JPH11149294A - Voice recognition device and voice recognition method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve recognition capacity by making best use of recognized results of the past, when voice is re-input. SOLUTION: A restatement judging part 20 carries out inter-input-pattern matching between this time input voice pattern and the past input voice pattern to judge whether this time vocalization is a restatement vocalization or not. A recognition processing part 30 carries out recognition matching of this time input voice pattern with the standard pattern to select a recognition candidate. If the result of the restatement judging part 20 is not the restatement vocalization, the recognition candidate selected this time is output as the recognized result as it is. In the case of the restatement vocalization, the adjusted recognition candidate is decided, based on double matching results using both of the past recognition candidates stored in a recognition candidate registration part 32 and the recognition candidate is obtained by the processing of this time.

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 and method, and more particularly to a speech recognition apparatus and method capable of improving the recognition ability when a speaker re-utters.

[0002]

2. Description of the Related Art Conventionally, a speech recognition apparatus for performing pattern matching between a speech pattern of an utterance input and a standard pattern has been known. The speech recognition device can be used as input means for various devices. For example, if a vehicle is provided with a voice recognition device, the driver can operate the in-vehicle device by uttering a voice command, thereby improving operability,
The burden on the driver can be reduced.

FIG. 1 shows a configuration of a conventional speech recognition apparatus. When a speaker speaks toward the microphone 10, an input voice signal from the microphone 10 is converted into a digital signal by the A / D converter 12 and input to the acoustic processing unit 14. In the acoustic processing unit 14, acoustic analysis is performed, an input voice pattern of a characteristic parameter (such as a cepstrum) is generated, and the input voice pattern is supplied to the recognition processing unit 16.

The recognition processing section 16 performs pattern matching between an input voice pattern and a standard pattern. For example, when performing word recognition, the recognition dictionary storage unit 18 stores a plurality of word data to be recognized. The input speech pattern and each word are individually matched, and the word with the best matching result is selected as the first recognition candidate. Further, lower recognition candidates (words) are appropriately selected in a descending order of matching. Then, the recognition processing unit 16
One or more selected recognition candidates are output as a recognition result.

[0005] As a method of pattern matching in the recognition processing unit 16, the probability of using a dynamic programming method (dynamic programming method, hereinafter referred to as DP method) or a hidden Markov model (hidden Markov model, hereinafter referred to as HMM) is used. Techniques are known. As is well known, the former D
The P method is suitable for recognizing a specific speaker's voice, and the latter H
The method using MM is suitable for recognizing the voice of an unspecified speaker. The DP method is also called DTW (Dynamic Time Warping).

[0006]

In a speech recognition apparatus, in order to inquire a speaker whether or not a recognition result is correct,
For example, a synthesized voice for notifying the recognition candidate is output (talkback), and, for example, the recognition candidate is displayed on the display. Then, if necessary, another utterance input is performed by the speaker to correct the recognition result. The utterance input performed once again after the utterance input in this way is referred to as “repetition input”.

[0007] The re-utterance input is performed when the speaker notices that he or she has said something wrong, or when the speaker gets stuck in words on the way, so that the voice is intended as intended by the speaker. For example, when it is not recognized. The re-utterance in such a case is a “paraphrase utterance” in which a different word is uttered.

[0008] In addition to the above paraphrased utterance, the re-utterance input includes "rephrase utterance". At present, the recognition accuracy rate of the speech recognition device does not reach 100%, and thus recognition candidates may be selected by mistake. At this time, a “restatement utterance” that utters the same word as the previous time is performed.

Conventionally, when a re-utterance input is performed, the same recognition processing as in a normal utterance input is performed. In the conventional processing, it is not considered whether the re-utterance input is a paraphrase utterance or a rephrase utterance. And in the case of rephrasing, even if the speaker utters the same word repeatedly,
The same erroneous recognition result may be output every time.

