JP3006496B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

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

[0002]

2. Description of the Related Art Word detection, compared to normal speech recognition,
This has the advantage that it is not necessary to determine the section in which the word exists in the input voice in advance.

Conventionally, as a method for detecting a specific word from a utterance, reference 1 (IEEE J67-D, 11p.1).
242-1249). In the method described in this paper, when the likelihood calculated independently for each frame candidate word exceeds a certain threshold, detection is performed independently of the detection of other candidate words. As another method, a method described in Document 2 (“Speech Recognition by Stochastic Model”, Seiichi Nakagawa, pp. 20-26) is known. In the method of this document, an optimal word is determined and output on a language model for each frame. Also, as a method for suppressing detection of words that overlap in time, reference 3 (IEICE Technical Report SP95-7)
7p. 31-38) are known. In the method of this document, provisional detection of each candidate word is determined in each frame, and detection that temporally overlaps can be suppressed using frame synchronization processing based on the likelihood of each detection.

[0004]

However, the method of Reference 1 allows a hypothesis that candidate words temporally overlap during utterance. Such a phenomenon does not occur in the method of Reference 2, but a phenomenon that a word similar to a part of the uttered word is detected in the middle of the utterance may occur. In order to prevent this, it is necessary to determine the range of time during which the word exists in the input voice by some means. The method of Ref. 3 does not require such a case, and does not allow a hypothesis in which candidate words temporally overlap during utterance. However,
As shown in Document 3, only in front of the recognition target word,
When using a model to which a model that accepts an arbitrary syllable string (a garbage model) is added, the likelihood of a word similar to a part of a spoken word is estimated to be higher than the likelihood of an actually spoken word. As a result, a phenomenon (partial matching) that detects a similar word and cancels the detection of an actually uttered word may occur, and a method of improving the phenomenon is not obvious. However, when using a model in which a garbage model is added before and after the recognition target word and using the method shown in Reference 2, partial matching is avoided by setting a low pattern matching score of the garbage model. It is known that it can.

[0005] It is an object of the present invention to use a method for reducing the detection of a plurality of candidates for one word utterance as described above for word detection, and a method which also has robustness to partial matching described in Reference 2. Accordingly, it is an object of the present invention to provide a word recognition method having the above-mentioned advantages of the methods described in Documents 1 to 3.

[0006]

SUMMARY OF THE INVENTION The present invention relates to a speech recognition apparatus for detecting candidate words for each frame from input speech data on a frame-by-frame basis, wherein a plurality of candidate words in one word are detected. It relates to a method for reduction.

[0007]

BEST MODE FOR CARRYING OUT THE INVENTION

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

[0008] A speech recognition apparatus according to a first embodiment of the present invention includes a speech analysis unit 101 for extracting a feature amount from an input speech for each predetermined time (hereinafter referred to as a frame), 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; A likelihood calculation unit 104 for obtaining the likelihood, a detection unit 105 that outputs a candidate word by using an output from the likelihood calculation unit 104 as an input, and a storage unit 106 associated therewith
Consisting of

The speech analysis unit 101 performs a frequency analysis for each frame of the input speech, and generates a feature vector (hereinafter referred to as a feature) for each frame. Power, power variation, mel cepstrum, mel cepstrum variation, mel cepstrum secondary variation, and the like are used as elements of the feature quantity. The feature amount for each frame is output to the likelihood calculation unit 104 for each frame.

The word dictionary 102 stores a word to be recognized in the form of a chain of unit acoustic models, for example, a chain of acoustic models representing syllables constituting a word ("O" in the case of "Osaka"). "O"-"sa"-"ka").

In the model generation unit 103, the word dictionary 102
A language model is constructed that accepts utterances with arbitrary syllable chains added before and after each word model in. The likelihood calculation unit 104 calculates an 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 optimal sequence and the likelihood output from the likelihood calculation unit 104 for each frame, determines candidate words and determines output, and outputs candidate words if necessary.

A speech recognition apparatus according to the first embodiment is a speech recognition apparatus according to the first, second, and fifth aspects.

(Explanation of Operation of First Embodiment) FIG. 2 is a diagram showing the operation of the speech recognition apparatus according to the first embodiment of the present invention.

