JPH09311694A - Speech recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a word spotting device with less erroneous detection and a means thereof. SOLUTION: By a likelihood computing unit 104, frame synchronous DP matching based on a model, to which an optional syllable chain is added before/ after a proposed word, is carried out, and the result is outputted to a detection unit 105. By an intermediate likelihood computing unit 108, synchronous DP matching based on a model, to which an optional syllable chain is added before the proposed word, is carried out, and the intermediate likelihood of each proposed word is outputted to the detection unit 105. In the detection unit 105, if an optimum word outputted from the likelihood computing unit 104 is provided with likelihood of the threshold value or more, an opposite proposed word with an intermediate likelihood of a fixed value or more is searched from outputs from the intermediate likelihood computing unit 108, and if this kind of word is detected, detection of the optimum word is retained. The detection is established when all of the opposite proposed words are eliminated. Alternately, disposal is carried out when either of the opposite proposed words is detected.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for detecting a specific word in utterance.

[0002]

2. Description of the Related Art Word detection, as compared with ordinary speech recognition,
It has an advantage that it is not necessary to determine the existing section of the word in the input voice in advance.

[0003] Conventionally, as a method for detecting a specific word in utterance, Reference 1 (Journal of theology (D) J67-D, 11p.
242-1249) is known. In the method described in this paper, if the likelihood calculated independently for each frame candidate word exceeds a certain threshold, it is detected regardless of the detection of other candidate words. Further, as another method, a method described in Document 2 (“Speech recognition by probabilistic model”, Seiichi Nakagawa p.20-26) is known. In the method of this document, the optimum word in the language model is determined and output for each frame. In addition, as a method for suppressing the detection of words that temporally overlap with each other, Reference 3 (Science Technical Report SP95-7
7p. 31-38) are known. According to the method of this document, it is possible to determine the provisional detection of each candidate word in each frame, and suppress the detection of temporal overlap using the frame synchronization processing from the likelihood of each detection.

[0004]

However, the method of Reference 1 allows a hypothesis that candidate words are temporally overlapped during utterance. Although such a phenomenon does not occur in the method of Document 2, a phenomenon may occur in which a word similar to a part of the uttered word is detected in the middle of utterance. In order to prevent this, it is necessary to somehow determine the range of time in which the words exist in the input voice. The method of Reference 3 does not require such a case, and does not allow the hypothesis that candidate words temporally overlap during utterance. However,
Only in front of the recognition target word, as shown in Reference 3,
When using a model that accepts an arbitrary syllable sequence (garbage model), estimate the likelihood of a word similar to a part of the uttered word higher than the likelihood of the actually uttered word. As a result, a similar word may be detected and the detection of the actually uttered word may be canceled (partial matching), and a method of improving it may not be obvious. However, when a model in which a garbage model is added before and after the recognition target word is used and the method shown in Reference 2 is used, partial matching is avoided by setting the pattern matching score of the garbage model low. It is known that you can.

It is an object of the present invention to use a method for detecting a plurality of candidates for utterance of one word as described above for word detection, and a method for retaining robustness against partial matching of Document 2. Then, there is a need to provide a word recognition method that has the above-mentioned advantages of the methods described in Documents 1 to 3.

[0006]

According to the present invention, in a voice recognition device for detecting a candidate word from input voice data on a frame-by-frame basis using a likelihood as a reference, a plurality of candidate words in one word is detected. It relates to a method for reducing.

[0007]

BEST MODE FOR CARRYING OUT THE INVENTION

(First Embodiment) FIG. 1 is a block diagram showing the arrangement of a speech recognition apparatus according to the first embodiment of the present invention.

The speech recognition apparatus according to the first aspect of the present invention includes a speech analysis unit 101 for extracting a feature amount for each constant time (hereinafter, referred to as a frame) from an input speech, and a word dictionary 102 storing candidate words. A model generation unit 103 that generates a language model from candidate words in a word dictionary; an optimal word sequence (hereinafter referred to as an optimal sequence) that applies to the language model in each frame based on the feature amount and the language model; Likelihood calculation section 104 for obtaining the likelihood, detection section 105 for outputting a candidate word using the output from likelihood calculation section 104 as an input, and storage section 106 associated therewith.
Consists of.