In the speech recognition system described in Japanese Patent Application Laid-Open No. 1-161299, the same part of the recognition result for the re-uttered input and the recognition result for the previous uttered input is extracted, and a dictionary is extracted based on the same part. A search is performed. For example, if two characters of a four-character word are the same,
A dictionary search is performed based on the characters. However, in this system, words that have nothing to do with the two utterances are chosen because they contain the same part, which can lead to misrecognition. Further, even in the conventional system, it is not considered whether re-utterance input is re-utterance speech or paraphrase speech, and it cannot be adapted to the fact that the speaker does not always speak the same word twice in succession.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and has as its object to provide a speech capable of improving recognition ability by effectively using past recognition results when a re-utterance input is performed. It is to provide a recognition device and a method.

[0012]

(1) A speech recognition apparatus according to the present invention comprises: an acoustic processing means for generating an input speech pattern based on an uttered speech signal; and recognition of the input speech pattern and a standard pattern. Recognition processing means for selecting a recognition candidate for voice recognition by performing matching; restatement determination means for determining that the current input voice pattern is a pattern obtained by restatement of past utterances; When it is determined that rephrasing has been performed by the means, a recognition candidate adjusting means for adjusting the recognition candidate selected by the recognition processing means, and recognition candidate adjusting means, as the adjustment processing, An adjusted recognition candidate is determined using both the recognition candidate of the current recognition matching result and the recognition candidate of the past recognition matching result.

According to the present invention, when the speaker re-utters, the re-phrase determining means determines whether or not the current input voice pattern is a pattern obtained by re-phrase of past utterances. Is determined. In the case of utterance rephrasing, by referring to recognition candidates selected by recognition matching in past utterances, recognition candidates can be determined more accurately, and erroneous recognition can be reduced. On the other hand, if it is not a rephrase, the recognition candidate adjustment process is not performed. Therefore, it is possible to avoid useless processing of referring to past recognition results in a meaningless manner, and to avoid occurrence of erroneous recognition caused by the above-mentioned meaningless reference. As described above, according to the present invention, since it is determined whether or not the utterance is a utterance restatement, the past recognition candidates are used in the case of the utterance restatement. Can improve the recognition ability.

(2) Preferably, the recognition candidate adjusting means performs a predetermined calculation process based on a parameter representing a past and current recognition matching result to determine an adjusted recognition candidate. The parameter representing the recognition matching result is
For example, the ranking and similarity of recognition candidates. As is well known, the similarity is expressed by the distance and likelihood between the input voice pattern and the standard pattern, the recognition score based on the distance, and the like. According to the present invention, since the process of extracting the same part of the recognition candidate twice is not performed as in the related art, there is no erroneous recognition caused by such a process, and the past recognition result can be effectively used.

(3) In a preferred aspect of the present invention, the rephrasing determining means includes a voice pattern storing means for storing an input voice pattern of a past utterance, and a past input voice pattern and a present input voice pattern. And a similarity determining means for performing the matching between the input patterns to determine the similarity between the two input voice patterns.If the predetermined similarity is obtained, the input voice pattern of this time is obtained by rephrasing the utterance. It is determined that it is.

According to the present invention, when a re-utterance input is performed, a matching process for the present and past input voice patterns is performed, so that a determination as to whether or not the voice is a utterance is made. Will be Rather than comparing the input voice pattern with the standard pattern, the input voice patterns obtained by two utterances of the same speaker are compared with each other. It becomes possible.

(4) Preferably, the recognition matching by the recognition processing means is a process adapted to an unspecified speaker, for example, a hidden Markov model process. Preferably, the matching between the input patterns by the similarity determination means is processing adapted to a specific speaker, for example, dynamic programming processing.

This embodiment works particularly suitably when the present invention is applied to a speech recognition apparatus for an unspecified speaker. For recognition matching for selecting a recognition candidate, of course, processing suitable for an unspecified speaker is suitable. However, comparison targets in input pattern matching for rephrasing determination are two speech signals by the same speaker. A process suitable for a specific speaker is suitable for such a comparison target, and the similarity determination can be performed with high accuracy by adopting the process. As described above, according to the present invention, in the speech recognition apparatus for an unspecified speaker, the restatement determination is performed with high accuracy by partially using the matching process adapted to the specific speaker, thereby improving the recognition ability. Improvement can be achieved.

(5) In a preferred aspect of the present invention, when two or more repetitions are performed by the speaker, the similarity determination means determines whether or not each of the plurality of combinations of the voice patterns has been obtained. The similarity is obtained, and the recognition candidate adjusting means determines an adjusted recognition candidate based on a combination of voice patterns having the highest similarity.