The speech analysis unit 101 performs a frequency analysis for each frame of the input speech to generate a feature amount,
Output to likelihood calculation section 104 every frame. The likelihood calculation unit 104 uses the automaton control O
Using the method of the ne Pass DP method, by performing pattern matching of the language model generated by the model generation unit 103 and the feature amount in each frame output from the speech analysis unit 101, the optimal sequence of each frame and its The likelihood is calculated, and the optimal sequence 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 greater than the threshold, the candidate word in the optimal sequence output from the likelihood calculating unit 104 (hereinafter, referred to as an optimal word) is stored in the storage unit 1
Compared with the candidate word stored in 06, and if different, the optimum word is output (step 2).

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

Next, the effects of the first embodiment of the present invention will be described. The method disclosed in Document 2 shows a means for obtaining the optimal sequence and likelihood for each frame. However, if these are directly used as the recognition results of word speech, the recognition results are output for each frame, which is inconvenient. The first mode 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, a model shown in FIG. 3 is used as a language model. The model of FIG. 3 is constructed in the form of a concatenation of three models: a front garbage that accepts any syllable string including an empty string, a candidate word, and a rear garbage that accepts any syllable string including an empty string; All utterances that pass through each model in this order, that is, utterances with an arbitrary syllable string added before and after one candidate word, are accepted. It is assumed that the model generation unit 103 previously creates and stores the model in FIG. 3 from each candidate word in the word dictionary 102. It is assumed that the contents of the word dictionary 102 are three words of “Seki” (seki), “Hekinan” (Hekinan), and “Naha” (Naha). The likelihood calculating unit 104 performs pattern matching between the feature amount sequence from the first frame to the corresponding frame for each frame and the language model of FIG. 3 according to the method described in Literature 2, and determines the optimal sequence on the relevant language model and Calculate the likelihood. However, the natural logarithm of the probability value in the final state of the language model is used as the likelihood. The storage unit 106 stores a threshold value of −100.0 in advance, and stores a null word as an initial value of the candidate word. In this embodiment, the frame interval is 10 ms.

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

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

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

In the 121st frame, step 1
, The likelihood -99.0 is compared with the threshold value -100.0. Since the likelihood is equal to or larger than the threshold value, in step 2, the optimal word “Hekinan” in the 121st frame is compared with the candidate word (blank word) stored in the storage unit 106, and “Hekinan” is output because of the difference. I do. Thereafter, in step 3A, the optimum word “Hekinan” is stored in the storage unit 106.

Since the likelihoods are all greater than or equal to the threshold after the 122nd frame, Step 1 is performed in the same manner as in the 121st frame, and in Step 2, the optimal word “Hekinan” and the candidate word stored in the storage unit 106 are "Hekinan" is compared, and the output of the optimum word is not performed because they are the same. In step 3A, the optimum word “Hekinan” is stored in the storage unit 106.

(Second Embodiment) FIG. 6 is a diagram showing a configuration 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 predetermined time (hereinafter referred to as a frame) from an input speech.
A voice analysis unit 101 for extracting a feature amount of each word, a word dictionary 102 storing words, a model generation unit 103 for generating a language model from words in the word dictionary 102, A likelihood calculating unit 104 for finding an optimal word sequence (hereinafter referred to as an optimal sequence) applicable to the language model in each frame and its likelihood;
An intermediate model generating unit 107 for generating an intermediate model from words in the word dictionary 102; an intermediate likelihood calculating unit 108 for obtaining the likelihood of the intermediate model in each frame from the feature amount and the intermediate model; It comprises a detection unit 105 that outputs the candidate words by using the outputs from the degree calculation unit 104 and the intermediate likelihood calculation unit 108 as inputs, and a storage unit 106 associated therewith.

The same speech analyzer 101, word dictionary 102, model generator 103, and likelihood calculator 104 as those described in the first embodiment are used.

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

The midway likelihood calculating section 108 performs pattern matching of all the midway models generated by the midway model generating section 107 and the feature amount of each frame which is the output of the speech analyzing section 101. Is calculated. Then, for each candidate word, the largest of the likelihoods of the model in the middle of the candidate word (hereinafter referred to as the middle likelihood) is obtained and output to the detection unit 105.

The storage unit 106 stores a predetermined threshold value and a predetermined preliminary threshold value. Also, a candidate word, a pending optimal word, and a list of its opposing candidate words are stored.