The voice analysis unit 101 analyzes the frequency of each frame of the input voice and generates a feature quantity vector (hereinafter referred to as a feature quantity) for each frame. As the element of the feature amount, power, power change amount, mel cepstrum, mel cepstrum change amount, mel cepstrum secondary change amount, or the like is used. The feature amount for each frame is output to the likelihood calculation unit 104 for each frame.

The word dictionary 102 forms a chain of unit acoustic models of a word to be recognized, for example, a chain of acoustic models representing syllables that form a word (“O” in the case of “Osaka”). It is stored in the form of "o"-"sa"-"ka").

In the model generation unit 103, the word dictionary 102
We construct a language model that accepts utterances with arbitrary syllable chains added before and after each word model in. The likelihood calculation unit 104 calculates the optimal sequence of each frame and its likelihood from the language model generated by the model generation unit 103 and the feature amount of each frame output from the speech analysis unit 101,
Output to the detection unit 105.

The storage unit 106 stores a predetermined threshold value and candidate words output from the detection unit 105. The detection unit 105 receives the optimum sequence output from the likelihood calculation unit 104 and its likelihood for each frame, determines a candidate word and determines the output, and outputs the candidate word if necessary.

A voice recognition device according to a first aspect is a voice recognition device according to claims 1, 2, and 5.

(Description of Operation of First Mode) FIG. 2 is a diagram showing an operation of the speech recognition apparatus of the first mode of the present invention.

In the voice analysis unit 101, frequency analysis is performed for each frame of the input voice to generate a feature quantity,
Each frame is output to the likelihood calculating unit 104. In the likelihood calculation unit 104, the automaton control O shown in Document 2 is performed.
The ne Pass DP method is used to perform pattern matching between the language model generated by the model generation unit 103 and the feature amount in each frame output from the speech analysis unit 101, and the optimal sequence of each frame and its The likelihood is calculated, and the optimum column and its likelihood are output to the detection unit 105.

The detection unit 105 first compares the likelihood output from the likelihood calculation unit 104 with the threshold value stored in the storage unit 106 for each frame (step 1). If the corresponding likelihood is equal to or more than the threshold value, the storage unit 1 stores the candidate word (hereinafter, referred to as an optimum word) in the optimum sequence output from the likelihood calculation unit 104.
The candidate word stored in 06 is compared, and if different, the optimum word is output (step 2).

Next, if the corresponding likelihood is greater than or equal to the threshold value, the optimum word is stored in the storage unit 106 (step 3A). If it is less than the threshold value, an empty word is stored in the storage unit 106 (step 3B).

Next, the effect of the first embodiment of the present invention will be described. The method shown in Document 2 shows a means for obtaining the optimum sequence and the likelihood for each frame, but if these are directly used as the recognition result of the word speech, the recognition result is output for each frame, which is inconvenient. The first form provides a means for extracting a single recognized word from the optimal sequence and likelihood for each frame.

(Example of First Embodiment) Next, the operation of one example of the first embodiment of the present invention will be described in detail.

In this embodiment, the model shown in FIG. 3 is used as the language model. The model of FIG. 3 is configured in the form of a concatenation of three models of forward garbage that accepts an arbitrary syllable sequence including an empty sequence, candidate words, and backward garbage that accepts an arbitrary syllable sequence including an empty sequence, All utterances that pass through each model in this order, that is, utterances in which an arbitrary syllable string is added before and after one candidate word, are all accepted. It is assumed that the model generation unit 103 previously creates and stores the model of FIG. 3 from each candidate word of the word dictionary 102. It is assumed that the contents of the word dictionary 102 are three words, “Seki” (cough), “Hekinan”, and “Naha”. The likelihood calculation unit 104 performs pattern matching between the feature quantity sequence from the first frame to the corresponding frame and the language model of FIG. 3 for each frame by the method described in Document 2 to determine the optimum sequence on the relevant language model. Calculate the likelihood. However, the likelihood uses the natural logarithm of the probability value in the final state of the corresponding language model. Further, the storage unit 106 stores threshold value -100.0 in advance, and stores an empty word as an initial value of a candidate word. In addition, in this embodiment, the frame interval is set to 10 ms.

