determined by using text data similar to the recognition task. You can also.

When the processing in step S6 is completed, the user voice-inputs the next word to be input (step S1). Thereafter, every time a word is input by speech, the same processing as described above is performed.

As a method of updating the adaptive score storage unit 5, the following methods (B) to (I) can be adopted in addition to the method (A) described above.

(B) The value of the adaptive score resulting from the correction is always set to a constant value A.

(C) A constant value B is always added to the value of the adaptive score resulting from the correction.

(D) If the adaptation score of the correction result is an initial value, the value of the adaptation score of the recognition result candidate is set to a constant value A
To If the adaptive score of the correction result is other than the initial value, a certain value B is added to the value of the adaptive score of the recognition result candidate.

(E) In the above method (B), (C) or (D), a fixed value C is obtained from the adaptive score of the recognition result candidate positioned higher than the correction result among the recognition result candidates.
Reduce.

(F) In the above method (B) or (C), if there is an adaptive score other than the initial value for the recognition result candidate positioned higher than the correction result, a certain value C is reduced.

(G) The above methods (A), (B),
In (C), (D), (E) or (F), the value of the adaptive score stored in the adaptive score storage unit 5 becomes positive each time the user inputs a word by voice. The value of the adaptive score is reduced by a constant value D. When this method is adopted, for example, each time the recognition result is displayed on the display unit 6 in the candidate selecting unit 4, the user correcting unit 7 is notified. It is sufficient that the value of the adaptive score stored in the adaptive score storage unit 5 is subtracted by a constant value D from the adaptive score having a positive value.

(H) The above methods (A), (B),
In (C), (D), (E) or (F), every time the user correction unit 7 corrects the recognition result, the values of all the adaptive scores stored in the adaptive score storage unit 5 are kept constant. An attenuation process of reducing by the value D is performed.

(I) The above methods (A), (B),
In each of (C), (D), (E) and (F), each time the user correction unit 7 corrects the recognition result, a fixed value is set from the positive value adaptive score stored in the adaptive score storage unit 5. An attenuation process of reducing the value D is performed.

The voice input device of this embodiment is very effective when used, for example, in a voice input document creation device. In general, a user creates documents in various fields using a voice input document creation device, and when inputting a group of documents, the same word is often used repeatedly in the documents. In the present embodiment, if the recognition result contains an error and the user corrects the recognition result, the candidate selection unit 4
Since the contents of the adaptive score storage unit 5 are updated so that the correction result can be easily selected in the method described in JP-A-4-291.
No text data similar to the task (recognition task) of the document to be created is prepared as in the prior art described in Japanese Patent Application Publication No. 399, and documents in various fields are smoothly read without causing recognition errors. It becomes possible to input.

Also, the voice input device of the present embodiment is applied to a system that presents a plurality of candidates for utterance by, for example, a command input by voice, and allows the user to select a command to be input from the candidates. Is also very effective.
If the voice input device of this embodiment is used, when a previously misrecognized command is input again by voice, the command is presented at a higher position, and the user interface can be improved. However, in such a case, it is necessary to slightly change the configuration of the candidate selection unit 4 so that not only the top command but also the remaining candidates are displayed on the display unit 6 in descending order of the linear sum. It is.

When the above methods (G), (H), and (I) are used as a method of updating the adaptive score storage unit 5, recognition errors frequently occur immediately after the recognition task becomes different. Can not be. For example,
If the recognition task to be input first is “an invention related to a voice input device” and the next recognition task to be input is “a topic related to a hot spring”, if “an invention related to a voice input device” is Assume that "voice" is erroneously recognized as "hot spring", and the adaptive score of "voice" is increased. In such a case, if you keep the adaptive score of "voice" high, the next time you enter "topics on hot springs"
The risk that "hot spring" is erroneously recognized as "voice" is increased.
In the above methods (G), (H), and (I), each time the user inputs a word or every time the user corrects the recognition result, the adaptive score stored in the adaptive score storage unit 5 is calculated. Since the value is reduced, the value of the adaptation score of “voice” is likely to be 0 when “topic on hot spring” is input. Therefore, according to the above methods (G), (H), and (I), it is possible to prevent recognition errors from occurring frequently immediately after a different recognition task.

Although not described in the above-described embodiment, recognition results and recognition result candidates for utterances before the latest utterance are stored, and past results are stored based on the stored contents. The recognition result can be made modifiable.