The detecting unit 105 receives the output of the likelihood calculating unit 104 and the output of the intermediate likelihood calculating unit 108,
The output word is determined and output using the information in 06.

The speech recognition apparatus according to the second embodiment realizes a speech recognition apparatus according to claims 1, 2, 3, 6, 7, and 14.

(Explanation of Operation of Second Embodiment) FIG.
FIG. 7 is a diagram illustrating an operation of the voice recognition device according to 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 corresponding likelihood is equal to or greater than the threshold, the likelihood calculating unit 10
4 is compared with the candidate words stored in the storage unit 106 (step 2). If the likelihood is less than the threshold, the empty word is stored in the storage unit 106 (step 3B). If the likelihood is equal to or larger than the threshold, the optimum word is stored in the storage unit 106 (Step 3A). If the likelihood is equal to or greater than the threshold and the optimal word is different from the candidate word stored in the storage unit 106, the candidate word whose intermediate likelihood is equal to or greater than the preliminary threshold in the optimal word and the current frame (hereinafter, referred to as List of all alternative words)
When the past conflict candidate list stored in the storage unit 106 includes the best word of the current frame, the list and the corresponding past best word information are stored in the past conflict candidate list stored in the storage unit 106 (step 4). The data is deleted from the storage unit 106 (step 4-2). As post-processing, all the candidate frames for which the intermediate likelihood is lower than the preliminary threshold in the current frame are removed from the past candidate word list stored in the storage unit 106 (step 5). Next, among the optimum words stored in the storage unit 106, those having an empty candidate word list are output, and the corresponding information is deleted from the storage unit 106 (step 6). However, here, candidate words for comparing the intermediate likelihood 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, when a word similar to a part of a candidate word is uttered, for example, when “Hekinan” is uttered, “Seki” similar to “Heki” of “Hekinan” is used. )
May be detected. In such a case, the second embodiment suspends the detection of "Seki" if it can be estimated that the word is in the middle of another word, that is, "Hekinan" (step 4). Upon completion, the detection of "Seki" is canceled (step 4-2). If the likelihood on the way to “Hekinan” is sufficiently reduced, the estimation that it is on the way to “Hekinan” is discarded (Step 5), and the reserved “Seki” is detected (Step 6). In this way, it is possible to prevent a subsequence similar word from being detected. In addition, since the output of the optimum word is presumed to be in the middle of another candidate word in the discarded frame, the delay time caused by the detection of the subsequence similar word is minimal.

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

The intermediate model is composed of a forward garbage, which accepts an arbitrary syllable string including an empty string, and a concatenation of two models of candidate words, as shown in FIG. That is, all utterances with an arbitrary syllable string 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, an intermediate likelihood can be obtained by extracting the likelihood of the state indicated by the black point in FIG.
It is assumed that the intermediate model generation unit 108 creates and stores the intermediate 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 is obtained by appropriately extracting the likelihood of the intermediate state of the language model in FIG. can get. Therefore, the present embodiment can be realized by a method in which the model in 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 is set to -101.5. Since the preliminary threshold selects candidate words that may be in the middle of the current frame, the lower the lower the threshold, the more the number of alternative candidate words increases, and the lower the possibility of estimation omission in the middle of the word. Word output delay time becomes longer. Usually, a value smaller than the threshold is selected. In this embodiment, it is assumed that the conflict candidate word does not include the optimum word of the corresponding frame.

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

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

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

In the 91st frame, in step 1, the likelihood -99.0 is compared with the threshold value -100.0. Since the likelihood is equal to or larger than the threshold, the ninth
The optimal word “Seki” of one frame is stored in the storage unit 106 (Step 3A). Next, a list of conflict candidate words (in this case, only “Hekinami”) is created, the optimal word “Seki” is stored in the storage unit 106, and a list including only “Hekinami” is stored (step 4).

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

From the 92nd frame to the 120th frame, since the optimum word remains unchanged at “Seki”, only the operation of writing “Seki” into the storage unit 106 is performed from Step 1 to Step 4. In step 4-2, nothing is performed because the optimal word “Seki” of the current frame does not exist in the past candidate list.

In step 5, the middle likelihood of "Hekinan", which is an element of the alternative candidate word list, is equal to or greater than the preliminary threshold, so that nothing is performed. In step 6, nothing is performed because the list includes "Hekinan".