As shown in FIG. 4, it is assumed that the utterance "Hekinan" is made. The detection unit 105 receives the optimal sequence output from the likelihood calculation unit 104 and its likelihood for each frame. The candidate words included in the optimum sequence are shown in FIG.
It is assumed that the series is as shown in.

In the first frame, first, in step 1, likelihood -200.0 is compared with threshold -100.0. Since the likelihood is less than the threshold, nothing is done in step 2. Then, in step 3B, the empty word is stored in the storage unit 106.

After that, since the likelihoods are all less than the threshold value up to the 120th frame, only the operation of storing in the storage unit 106 is performed in step 3B as in the first frame.

In the 121st frame, step 1
In, the likelihood -99.0 and the threshold -100.0 are compared. Since the likelihood is greater than or equal to the threshold value, in step 2, the optimum word “Hekinan” in the 121st frame is compared with the candidate word (empty word) stored in the storage unit 106, and “Hekinan” is output because it is different. To do. Then, in step 3A, the optimum word "Binan" is stored in the storage unit 106.

Since the likelihoods after the 122nd frame are all above the threshold, step 1 is performed in the same manner as in the 121st frame, and in step 2, the optimum word "Hekinan" and the candidate word stored in the storage unit 106 are stored. The comparison of "Binan" is performed, and since they are the same, the optimum word is not output. In step 3A, the optimum word “Hekinan” is stored in the storage unit 106.

(Second Embodiment) FIG. 6 is a diagram showing the structure of a speech recognition apparatus according to a second embodiment of the present invention.

The speech recognition apparatus according to the second embodiment of the present invention has a fixed time (hereinafter, referred to as a frame) from the input speech.
A speech analysis unit 101 that extracts a feature amount for each of the words, a word dictionary 102 that stores words, a model generation unit 103 that generates a language model from the words in the word dictionary 102, and a feature amount and the language model. , A likelihood calculation unit 104 for obtaining an optimal word sequence (hereinafter referred to as an optimal sequence) that fits the language model in each frame and its likelihood,
An intermediate model generation unit 107 that generates an intermediate model from the words in the word dictionary 102; an intermediate likelihood calculation unit 108 that calculates the likelihood of the intermediate model in each frame from the feature amount and the intermediate model; The detection unit 105 and the storage unit 106 associated with the detection unit 105 output the candidate words with the outputs from the degree calculation unit 104 and the intermediate likelihood calculation unit 108 as inputs.

The voice analysis unit 101, the word dictionary 102, the model generation unit 103, and the likelihood calculation unit 104 are the same as those described in the first embodiment.

The intermediate model generation unit 107 uses the word dictionary 10
For all the first half subsequences of each word model in 2, a language model (intermediate model) that accepts utterances with an arbitrary syllable chain added immediately before is generated.

The intermediate likelihood calculation unit 108 performs pattern matching of all the intermediate models generated by the intermediate model generation unit 107 and the feature amount for each frame output from the speech analysis unit 101, and each intermediate model for each frame. Compute the likelihood of. Then, for each candidate word, the maximum likelihood (hereinafter referred to as the intermediate likelihood) of the likelihoods of the intermediate model of the relevant candidate word is obtained and output to the detection unit 105.

The storage unit 106 stores a predetermined threshold value and a predetermined preliminary threshold value. It also stores candidate words, pending optimal words, and their conflict candidate word list.

The detection unit 105 receives the output of the likelihood calculation unit 104 and the output of the intermediate likelihood calculation unit 108, and stores them in the storage unit 1.
The output word is determined and output using the information in 06.

The voice recognition device according to the second embodiment realizes a voice recognition device corresponding to claims 1, 2, 3, 6, 7, and 14.