According to the present invention, when two or more repetitions, that is, three or more utterances, are performed, the combination having the highest similarity is selected based on the result of pattern matching of a plurality of combinations. Then, a final recognition candidate is selected using the input voice pattern of the combination. Since the combinations that can obtain highly reliable recognition candidates are narrowed down, the recognition ability can be improved while reducing the processing load on the recognition candidate adjustment unit.

(6) A voice recognition device according to another aspect of the present invention is a device for performing voice recognition by performing recognition matching between an input voice pattern and a standard pattern, and performs input recognition of a past voice and current time of the input voice pattern. Is performed based on the similarity between the two utterances and the input utterance pattern, and based on the similarity between the two utterances, it is determined that the current input utterance pattern is a pattern obtained by re-uttering the past utterance. It has a correction determining means.

(7) In the voice recognition method of the present invention, an audio processing step of generating an input voice pattern based on a voice signal input as a voice, and a current input voice pattern is performed by rephrasing a past voice. A restatement determining step of determining that the pattern is an obtained pattern, a recognition matching step of performing recognition matching between the input voice pattern and the standard pattern, and a step of determining that the current utterance is not a restatement utterance. Determining the recognition candidate based on the recognition matching result of the above, and, when it is determined that the current utterance is a restatement utterance, the recognition candidate is determined using both the current recognition matching result and the past recognition matching result. Determining. According to this aspect, the above effects of the present invention can be obtained in the form of a method.

[0023]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention (hereinafter, referred to as embodiments) will be described below with reference to the drawings. In the present embodiment, the present invention is applied to a voice recognition device for an unspecified speaker. In the present embodiment, as an example, word recognition is performed.

FIG. 2 is a block diagram showing the configuration of the speech recognition apparatus of the present embodiment. As in the related art, when the speaker speaks toward the microphone 10, an input voice signal from the microphone 10 is converted into a digital signal by the A / D converter 12 and input to the acoustic processing unit 14. In the sound processing unit 14, sound analysis is performed, and an input sound pattern of a feature parameter (such as a cepstrum) is generated.

The generated input voice pattern is supplied to the rephrasing judgment section 20 and stored in the feature parameter registration section 22. The rephrasing determination unit 20 and the feature parameter registration unit 22 function as the rephrasing determination unit of the present invention. The feature parameter registration unit 22 is a storage unit such as a RAM, and stores an input voice pattern. The rewording determination unit 20 performs pattern matching based on the DP method, and compares the past input voice pattern stored in the feature parameter registration unit 22 with the current input voice pattern supplied from the rewording determination unit 20. Collate. By collation, 2
The degree of similarity between two input voice patterns (represented by the distance between them and the score based on the distance) is obtained. If the similarity is equal to or greater than a predetermined threshold value, it is determined that restatement has been performed this time. The rephrasing determining unit 20 supplies the rephrasing determination result to the recognition processing unit 30 together with the current input voice pattern. Note that the similarity may be supplied to the recognition processing unit 30 instead of the determination result, and the recognition processing unit 30 may determine the restatement.

The recognition processing section 30 performs pattern matching between the input voice pattern and the standard pattern. Here, the matching process using the HMM is performed. FIG.
As described with reference to FIG.
A plurality of word data to be recognized are stored. The input speech pattern and each word are individually matched, and the word with the best matching result is selected as the first recognition candidate. Further lower recognition candidates (words) are selected in descending order of matching. In the present embodiment, recognition candidates in a predetermined order, for example, up to tenth are selected. The selected recognition candidates are sent to the recognition candidate registration unit 32. The recognition candidate registration unit 32 is a storage unit such as a RAM, and stores recognition candidates for adjustment of recognition candidates when re-speaking is performed later. Along with the recognition candidates, the ranking of each recognition candidate and the similarity of the matching result are also stored. The recognition candidate registration unit 32 and the feature parameter registration unit 22 may be integrated.