In the 121st frame, the optimum word is replaced with "Hekinan". In steps 1 and 2, “Hekinan” is stored in the storage unit 106. In steps 3 and 4, the contending candidate word list including only the optimal word “Hekinan” and the element “Naha” is stored in the storage unit 106.

In step 4-2, since the candidate word list of the past optimal word “Seki” includes the optimal word “Hekinan” in the 121st frame, the storage unit 106 stores the past optimal words “Seki” and “Seki”. The conflict candidate word list is deleted. In steps 5 and 6, there is no corresponding candidate word list, so nothing is done.

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

In the 141st frame, since there is no change in the optimum word, steps 1 to 4-2 move in the same manner as the immediately preceding frame. In step 5,
Since the intermediate likelihood of “Naha” is below the preliminary threshold, it is removed from the list of candidate words for the past optimal word “Hekinan”. As a result, the candidate word list of “Hekinan” becomes empty, and “Hekinan” is detected in step 6.
In this way, the output of the subsequence analogy word “Seki” of “Hekinan” is not performed, and the output of the uttered word “Hekinan” is performed as soon as the midway estimation of the alternative 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, an operation is performed to output not only the candidate word for which the conflict candidate word list is empty, but also the one whose frame number stored in the storage unit 106 is a predetermined time or more past the current frame. Conceivable. When such an operation is adopted, the delay time from the utterance of the candidate word to the output thereof can be forcibly suppressed within a certain time, which is effective in realizing a speech recognition device that wants a recognition result as quickly as possible. .

This modified example is described in claims 1, 2, 6, 7, and 1.
A speech recognition device corresponding to 3, 14 is realized.

As a further modified example, in step 6 of the above modified example, when the frame number stored in the storage unit 106 is equal to or more than a certain value, an operation of outputting a corresponding optimum word may be considered. This is effective as a device that accepts only an input within a predetermined time from the activation of the voice recognition device.

This modified example is described in claims 1, 2, 6, 7, 1.
A speech recognition device corresponding to 0 and 14 is realized.

Further, as another modified example of the second embodiment, in step 6 in the detecting unit 105, not only the candidate word list is empty but also the external input device, that is, the keyboard in the corresponding frame. When an input signal from a switch, a switch, another speech recognition device, or the like is received, an operation of outputting the optimum word stored in the storage device 106 is conceivable. Thereby, the existence section of the word is limited by the speaker determining the timing of the recognition,
It can be expected that recognition performance will improve. For example, it is effective when the speaker can be requested to perform an operation such as pressing a button after the utterance is completed.

This modified example is described in claims 1, 2, 6, 7, 1.
1 and 14 are realized.

In this case, as an external input device,
It is conceivable to use a voice detection device. As the voice detection device, for example, a device described in JP-A-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 start and end of a voice section in conjunction with the speaker's utterance, using a voice detection device instead of a button etc. allows a voice recognition device that does not require the speaker to perform extra operations. realizable. In this case, the detection unit 105 considers the case where the voice section end detection by voice detection fails.
In step 4 of, the number of the current frame is stored together in the storage unit 106, and in step 6,
Even if there is no signal from the voice detection device, it is conceivable to perform an operation of outputting a frame number stored in the storage unit 106 that is past a current frame for a predetermined time or more. In this case, the first and second aspects of the present invention are described.
A speech recognition device corresponding to 6, 7, 17 can be realized. In this case, the output conditions adopted in claim 17 are the same as those in claim 1.
These are output conditions in 2 and 14.

Further, as a modification of the second embodiment, after the conflict candidate word list is created in step 4 of the detection unit 105, the optimum words are not changed and the end of the optimum row ends with the corresponding optimum word. During the interval, an operation of adding all the candidate words whose intermediate likelihoods have reached the preliminary threshold in the corresponding frame to the list may be performed to more reliably list the opposing candidate words. With the addition of this operation, it is possible to adopt an alternative candidate word whose likelihood happens to be low in the frame where the optimum word first appears, without missing from the list.