(Explanation of the Operation of the Second Embodiment) FIG.
[Fig. 8] is a diagram showing an operation of the voice recognition device in the second exemplary embodiment of the present invention. The voice analysis unit 101, the word dictionary 102, the model generation unit 103, and the likelihood calculation unit 104 perform the same operations as in the first embodiment.

First, step 1 of the first embodiment is executed. If the relevant likelihood is greater than or equal to the threshold, the likelihood calculation unit 10
The optimum word output from No. 4 is compared with the candidate word stored in the storage unit 106 (step 2). If the likelihood is less than the threshold value, an empty word is stored in the storage unit 106 (step 3B). If the likelihood is equal to or more than the threshold value, the optimum word is stored in the storage unit 106 (step 3A). When the likelihood is equal to or higher than the threshold and the optimum word is different from the candidate word stored in the storage unit 106, the optimum word and the candidate word whose midway likelihood is equal to or higher than the preliminary threshold in the current frame (hereinafter, All the lists)
If there is a past conflict candidate list stored in the storage unit 106 (step 4) and containing the optimum word of the current frame in the past candidate list stored in the storage unit 106, information on the list and the corresponding past optimum word is displayed. It is erased from the storage unit 106 (step 4-2). As post-processing, all the past candidate candidate words stored in the storage unit 106 are removed from the past candidate whose intermediate likelihood is below the preliminary threshold in the current frame (step 5). Next, among the optimum words stored in the storage unit 106, the optimum candidate word list is output, and the corresponding information is deleted from the storage unit 106 (step 6). However, here, the candidate words whose intermediate likelihoods are compared with the preliminary threshold are not all candidate words,
It is often preferable to exclude the best word in the current frame.

Next, the effect of the second embodiment of the present invention will be described.

In the first embodiment, a word similar to a part of a certain candidate word, for example, when "Hekinan" (hekinan) is uttered, "seki" (seki) similar to "hekinan" of "hekinan" is used. )
May be detected. In such a case, the second embodiment suspends the detection of “Seki” (step 4) when it can be estimated that it is in the middle of another word, that is, “Hekinan”, and thereafter, Upon completion, detection of "Seki" is stopped (step 4-2). When the likelihood in the middle of "Hekinan" is sufficiently reduced, the estimation that it is in the middle of "Hekinan" is discarded (step 5), and "Seki" that has been held is detected (step 6). In this way, detection of subsequence similar words can be prevented. In addition, the output of the optimum word is immediately performed in the frame in which it is estimated that it is in the middle of other candidate words, so that the delay time caused by the detection of the subsequence similar word is the minimum necessary.

(Example of Second Embodiment) Next, the operation of one example of the second embodiment of the present invention will be described in detail.

In this embodiment, the same model shown in FIG. 3 as that used in the first embodiment is used as the language model. Word dictionary 102, likelihood calculation unit 104, storage unit 106
The same numerical values, likelihood calculation methods, and frame intervals stored in advance are used as in the first embodiment.

The midway model is composed of a forward garbage that accepts an arbitrary syllable sequence including an empty sequence and a concatenation of two models of candidate words, as shown in FIG. That is, all utterances in which an arbitrary syllable string is added before and after one candidate word are accepted. Figure 8 for each frame
Likelihood of all states in the middle of a word in the model of
In the case of the syllable 1 state, the midway likelihood can be obtained by extracting the likelihood of the state shown by the black dots in FIG.
It is assumed that the midway model generation unit 108 has created and memorized the midway model in FIG. 8 from each candidate word in the word dictionary 102. However, since the intermediate model is a partial model of the language model shown in FIG. 3 and the pattern matching method is the same, the intermediate likelihood can be reduced by appropriately extracting the likelihood of the intermediate state of the language model of FIG. can get. Therefore, the present embodiment can be realized by a method in which the model of FIG. 3 is shared by the language model and the intermediate model, and the independent intermediate model generation unit 107 and the intermediate likelihood calculation unit 108 are not provided.