Further, the recognition processing section 30 functions as recognition candidate adjusting means of the present invention. As described above, the recognition processing unit 30 receives the determination result as to whether or not the current utterance input is “restatement utterance” from the restatement determination unit 20. In the case of “if not a rephrasing utterance”, the recognition processing unit 30
Does not adjust the recognition candidates. Therefore, the recognition candidate selected by the current matching is output as a recognition result as it is. The output recognition result is presented to the speaker by talkback or display.

On the other hand, in the case of "restatement utterance",
The recognition processing unit 30 reads, from the recognition candidate registration unit 32, the recognition candidate selected in the past utterance input. And
An adjusted recognition candidate is determined by comprehensive judgment using both the past recognition candidate and the current recognition candidate. The adjusted recognition candidate is output as a recognition result. Various methods can be applied to the above-mentioned comprehensive judgment, and the following methods are exemplified.

(1) The ranking of each recognition candidate (word) is given in the current candidate set and the past candidate set.
Therefore, an appropriate operation is performed on the two ranks to obtain a comprehensive rank. For example, a sum or a product of two ranks is obtained, and the recognition candidates are rearranged in ascending order of the sum or the product.
An overall ranking is determined according to the order after rearrangement. At this time, it is needless to say that a sum or a product cannot be obtained for recognition candidates that exist only in one candidate set.

(2) Instead of the ranking of (1), an appropriate operation is performed on two similarities of each recognition candidate (word),
An overall similarity is required. The similarity here indicates how similar the input voice pattern and the standard pattern are, and is represented by likelihood and a score based on the likelihood. This similarity is calculated by pattern matching. Here, for example, the sum or product of the two scores is calculated, and the calculation result is used as the total score. The recognition candidate having the largest total score is selected as the first candidate after adjustment. Hereinafter, each candidate is re-ranked based on the total score. When calculating the sum or the product, it is also preferable to weight the two scores. For example, it is conceivable to give a higher weight to a higher score or to give a higher weight to the score of the recognition result this time.

In each of the above-described determination methods, the results of two matchings are used in the form of rank and similarity. That is, the adjusted recognition candidates are determined using the rank and the similarity, which are parameters representing the matching result. Even when a determination method other than the above is employed, processing based on the results of two matchings is performed.

In addition, when determining the adjusted recognition candidates,
It is preferable that the top candidate in the past matching result is excluded from the recognition candidates. It is clear that this candidate is not the correct answer. In addition, as the adjusted recognition candidates of the present invention, only the first recognition candidate may be obtained, and the second or lower recognition candidate may not be obtained.

FIG. 3 shows a change in the order of recognition candidates due to the above processing. Suppose the speaker said "raise". In the past matching, “raise” was the second highest recognition candidate, and the wrong candidate “lower” was ranked first. In this match, “raise” was selected as third place, and “lower” was again ranked first. At this time, since it is known that "lower" is not the correct answer, it is excluded from the recognition candidates. Although the correct answer "raise" is not the first place, a moderately high rank (similarity) is obtained by performing the matching twice. Since the answer is correct, it is natural that a higher ranking will be given. “Raise”, which has been ranked high twice, is selected as the top candidate after adjustment. In this way, by referring to the two recognition results, the order of the recognition candidates is switched, and the accuracy of the selection of the recognition candidates increases.

The configuration of the speech recognition apparatus according to the present embodiment has been described above. In the above configuration, in contrast to the case where the pattern matching in the recognition processing unit 30 is performed for the purpose of selecting a recognition candidate, the word matching determination unit 20 performs the pattern matching for the purpose of the wording determination. Corresponding to this difference in purpose, the processing performed in the two matchings is completely different. That is, the recognition processing unit 30 individually compares one input voice pattern with a number of standard patterns (all words to be recognized), and specifies a word as a recognition candidate. On the other hand, the rephrase determination unit 20 compares the two input voice patterns to determine the similarity, and here, it is not necessary to specify the recognition candidate. Therefore, the power required for the rephrasing determination calculation process may be much smaller than that of the recognition process, and therefore, the rephrasing determination can be performed in a short time with a simple process.

In addition, the rephrasing determination unit 20 determines whether the same speaker has uttered the same word. Same speaker,
When DP matching between input voice patterns is performed for the same word, remarkably high similarity can be obtained as compared with the case of different words. Therefore, the rephrasing determination unit 20 of the present embodiment can determine with high accuracy whether the rephrasing utterance has been performed. Furthermore, it is also preferable to simplify the processing such as matching in the rewording determination unit 20 within a range where sufficient determination accuracy can be obtained, by utilizing the fact that a remarkably high similarity is obtained at the time of rewording. Thus, the processing can be further speeded up.