FIG. 5 is a block diagram of another embodiment of the present invention. The speech input device of the present embodiment has a configuration in which the speech input device shown in FIG. A holding unit 8 and a lattice forming unit 9 that forms a lattice from recognition result candidates given from the recognition result holding unit 8
It has. The lattice is composed of recognition result candidates for each utterance.

Further, the voice input device of the present embodiment is provided with a language processing unit 10 instead of the candidate selection unit 4 shown in FIG. The language processing unit 10 determines the distance between the lattice provided from the lattice configuration unit 9 and the standard pattern of each recognition result candidate of the input voice provided from the voice recognition unit 3,
Based on the value of the adaptive score of the recognition result candidate stored in the adaptive score storage unit 5 and the contents of the statistical language model (indicating the generation order of words) possessed by the language processing unit 10, each utterance in the lattice is It has a function of obtaining an optimum recognition result from each of the recognition result candidates. In FIG. 5, the same reference numerals as those in FIG. 1 represent the same parts.

As a specific example of the use of the adaptation score when selecting a recognition result from a plurality of recognition result candidates, a lattice in which recognition result candidates for discrete word utterances are arranged in time series is used as a statistical language model. The case where the bigram model is used will be described.

In the language processing unit 10, S (w_w_
, W_n) to maximize the word sequence w_1,.
_N is determined as a recognition result for the lattice.

[0064]

(Equation 1)

Here, n is the lattice width (the number of utterances to be processed), w_i is one of the recognition result candidates for the i-th utterance, P (w_i) is the appearance probability of w_i, and P (w
_I / w_ (i-1)) is a conditional occurrence probability of w_i when w_ (i-1) appears. U_sc
ore (w_i) is the adaptation score for w_i, a_s
core (w_i) is obtained by reversing the sign of the distance between the input speech and the standard pattern, which is obtained by the speech recognition unit 3 using the acoustic score for w_i. c is a predetermined constant.

The adaptive score reflects the contents of the user's utterance so far and the tendency of the recognition error of the voice recognition unit 3, and improves the performance as compared with the case where only the word bigram prepared in advance is used. Means can be provided. In addition, higher performance can be provided by using the word bigram as compared with the voice input device shown in FIG.

FIG. 6 is a block diagram showing a hardware configuration of the voice input device, which comprises a computer 61, a storage medium 62, a storage device 63, and a display device 64. The storage medium 62 is a disk, a semiconductor memory, or another storage medium, and stores a program for causing the computer 61 to operate as a voice input device.

In the case of realizing the voice input device shown in FIG. 1, the program recorded on the recording medium 62 is read by the computer 61 and the operation of the computer 61 is controlled so that the program shown in FIG. A parameter analysis unit 1, a speech recognition unit 3, a candidate selection unit 4, and a user correction unit 7 are realized. Note that the standard pattern storage unit 2 and the adaptive score storage unit 5 are realized on a storage device 63, and the display unit 6 is realized by a display device 64.

When the voice input device shown in FIG. 5 is realized, the program recorded on the recording medium 62 is read by the computer 61 and the operation of the computer 61 is controlled so that the program is stored on the computer 61 as shown in FIG. The parameter analysis unit 1, the speech recognition unit 3, the user correction unit 7, the lattice configuration unit 9, and the language processing unit 10 shown in FIG.
Note that the standard pattern storage unit 2 and the adaptive score storage unit 5 are realized on a storage device 63, and the display unit 6 is realized by a display device 64.

[0070]

As described above, according to the present invention, when a user corrects a recognition result, the correction is performed within the recognition result candidates.
Adaptation of recognition result candidates positioned higher than the later recognition results
Recognition after subtracting a fixed value from the score value or correcting it
Initial value for the recognition result candidate higher than the result
If there is an adaptation score other than, a certain value is reduced, so that the following first to third effects can be obtained.

The first effect is that the recognition performance can be improved by adapting to the utterance content without preparing text data similar to the utterance content in advance.

The second effect is that the recognition error tendency inherent to the user can be reflected, and the recognition error can be reduced.

The third effect is that the recognition performance can be improved by adapting to the utterance content and the tendency of the user to make mistakes as the user uses it. That is, the higher the use, the higher the recognition performance can be obtained.

Further, the present invention reduces the adaptive score value of each recognition result candidate stored in the adaptive score storage unit each time the user inputs a word by speech or corrects the recognition result. Therefore, immediately after the recognition tasks become different, it is possible to prevent recognition errors from occurring frequently.

[Brief description of the drawings]

FIG. 1 is a block diagram of one embodiment of the present invention.