In this embodiment, the preliminary threshold value is set to -101.5. Since the preliminary threshold is for selecting candidate words that may be in the middle of the current frame, the lower it is, the more the number of conflict candidate words increases and the possibility of omission of estimation in the middle of words decreases. The delay time of word output becomes long. Normally, select a value smaller than the threshold. In this embodiment, the opposition candidate word does not include the optimum word of the corresponding frame.

A specific example of the operation of the second embodiment will be shown below.

As shown in FIG. 9, it is assumed that the utterance "Hekinan" is made. The detection unit 105 receives the optimal sequence output from the likelihood calculation unit 104 and its likelihood for each frame. The candidate words included in the optimum sequence are shown in FIG. 1 for each frame.
It is assumed that the word is 0.

In the first frame, first, in step 1, the likelihood -200.0 is compared with the threshold -100.0. In step 3B, the empty word is stored in the storage unit 106. Since the storage unit 106 does not store the conflict candidate list corresponding to the optimum word, nothing is done in the subsequent steps. Thereafter, up to the 90th frame, the likelihoods are all less than the threshold, and the storage unit 106 does not store the set of the optimum word and the confrontation candidate list. Only the operation of storing in storage unit 106 (step 2) is performed.

In the 91st frame, in step 1, the likelihood -99.0 and the threshold -100.0 are compared. Since the likelihood is greater than or equal to the threshold, in step 2 the ninth
The optimum word “Seki” of one frame is stored in the storage unit 106 (step 3A). Next, a list of confrontation candidate words (in this case, only "Hekinan") is created, the optimum word "Seki" is stored in the storage unit 106, and a list including only "Hekinan" is stored (step 4).

In the 91st frame, since the past optimum word is not stored in the storage unit 106, nothing is done in the subsequent steps.

From the 92nd frame to the 120th frame, the optimum word remains unchanged as "Seki". Therefore, from Step 1 to Step 4, only the operation of writing "Seki" in the storage unit 106 is performed. In step 4-2, since the optimum word "Seki" of the current frame does not exist in the past conflict candidate list, nothing is done.

In step 5, nothing is done because the intermediate likelihood of "Binan", which is an element of the conflict candidate word list, is greater than or equal to the preliminary threshold. In step 6, since the list includes "Hekinan", nothing is done.

In the 121st frame, the optimum word is changed to "Hekinan". In steps 1 and 2, “Hekinan” is stored in the storage unit 106. In steps 3 to 4, a candidate candidate word list including only the optimum word “Hekinan” and “Naha” as an element is stored in the storage unit 106.

In step 4-2, since the conflict candidate word list of the past optimum word “Seki” includes the optimum word “Hekinan” of the 121st frame, the past optimum word “Seki” and the optimum word “Seki” are stored in the storage unit 106. The conflict candidate word list is deleted. In step 5 and step 6, since no corresponding conflict candidate word list exists, nothing is done.

From the 122nd frame to the 140th frame, there is no change in the optimum word and the conflict candidate word.
Only the operation of updating the optimum word in the storage unit 106 is performed.

Since there is no change in the optimum word in the 141st frame, steps 1 to 4-2 move similarly to the immediately preceding frame. In step 5,
Since the likelihood of “Naha” in the middle is below the preliminary threshold, it is removed from the past candidate word list of the optimum word “Binan”. As a result, the conflict candidate word list of "Hekinan" is emptied, so that "Hekinan" is detected in step 6.
In this way, the substring-similar word "Seki" of "Hekinan" is not output, and the spoken word "Hekinan" is output as soon as the halfway estimation of the conflict candidate word "Naha" is discarded.

(Modification of Second Embodiment) As a modification of the second embodiment, in step 4 of the detection unit 105, the number of the current frame is stored together with the storage unit 106.
In step 6, in addition to the empty candidate word list, the operation in which the frame number stored in the storage unit 106 is past the current frame by a certain time or more is output. Conceivable. By adopting such an operation, the delay time from the utterance of a candidate word to its output can be forcibly suppressed within a certain time, which is effective in realizing a voice recognition device that wants a recognition result as quickly as possible. .