In particular, in this embodiment, the rephrasing judging section 20
The DP method is adopted for the pattern matching in, and the HMM is adopted for the pattern matching in the recognition processing unit 30. The reason why the recognition processing unit 30 employs the HMM is that the target of the speech recognition apparatus of the present embodiment is an unspecified speaker, and the HMM is suitable for the speech recognition of the unspecified speaker. . HMM adopts the probability method,
Resistant to changes in input signals due to individual differences between speakers. on the other hand,
As described above, since the input speech patterns of the "same speaker" are compared in the rephrasing determination unit 20, pattern matching suitable for a specific speaker is suitable for this comparison processing. The determination accuracy is further improved by adopting. As described above, it is also preferable to simplify the processing of the DP method as long as sufficient accuracy is obtained and to increase the processing speed.

As described above, in the rephrasing determination process of the present embodiment, it is determined whether or not the same word is uttered by comparing two input voice patterns of the same speaker. It is not necessary to use HMM or the like for such processing. Based on the DP method, simple repetition determination is performed in a short time with simple processing, and a determination result necessary for determining whether or not to perform processing with reference to past recognition results in a later stage is obtained. Can be
Further, the DP method can be simplified to increase the processing speed.

Next, the speech recognition processing of this embodiment will be described with reference to the flowchart of FIG. The acoustic processing unit 14 performs an acoustic analysis on the input audio signal, and generates an input audio pattern of a characteristic pattern (S10). This input voice pattern is supplied to the rephrasing determination unit 20 and registered in the feature parameter registration unit 22 (S
12). The rephrase determination unit 20 reads the input voice pattern at the time of the previous utterance from the feature parameter registration unit 22. Then, the previous input voice pattern and the current input voice pattern are compared according to the DP method, the similarity between the two patterns is obtained (S14), and the similarity is compared with a predetermined value (threshold) (S16).

If the similarity is less than the predetermined value, it is determined that the present utterance is not a restatement utterance (S18). The result of this determination is sent to the recognition processing unit 30 together with the current input voice pattern. If the current utterance is the first utterance (if not a re-utterance), there is no input voice pattern to be compared in S16, the determination in S16 is NO, and the determination in S16 is NO.
Proceed to 18. The recognition processing unit 30 performs recognition processing of the input voice pattern, selects a recognition candidate (S20), and registers the recognition candidate in the recognition candidate registration unit 32 (S22). Then, recognition candidates are determined only from the current result (S2
4). That is, the recognition candidate selected in S20 is output as it is as a recognition result. If it is not determined that the recognition result needs to be corrected by talkback, etc.,
Subsequent processing is performed. For example, the recognition result is sent to another device connected to the speech recognition device.

On the other hand, if the similarity is equal to or more than the predetermined value in S16, it is determined that the present utterance is a restatement utterance (S26). The result of this determination is sent to the recognition processing unit 30 together with the current input voice pattern. Recognition processing unit 30
Performs an input voice pattern recognition process to select a recognition candidate (S28), and registers the recognition candidate in the recognition candidate registration unit 32 (S30). Then, the recognition processing unit 30 determines a final recognition candidate from both the previous recognition result and the current recognition result (S32). Here, the recognition candidate set obtained in the previous recognition processing is read out from the recognition candidate registration unit 32 together with the data on the ranking and similarity of each candidate. Then, as described above, comprehensive judgment is performed using the ranking and the similarity, and finally the adjusted recognition candidates are determined and output as the recognition result.

It is also preferable that the speech recognition device grasps whether or not the current utterance is a re-utterance as follows, for example. The recognition device notifies the speaker of the recognition result by talkback or display. The speaker knows the recognition result, determines the necessity of re-speaking, and operates an appropriate operation switch to notify the recognition device that re-speaking is to be performed. The recognizing device determines whether the subsequent re-utterance is a “paraphrase utterance” that utters a different word from the previous one or a “restate utterance” that utters the same word as before.