FIG. 2 is a flowchart illustrating a processing example of the voice input device illustrated in FIG. 1;

FIG. 3 is a diagram for explaining a process of rearranging recognition result candidates.

FIG. 4 is a diagram for explaining an update process for an adaptive score storage unit 5;

FIG. 5 is a block diagram of another embodiment of the present invention.

FIG. 6 is a diagram illustrating an example of a hardware configuration of a voice input device.

FIG. 7 is a block diagram of the related art.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Parameter analysis part 2 ... Standard pattern storage part 3 ... Speech recognition part 4 ... Candidate selection part 5 ... Adaptive score storage part 6 ... Display part 7 ... User correction part 8 ... Recognition result holding part 9 ... Lattice construction part 10 ... Language Processing unit 111 Voice signal input terminal 112 Feature extraction unit 113 Recognition unit 114 Recognition result output unit 115 Standard pattern memory 116 General-purpose statistical language model 117 Learning language model 118 Adaptive statistical language model

Continuation of front page (56) References JP-A-59-3951 (JP, A) JP-A-4-291399 (JP, A) JP-A-63-201698 (JP, A) JP-A-58-93099 (JP, A) JP-A-59-17595 (JP, A) JP-A-61-292200 (JP, A) JP-A-61-69099 (JP, A) Patent 2502880 (JP, B2) JP, B2) JP 1-43960 (JP, B2) JP 7-104679 (JP, B2) IEEE Transactions on Pattern Analysis is Machine Inteligence, Vol. PAMI-12, no. 6, June 1990, "A Cache-Based Natural Language Model for Speech Recognition", p. 570-583 Sharp Technical Report, Vol. 31, No. 31, Mitsuhiro Tootani, "Japanese Speech Input Device 10-8335," p. 97-103, issued March 20, 1985 (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 15/22 G10L 15/18 JICST file (JOIS)

Claims (12)