This modification is defined in claims 1, 2, 6, 7, and 1.
A voice recognition device corresponding to 3 and 14 is realized.

Further, as a further modified example, in step 6 of the modified example described above, an operation of outputting the corresponding optimum word when the frame number stored in the storage unit 106 is a certain number or more can be considered. This is effective as a device that accepts only input within a fixed time after the voice recognition device is activated.

This modification is defined in claims 1, 2, 6, 7, and 1.
A voice recognition device corresponding to 0 and 14 is realized.

Further, as another modification of the second embodiment, in step 6 in the detection unit 105, not only the conflict candidate word list is empty, but also in the corresponding frame, an external input device, that is, a keyboard is used. When receiving an input signal from a switch, another voice recognition device, or the like, an operation of outputting the optimum word stored in the storage device 106 can be considered. As a result, the speaker determines the recognition timing to limit the word existence section,
It can be expected that the recognition performance will be improved. For example, it is effective when the speaker can be requested to perform an operation such as pressing a button after utterance.

This modification is defined in claims 1, 2, 6, 7, and 1.
A voice recognition device corresponding to 1 and 14 is realized.

In this case, as an external input device,
It is conceivable to use a voice detection device. As the voice detecting device, for example, a device described in Japanese Patent Laid-Open No. 7-225592 is used, but any device may be used as long as it determines a vocal section in the input voice. Since voice detection is considered to determine the beginning and end of the voice section in conjunction with the speaker's utterance, a voice recognition device that does not require extra action to the speaker can be used by using a voice detection device instead of a button or the like. realizable. In this case, in consideration of the case where the voice section end detection by the voice detection fails, the detection unit 105
In step 4, the current frame number is stored together in the storage unit 106, and in step 6,
Even if there is no signal from the voice detecting device, it is possible to perform an operation of outputting the frame number stored in the storage unit 106 that is past the current frame by a certain time or more. With this configuration, claims 1, 2,
A voice recognition device corresponding to 6, 7, or 17 can be realized. In this case, the output condition adopted in claim 17 is the claim 1
2 and 14 are output conditions.

Further, as a modification of the second embodiment, after the opposition candidate word list is created in step 4 of the detection unit 105, the optimum words do not change and the end of the optimum sequence ends with the corresponding optimum word. In the meantime, it is conceivable that the operation of adding all candidate words whose intermediate likelihood has reached the preliminary threshold in the corresponding frame to the list to more reliably list the confrontation candidate words. With the addition of this operation, it becomes possible to adopt the conflicting candidate word whose likelihood is low in the frame in which the optimum word first appears without missing from the list.

This modified example is defined by claim 1, 2, 4, 6,
A voice recognition device corresponding to 8, 10, and 14 is realized. Note that this is an example in which the fixed time after the detection section in claim 8 is 0 ms.

In addition, in this case, after the opposition candidate word list is created in step 4 of the detection unit 105, the optimum words do not change and the end of the optimum sequence ends with the corresponding optimum word, and a fixed time thereafter. For example, during a period of about 200 ms, an operation is performed in which all candidate words whose intermediate likelihood has reached the preliminary threshold in the corresponding frame are added to the list, and the intermediate likelihoods of all conflict candidate words in the list fall below the preliminary threshold. It can also be designed to perform output operation for the first time. This means, for example, that multiple candidate words are 200 ms
This is a method of reducing false positives by using the assumption that it will not be continuously spoken within. This is claimed in claims 1, 2, 4, 6, 8, 10, 1
A voice recognition device corresponding to 7 is realized. This is an example in which the fixed time after the detection section in claim 8 is 200 ms.

Further, as a modification of the second embodiment,
It is conceivable that the conflict candidate word list stored in the storage unit 106 in step 4 of the detection unit 105 is always a list of all candidate words. In this case, after the detection of the word, the detection is suspended as long as the intermediate likelihood of some word exceeds the preliminary threshold. Also, when the optimum word is replaced, it is always included in the conflict candidate word list, so that detection of the optimum word in the past is abandoned. According to this modification as well, not only detection of overlapping words in time but also detection of proximity in time can be suppressed.