In the above description, the recognition process has been described on the assumption that the process is performed after re-speaking is performed after the first utterance. The process of FIG. 4 can be similarly applied when a re-speak is performed again after a certain re-speak. However, when re-speaking is performed continuously, it is more preferable to perform the following processing shown in FIG.

Here, it is assumed that N-1 utterances have already been performed (N-2 repetitions) and the Nth utterance has been performed. First, similarly to FIG. 4, the acoustic processing unit 14 performs an acoustic analysis on an input audio signal to generate an input audio pattern of a characteristic pattern (S50). This input voice pattern is supplied to the rewording determination unit 20 and registered in the feature parameter registration unit 22 (S5).
2). The rephrase determination unit 20 reads out the input voice pattern at the time of the past N-1 utterances from the feature parameter registration unit 22. Then, as shown in FIG. 6, each of the N-1 input voice patterns and the current input voice pattern are
The comparison is individually performed according to the DP method, and the similarity of N-1 is obtained (S54). Then, the maximum similarity is compared with a predetermined value (threshold) (S56).

If the maximum similarity is less than the predetermined value, it is determined that this utterance is not a restatement of any past utterance (S58). Therefore, the same processing as the processing after S20 in FIG. 4 is performed. The determination result of S56 is sent to the recognition processing unit 30 together with the current input voice pattern. The recognition processing unit 30 selects a recognition candidate from the current input data (S60) and registers it in the recognition candidate registration unit 32 (S62). Then, a recognition candidate is determined only from the current result (S64).

On the other hand, if the maximum similarity is equal to or more than the predetermined value, it is determined that the combination of the input voice pattern giving the maximum similarity and the current voice pattern is the optimum combination (S66). Then, the same processing as the processing from S28 onward in FIG. 4 is performed on the two audio patterns of the optimal set. The determination result of S56 is sent to the recognition processing unit 30 together with the current input voice pattern. The recognition processing unit 30
A recognition candidate is selected by performing recognition processing of the input voice pattern (S68), and the recognition candidate is registered in the recognition candidate registration unit 32 (S70). Then, the recognition processing unit 30 reads out from the recognition candidate registration unit 32 the recognition candidate set stored as the recognition result of the input speech pattern of the optimal set, together with the data on the order and the similarity. A final recognition candidate is determined from both the read past recognition result and the current recognition result (S72). At this time, the recognition candidates selected as the first place in the past N-1 recognition processes are all excluded from the recognition candidates.

In S54 of FIG. 5, the current input voice pattern is compared with the previous input voice pattern.
N-1 similarities were determined. On the other hand, in S54, a combination to be compared may be selected based on another criterion. For example, comparison of all combinations of _N C ₂ kinds may be performed.

The preferred embodiment of the present invention has been described above. As described above, according to the present embodiment, the past recognition candidates are effectively used based on the rephrasing determination result, so that the accuracy of the recognition at the time of re-inputting the voice is increased, thereby improving the recognition ability. can do. From the viewpoint of the speaker, it is possible to avoid a situation in which the same recognition is obtained no matter how many restatements are made.

In the rephrasing determination process of the present embodiment, two input voice patterns of the same speaker are compared with each other to determine whether the same word is uttered. Therefore, D
By adopting the P method or the like, highly accurate rephrasing determination can be performed in a short time, and the determination processing can be simplified to speed up the processing. In particular, although the recognition device of the present embodiment is for an unspecified speaker, a DP adapted to a specific speaker is partially used.
By applying the matching, the restatement determination is performed with high accuracy.

Furthermore, according to the present embodiment, as described with reference to FIG. 5, when re-speech is performed a plurality of times, the combination having the highest similarity is selected. Since narrowing down to combinations that can obtain highly reliable recognition candidates is performed, while reducing the processing load on the recognition processing unit 30,
The recognition ability can be improved.

Hereinafter, a modified example of this embodiment will be described.

(1) The recognition target of the speech recognition device is not limited to words. For example, anything, such as a sentence, a character, or a number, is acceptable.

(2) Recognition processing unit 30 and rewording determination unit 20
Is not limited to the HMM or DP method,
Any other technique may be applied. H in the rephrasing judgment unit 20
It is also possible to employ the MM or the DP method for the recognition processing unit 30. However, some effects described above cannot be obtained.