(57) [Claims] 1. A speech input device for selecting an optimum one from a plurality of recognition result candidates obtained for an input speech and for obtaining a recognition result for the input speech, comprising: An adaptive score storage unit in which an adaptive score is stored; and selecting a recognition result from a plurality of recognition result candidates obtained for the input voice, when selecting the recognition result candidates stored in the adaptive score storage unit. means for selecting a recognition result adaptive score even considering the recognition result selected by said means as well as modified in accordance with an instruction from the user, among the recognition result candidates, certified after correction
Adaptive score of the recognition result candidate higher than the recognition result
And a user correction unit for subtracting a constant value from the value of the voice input device. (2)Instead of the user correction unit, The recognition result selected by the above means is converted into a user instruction.
Therefore, correct it and place it higher than the corrected recognition result.
Adaptive score other than the initial value for the recognition result candidate
That there is a user correction unit that reduces a certain value if there is
The voice input device according to claim 1, wherein: 3. A parameter analysis unit for analyzing an input voice and outputting an analysis result, a standard pattern storage unit in which standard patterns for a plurality of words are stored, an analysis result of the parameter analysis unit, and storage of the standard pattern. A voice recognition unit that calculates a distance between each of the standard patterns stored in the unit and obtains a plurality of recognition result candidates based on the obtained distances; and an adaptive score of a plurality of words that can be a recognition result candidate. An adaptive score storage unit that is stored; an adaptive score of each recognition result candidate stored in the adaptive score storage unit; and a distance between a standard pattern of each recognition result candidate and the speech recognition. A candidate selecting unit for selecting a recognition result from among a plurality of recognition result candidates obtained by the unit, and correcting the recognition result selected by the candidate selecting unit in accordance with a user instruction Rutotomoni, of the recognition result candidate, Modify
Adaptation of recognition result candidates positioned higher than the later recognition results
A voice input device comprising: a user correction unit for subtracting a fixed value from a score value . (4)Instead of the user correction unit, The recognition result selected by the candidate selection unit is
Make corrections according to the instructions, and
Adaptation other than the initial value to the recognition result candidates at the top
Equipped with a user correction unit that reduces a certain value if there is a score
4. The voice input device according to claim 3, wherein: 5. A parameter analysis unit for analyzing an input voice and outputting an analysis result, a standard pattern storage unit storing standard patterns for a plurality of words, an analysis result of the parameter analysis unit and storing the standard pattern. A voice recognition unit that calculates a distance between each of the standard patterns stored in the unit and obtains a plurality of recognition result candidates based on the obtained distances; and an adaptive score of a plurality of words that can be a recognition result candidate. A stored adaptive score storage unit, a statistical language model indicating an order of occurrence of words, an adaptive score of each recognition result candidate stored in the adaptive score storage unit, and a standard pattern of each recognition result candidate. Language processing for selecting a recognition result from among a plurality of recognition result candidates obtained by the speech recognition unit based on a distance between the recognition result and the contents of the statistical language model When the recognition result selected by該言word processor as well as modified according to a user instruction, of the recognition result candidates, modified
Adaptation of recognition result candidates positioned higher than the later recognition results
A voice input device comprising: a user correction unit for subtracting a fixed value from a score value . 6.Instead of the user correction unit, The recognition result selected by the language processing unit is
Make corrections according to the instructions, and
Adaptation other than the initial value to the recognition result candidates at the top
Equipped with a user correction unit that reduces a certain value if there is a score
Characterized by The voice input device according to claim 5. 7. A computer provided with an adaptive score storage unit for storing adaptive scores of a plurality of words that can be recognition result candidates is selected from a plurality of recognition result candidates obtained from input speech. A machine-readable recording medium that records a program for functioning as a speech input device that serves as a recognition result for the input speech, wherein the computer is configured to select a plurality of recognition result candidates obtained for the input speech. Means for selecting a recognition result in consideration of the adaptive score of each recognition result candidate stored in the adaptive score storage unit when selecting a recognition result; correcting the recognition result selected by the means in accordance with a user instruction And the corrected recognition result among the recognition result candidates.
Adaptive score of the recognition result candidate higher than the recognition result
A machine-readable recording medium on which a program for functioning as a user correcting unit for subtracting a certain value from the value of the above is recorded. 8.The computer, the user correction unit
Instead of, The recognition result selected by the above means is converted into a user instruction.
Therefore, correct it and place it higher than the corrected recognition result.
Adaptive score other than the initial value for the recognition result candidate
Function as a user correction unit that reduces a certain value if there is
Claims recorded a program for the
A machine-readable record that records the program according to 7.
Medium. 9. A computer comprising: a standard pattern storage unit for storing standard patterns for a plurality of words; and an adaptive score storage unit for storing adaptive scores of a plurality of words that may be recognition result candidates. And a parameter analysis unit that outputs an analysis result.A distance between the analysis result of the parameter analysis unit and each of the standard patterns stored in the standard pattern storage unit is obtained. Based on each of the obtained distances, A voice recognition unit that obtains a plurality of recognition result candidates, and among the plurality of recognition result candidates obtained by the voice recognition unit, the adaptive score of each recognition result candidate stored in the adaptive score storage unit; A candidate selecting unit for selecting a recognition result based on a distance between each recognition result candidate and a standard pattern; With modified according, among the recognition result candidates, modified
Adaptation of recognition result candidates positioned higher than the later recognition results
A machine-readable recording medium on which a program for functioning as a user correction unit for subtracting a certain value from a score value is recorded. 10.Modify the computer with the user
Instead of department, The recognition result selected by the candidate selection unit is
Make corrections according to the instructions, and
Adaptation other than the initial value to the recognition result candidates at the top
If there is a score, it can be used as a user correction unit that reduces a certain value.
Characterized by recording a program for functioning
Machine readable recording the program according to claim 9
Recording medium. 11. A standard pattern storage unit for storing standard patterns for a plurality of words, an adaptive score storage unit for storing adaptive scores of a plurality of words that may be recognition result candidates, and statistics indicating the order of occurrence of the words. A parameter analysis unit that analyzes an input voice and outputs an analysis result between the analysis result of the parameter analysis unit and each standard pattern stored in the standard pattern storage unit. A voice recognition unit that obtains a distance and obtains a plurality of recognition result candidates based on the obtained distances; and a plurality of recognition result candidates obtained by the voice recognition unit. Based on the adaptive score of each recognition result candidate and the distance between the standard pattern of each recognition result candidate and the contents of the statistical language model. A language processing unit to be selected; a recognition result selected by the language processing unit is corrected in accordance with a user instruction ;
Adaptation of recognition result candidates positioned higher than the later recognition results
A machine-readable recording medium on which a program for functioning as a user correction unit for subtracting a certain value from a score value is recorded. 12.Modify the computer with the user
Instead of department, The recognition result selected by the language processing unit is
Correct according to instructions Together with the corrected recognition result
Adaptation other than the initial value to the recognition result candidates at the top
If there is a score, it can be used as a user correction unit that reduces a certain value.
Characterized by recording a program for functioning
Machine readable recording the program according to claim 11
Recording media.