This modified example is defined by claim 1, 2, 4, 6,
It is a voice recognition device corresponding to 8 and 16. Note that this is an example in which the fixed time after the detection section in claim 8 is 0 ms.

As a further modification of this modification,
It is conceivable to add a voice detection device as an external input device and add a condition that the voice end detection signal from the corresponding voice detection device is given to the output condition to make the output condition more strict. This makes it possible to enhance the detection hold by using information on both temporal proximity and the possibility of other candidate words being spoken.

This modified example is defined by claim 1, 2, 4, 6,
The output conditions in claims 18 and 18 correspond to the output conditions in claims 12 and 16. This is also an example in which the constant time after the detection section in claim 8 is 0 ms.

[0067]

As described above, according to the present invention, it is possible to reduce the detection of overlapping words and the detection of words that are too close to each other, and to realize a voice recognition device with less erroneous detection with minimum delay. can do.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a voice recognition device according to a first embodiment of this invention.

FIG. 2 is a flowchart showing an operation of the voice recognition device in the first exemplary embodiment of the present invention.

FIG. 3 is a network showing a language model used in the speech recognition apparatus according to the first embodiment of the present invention.

FIG. 4 is a speech recognition device 1 according to the first embodiment of the present invention.
It is a figure showing the change with time of the optimal word and likelihood in an Example.

FIG. 5 is a speech recognition device 1 according to the first embodiment of the present invention.
7 is a table showing optimal words and likelihoods for each frame in the example.

FIG. 6 is a configuration diagram of a voice recognition device according to a second embodiment of the present invention.

FIG. 7 is a flowchart showing the operation of the voice recognition device in the second exemplary embodiment of the present invention.

FIG. 8 is a network showing a language model used in the speech recognition apparatus according to the second embodiment of the present invention.

FIG. 9 is a voice recognition device 1 according to a second embodiment of the present invention.
It is a figure showing the time-dependent change of the optimal word, likelihood, and halfway likelihood in an Example.

FIG. 10 is a table showing optimal words, likelihoods, and conflict candidate words for each frame in one example of the speech recognition apparatus according to the second exemplary embodiment of the present invention.

Claims (18)