This modified example is described in claims 1, 2, 4, 6,
A speech 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 this case, after the alternative candidate word list is created in step 4 of the detection unit 105, the optimum words are not changed and the last of the optimum row ends with the corresponding optimum word, and for a certain period thereafter, For example, during the period of about 200 ms, an operation of adding all the candidate words whose intermediate likelihood has reached the preliminary threshold in the corresponding frame to the list is performed, and the intermediate likelihood of all the opposing candidate words in the list is lower than the preliminary threshold. It is also possible to design such that the output operation is performed for the first time. This means, for example, that a plurality of candidate words are 200 ms.
This is a method for reducing erroneous detection by using the assumption that it is possible to make an assumption that the user will not be continuously spoken within this time. This is defined in claims 1, 2, 4, 6, 8, 10, 1
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 alternative 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 while the likelihood of some word exceeds the preliminary threshold. Further, when the optimal word is replaced, it is always included in the alternative candidate word list, so that the detection of the past optimal word is abandoned. According to this modification as well, not only detection of temporally overlapping words but also detection of temporal proximity can be suppressed.

This modification is described in claims 1, 2, 4, 6,
8 and 16 are speech recognition devices. 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,
A voice detection device may be added as an external input device, and a condition that a voice termination detection signal from the voice detection device is provided may be added to the output condition to make the output condition more strict. As a result, it is possible to enhance the detection suspension by using both the information on the temporal proximity and the possibility of another candidate word being uttered.

This modification is described in claims 1, 2, 4, 6,
The output conditions according to claims 18 and 18 correspond to the output conditions according to claims 12 and 16. This is also an example in which the fixed 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 realize a speech recognition apparatus with less erroneous detection with a minimum required delay. can do.

[Brief description of the drawings]

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

FIG. 2 is a flowchart illustrating an operation of the voice recognition device according to the first exemplary embodiment of the present invention.

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

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

FIG. 5 illustrates a speech recognition device according to the first embodiment of the present invention.
6 is a table showing an optimum word and likelihood for each frame in the embodiment.

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

FIG. 7 is a flowchart illustrating an operation of the voice recognition device according to the second exemplary embodiment of the present invention.

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

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

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

Continuation of front page (56) References JP-A-9-166995 (JP, A) JP-A-8-83091 (JP, A) JP-A-6-266386 (JP, A) JP-A-5-19784 (JP) , A) JP-A-4-362698 (JP, A) JP-A-9-179581 (JP, A) JP-A-9-106297 (JP, A) Proceedings of the Acoustical Society of Japan, Spring Meeting, 1996 I ▼ 3-5-2 “Word spotting for word recognition based on syllable units” p. 111-112 (issued March 26, 1996) IEICE Technical Report [Voice] Vol. 95, No. 355, SP95-77, “Word recognition by word spotting using semisyllables” p. 31-38 (Issued November 16, 1995) Proceedings of 1985 IEEE International Conference on Acoustics, Speech and Signal Processing, Vol. 3 "Keyword Recognition using Template update Concatenation" p. 1233-1236 (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 3/00-9/20 JICST file (JOIS)

Claims (18)