(3) The present invention may be applied to a voice recognition device for a specific speaker. The matching processing of the recognition processing unit 30 and the rewording determination unit 20 is appropriately changed according to the application.

(4) The function of each component shown in FIG. 2 may be realized by hardware or software.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a conventional voice recognition device.

FIG. 2 is a block diagram illustrating a configuration of a speech recognition device according to an embodiment of the present invention.

FIG. 3 is a diagram showing a change in the order of recognition candidates when a recognition candidate is determined based on the results of two previous recognitions;

FIG. 4 is a flowchart of a voice recognition process of the device of FIG. 2;

FIG. 5 is a flowchart of a speech recognition process of the apparatus of FIG. 2, which is suitable for a case where re-speaking is performed a plurality of times.

FIG. 6 is a diagram showing a combination of a plurality of similarity determinations of an input voice pattern in the process of FIG. 5;

[Explanation of symbols]

14 acoustic processing unit, 18 recognition dictionary storage unit, 20 rewording determination unit, 22 feature parameter registration unit, 30 recognition processing unit, 32 recognition candidate registration unit.

Claims (8)

[Claims] 1. An audio processing unit for generating an input voice pattern based on a voice signal input and uttered, and a recognition processing unit for performing recognition matching between the input voice pattern and a standard pattern to select a recognition candidate for voice recognition. And a restatement judging means for judging that the input voice pattern of this time is a pattern obtained by restatement of past utterances, and when it is judged that the restatement utterance is performed by the judgment means, And a recognition candidate adjustment unit that performs an adjustment process of the recognition candidate selected by the recognition processing unit. The recognition candidate adjustment unit includes, as the adjustment process, a recognition candidate of a current recognition matching result and a recognition candidate of a past recognition matching result. A speech recognition device, wherein an adjusted recognition candidate is determined using both of the recognition candidates. 2. The apparatus according to claim 1, wherein the recognition candidate adjustment unit performs a predetermined calculation process based on a parameter representing a past and current recognition matching result to determine an adjusted recognition candidate. A speech recognition device characterized by the following. 3. The apparatus according to claim 1, wherein the rephrasing determining unit includes: a voice pattern storing unit that stores an input voice pattern of a past utterance; And a similarity determining unit that performs matching between the input voice patterns with the input voice pattern to determine the similarity between the two input voice patterns. If the predetermined similarity is obtained, the input voice pattern of this time is rephrased. A speech recognition device for determining that the speech is obtained by utterance. 4. The apparatus according to claim 3, wherein the recognition matching in the recognition processing unit is a process adapted to an unspecified speaker, and the matching between the input patterns in the similarity determination unit is: A speech recognition device characterized by processing adapted to a specific speaker. 5. The apparatus according to claim 3, wherein the recognition matching in the recognition processing means is Hidden Markov model processing, and the matching between the input patterns in the similarity determination means is dynamic programming processing. A speech recognition device, characterized in that: 6. The apparatus according to claim 3, wherein when the speaker performs re-speaking more than once, the similarity determining unit determines a plurality of combinations of the voice patterns. A speech recognition apparatus, wherein a similarity is obtained for each of them, and the recognition candidate adjusting means determines an adjusted recognition candidate based on a combination of speech patterns having the highest similarity. 7. A voice recognition apparatus for performing voice recognition by performing recognition matching between an input voice pattern and a standard pattern, wherein matching between input patterns of an input voice pattern of a past voice and an input voice pattern of a current voice is performed. The input voice pattern based on the similarity, and based on the similarity, determine that the current input voice pattern is a pattern obtained by rephrasing the past voice. Characteristic speech recognition device. 8. An audio processing step of generating an input voice pattern based on a voice signal input by utterance, and judging that the current input voice pattern is a pattern obtained by restatement of past utterances. A rephrasing determination step, a recognition matching step of performing recognition matching between the input voice pattern and the standard pattern, and if it is determined that the present utterance is not a rephrase utterance,
When the recognition candidate is determined based on the recognition matching result of this time, and when the utterance of this time is determined to be a restatement utterance,
Determining a recognition candidate using both the current recognition matching result and the past recognition matching result.