[Claims] 1. A voice analysis unit for extracting a feature amount for each frame from a frequency analysis of an input voice for each fixed time (frame), and a voice model for accepting an arbitrary syllable string before and after a candidate word. A likelihood calculation unit that performs pattern matching between the language model and the feature amount for each frame, selects an optimal word sequence (optimal sequence) on the language model for each frame, and calculates the likelihood thereof; A detection unit that determines and outputs the optimum word for each frame based on the optimum sequence and the likelihood, and when the optimum word continues one or more times, the optimum word within the corresponding time (detection section) is at most A voice recognition device characterized by outputting once. 2. A voice analysis unit for extracting a feature amount for each frame by frequency analysis of an input voice for each fixed time (frame), and receiving an arbitrary syllable string before and after a predetermined candidate word. Likelihood calculation for performing pattern matching between the language model added with a speech model and the feature quantity for each frame, selecting an optimal word sequence (optimal sequence) on the language model for each frame, and calculating the likelihood thereof. Unit, a detection unit that determines and outputs the optimum word for each frame based on the optimum sequence and the likelihood, and when the optimum word continues one or more times, the corresponding time (detection section)
A voice recognition device characterized in that when the likelihood reaches a predetermined constant value (threshold value) even once, the optimum word in the detection section is output at most once. 3. The voice recognition device according to claim 1, wherein the candidate word included in the optimum sequence is an optimum word. 4. The optimum word is output once as the optimum word if the last word of the optimum sequence is the candidate word, and the optimum word is not output otherwise. The voice recognition device according to item 1 or 2. 5. The voice recognition device according to claim 1, wherein the time at which the optimum word is output is the first time of the detection section. 6. The voice recognition device according to claim 1, wherein the time to output the optimum word is the first time when a predetermined condition (output condition) is satisfied. . 7. A language model (intermediate model) for accepting a utterance in which an arbitrary syllable sequence is added immediately before all the first half subsequences of all the candidate words, and in each of the frames, all the intermediate models A time when the optimal word should be detected, which is provided with a device that calculates the likelihood of pattern matching of the feature quantity and calculates the maximum likelihood (halfway likelihood) of all the halfway models of each candidate word. In the above, output is suspended only when there is the candidate word (opposite candidate word) whose intermediate likelihood has reached a predetermined constant value (preliminary threshold value), and thereafter, the predetermined condition (output condition) is The voice recognition device according to claim 1, wherein the voice recognition device outputs at the first time when it is satisfied. 8. A language model (intermediate model) for accepting a utterance in which an arbitrary syllable sequence is added immediately before all the first half subsequences of all of the candidate words is provided, and all intermediate models and the intermediate model are provided in each frame. A device for calculating the likelihood of pattern matching of the feature quantity and calculating the maximum likelihood (midway likelihood) of all the midway models of the candidate word, and the midway likelihood within the detection section. Output is suspended only when there is the candidate word (opposite candidate word) that has reached a predetermined constant value (preliminary threshold value) even once, and thereafter, the first condition in which the predetermined condition (output condition) is satisfied Outputting at time, Claim 1, 2, 3
Alternatively, the voice recognition device according to item 4. 9. A language model (intermediate model) for accepting a utterance in which an arbitrary syllable sequence is added immediately before all the first half subsequences of all of the candidate words is provided, and all intermediate models and the intermediate model are provided in each frame. A likelihood is calculated by performing pattern matching of feature amounts, and a device that calculates the maximum likelihood (intermediate likelihood) of all intermediate models of the candidate word is provided, and the intermediate likelihood is within the detection section and Output is suspended only when there is a candidate word (opposite candidate word) that has reached a predetermined constant value (preliminary threshold value) in all frames within a predetermined time period thereafter, and thereafter, in advance. The output is performed at the first time when the defined condition (output condition) is satisfied.
Alternatively, the voice recognition device according to item 4. 10. The output condition is that there is no one of the conflict candidate words that has become the optimum word since the detection time in the frame in which a predetermined time has elapsed after the device is activated. The voice recognition device according to Item 7, 8 or 9. 11. An output device comprising an input device for receiving an input from the outside, wherein the output condition is that a signal from the input device is given after a detection time. The voice recognition device described. 12. The input device is a voice detection device for determining the range of a voice section, and the output condition is that a signal indicating the end of the voice section is given from the voice detection device. The voice recognition device according to claim 11. 13. The output condition is that there is no one of the conflict candidate words that has become the optimum word since the detection time in the frame in which a certain time has elapsed after the detection time. The voice recognition device according to 7, 8, or 9. 14. After the detection time, each of the intermediate likelihoods of all the conflict candidate words once falls below the preliminary threshold, and the corresponding conflict candidate word has never been the optimum word since the detection time. 10. The voice recognition device according to claim 7, 8 or 9, wherein the output condition is set as follows. 15. In a certain frame after the detection time, the intermediate likelihoods of all the conflict candidate words are lower than the preliminary threshold, and the conflict candidate word is the optimum word even after the detection time. 10. The voice recognition device according to claim 7, wherein the output condition is that the voice recognition is not performed. 16. In a certain frame after the detection time, the intermediate likelihoods of all the other candidate words are lower than the preliminary threshold, and the other candidate words are the optimum words even once after the detection time. 10. The voice recognition device according to claim 7, wherein the output condition is that the condition is not satisfied. 17. The output condition is that any one of a plurality of the output conditions selected in advance from the output conditions according to claim 10, 11, 12, 13, 14 or 15 is satisfied. The voice recognition device according to claim 7, 8, or 9. 18. A method according to claim 10, 11, 12, 13, 14,
10. The voice recognition device according to claim 7, wherein the output condition is that all of a plurality of arbitrary output conditions selected in advance from the output conditions of 15 or 16 are satisfied.