(57) [Claims] 1. A speech analysis unit for extracting a feature amount for each frame from a frequency analysis of a fixed time (frame) of an input speech, and a speech model for receiving an arbitrary syllable string before and after a candidate word are added. A likelihood calculation unit that performs pattern matching between the language model and the feature amount for each frame, selects an optimal word sequence (optimum sequence) on the language model for each frame, and calculates the likelihood thereof; A detection unit that determines and outputs an optimal word for each frame from the optimal sequence and the likelihood, and when the optimal word is repeated once or more, the optimal word within a corresponding time (detection section) is determined at most A speech recognition device characterized by outputting once. 2. A speech analysis unit for extracting a feature amount for each frame from a frequency analysis for each fixed time (frame) of an input speech, and an arbitrary syllable string before and after a predetermined candidate word is received. Pattern matching between the language model to which the speech model is added and the feature amount is performed for each frame, an optimal word sequence (optimum sequence) on the language model is selected for each frame, and likelihood calculation for calculating the likelihood is performed. And a detection unit that determines and outputs an optimal word for each frame from the optimal sequence and the likelihood, and when the optimal word is repeated one or more times, a corresponding time (detection section)
A speech recognition device for outputting the optimum word in the detection section at most once when the likelihood reaches a predetermined constant value (threshold) even once within the range. 3. The speech recognition apparatus according to claim 1, wherein the candidate words included in the optimum sequence are set as optimum words. 4. The method according to claim 1, wherein the candidate word is output once as the optimal word if the last word in the optimal sequence is the candidate word, and otherwise the optimal word is not output. Item 3. The speech recognition device according to item 1 or 2. 5. The speech recognition apparatus according to claim 1, wherein the time at which the optimum word is output is the first time of the detection section. 6. The speech recognition apparatus according to claim 1, wherein the time at which the optimum word is output is the first time at which a predetermined condition (output condition) is satisfied. . 7. A language model (intermediate model) for receiving an utterance to which an arbitrary syllable string is added immediately before all of the first half subsequences of all of the candidate words, and in each of the frames, A device for calculating the likelihood obtained by performing the pattern matching of the feature amount, and calculating the maximum likelihood (intermediate likelihood) among the likelihoods of all the intermediate models of each of the candidate words; In, the output is suspended only when there is the candidate word (alternative candidate word) in which the intermediate likelihood has reached a predetermined constant value (preliminary threshold), and thereafter, the predetermined condition (output condition) 5. The speech recognition device according to claim 1, wherein the output is performed at the first time when the speech is satisfied. 8. A language model (intermediate model) for receiving an utterance to which an arbitrary syllable string is added immediately before all of the first half substrings of all of the candidate words, and in each of the frames, all of the intermediate models and the An apparatus for calculating the likelihood obtained by performing the pattern matching of the feature amount, and calculating the maximum likelihood (intermediate likelihood) among all the intermediate models of the candidate word, wherein the intermediate likelihood is included in the detection section. Suspends output only when there is a candidate word (alternative candidate word) that has reached a predetermined constant value (preliminary threshold) even once, and thereafter, the first time the predetermined condition (output condition) is satisfied 4. An output at a time.
Or the voice recognition device according to 4. 9. A language model (intermediate model) for receiving an utterance to which an arbitrary syllable string is added immediately before all of the first half substrings of all of the candidate words, and in each of the frames, all of the intermediate models and the A device for calculating the likelihood obtained by performing pattern matching of the feature amount, and calculating the maximum likelihood of all the intermediate models of the candidate word (intermediate likelihood), wherein the intermediate likelihood is within the detection interval and Thereafter, the output is suspended only when there is a candidate word (alternative candidate word) which has reached a predetermined constant value (preliminary threshold value) in all frames within a predetermined time period. 4. An output at the first time when a predetermined condition (output condition) is satisfied.
Or the voice recognition device according to 4. 10. The output condition is that, in the frame after a predetermined time has elapsed since the apparatus was started, there is no conflict candidate word that has become the optimum word since the detection time. Item 10. The speech recognition device according to item 7, 8 or 9. 11. An apparatus according to claim 6, further comprising: an input device for receiving an external input, wherein the output condition is that a signal from the input device is supplied after the detection time. The speech recognition device according to the above. 12. The apparatus according to claim 1, wherein said input device is a voice detection device for determining a range of a voice section, and said output condition is that a signal of a voice section end from said voice detection device is given. The speech recognition device according to claim 11. 13. The output condition is that, in the frame for which a predetermined time has elapsed after the detection time, there is no conflict candidate word that has become the optimum word since the detection time. The speech recognition device according to 7, 8, or 9. 14. After the detection time, the intermediate likelihoods of all of the alternative candidate words once respectively fall below the preliminary threshold value, and the relevant candidate word has not become the optimum word even once from the detection time. 10. The speech recognition device according to claim 7, wherein the output condition is set as the output condition. 15. In a certain frame after the detection time, the halfway likelihood of all of the alternative candidate words is lower than the preliminary threshold value, and the relevant alternative candidate word becomes the optimum word at least once from the detection time. 10. The speech recognition apparatus according to claim 7, wherein the output condition is that no voice recognition is performed. 16. In a certain frame after the detection time, the intermediate likelihood of all other candidate words is lower than the preliminary threshold value, and the other candidate word is at least once the optimal word. 10. The speech recognition device according to claim 7, wherein the output condition is that the condition is not satisfied. 17. An output condition in which any one of a plurality of output conditions selected in advance from among the output conditions according to claim 10, 11, 12, 13, 14, or 15 is satisfied. The speech recognition device according to claim 7, 8 or 9, wherein 18. The method of claim 10, 11, 12, 13, 14,
10. The speech recognition apparatus according to claim 7, wherein the output condition is that all of a plurality of output conditions selected in advance from among the output conditions in 15 or 16 are satisfied.