JP3112037B2 - 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 apparatus for recognizing inputted speech.

[0002]

2. Description of the Related Art In general, in a speech recognition apparatus, a feature pattern of an inputted speech is collated with various standard patterns registered in a dictionary or the like in advance, and the input speech among the standard patterns is compared. Are selected as candidates (recognition results), and the first candidate having the highest similarity among the selected candidates is basically selected as the final recognition result. However, conventionally, even when the first candidate having the highest similarity is obtained, for example, if the recognition reliability of the first candidate is low, the final recognition result is rejected. Or asking the user to confirm the recognition result, and outputting only the confirmed result as a correct recognition result. By adding such a function, it is possible to prevent an erroneous recognition operation from being performed when an acoustic signal other than the voice to be recognized is input or when the input voice is unstable.

[0003]

As described above, in the conventional speech recognition apparatus, when the recognition reliability of the obtained candidate is lower than a predetermined threshold value, for example, erroneous recognition is effectively prevented by a function such as rejection. Can be
On the other hand, since these functions are added, there is a problem that the voice of a specific word is easily rejected even if it is input, and a correct recognition result cannot be obtained immediately. For example, when there are words that are acoustically similar in the vocabulary to be recognized, the recognition reliability is low. In this case, the cause of the low recognition reliability cannot be eliminated even if repetition is performed many times. In addition, since rejected words are likely to be rejected at a similar frequency, a correct recognition result cannot be obtained immediately.

SUMMARY OF THE INVENTION The present invention solves such a conventional drawback. In the case where the recognition result is not adopted and restatement is performed, a correct recognition result can be easily obtained for the reworded word, and the correct recognition result can be obtained. It is an object of the present invention to provide a speech recognition device that can be obtained early.

[0005]

According to a first aspect of the present invention, there is provided a voice input unit for inputting a voice, a recognition unit for recognizing the input voice, and a recognition unit. Determining means for determining whether or not to use the recognition result; and if the recognition result is not used by the determining means, the recognition means recognizes the next input voice by changing the vocabulary to be recognized. It is supposed to
In this case, the recognizing means sets the recognition target vocabulary after the change
It is characterized in that it is set as a top candidate of the recognition result .

[0006]

[0007] According to the second aspect of the present invention, audio is input.
Voice input means for inputting and recognition for recognizing the input voice
Means and the recognition result obtained by the recognition means is adopted
Determination means for determining whether or not the
And the recognition result is not adopted,
Changes the vocabulary to be recognized for the next input speech.
In this case, the recognition means is characterized in that the recognition target vocabulary after the change is a high-rank candidate of the previous recognition result and a synonym of the high-rank candidate.

[0008] The invention according to claim 3 has voice input means for inputting voice, recognition means for recognizing the input voice, and determination means for determining whether or not to use the recognition result. When the recognition result is not adopted by the determination means, the recognition means obtains an integrated score by combining the score of the recognition result of the next input voice and the score of the previous recognition result. It is characterized by:

In the invention according to claim 4 , when the recognition result is not adopted in the determination means, the recognition means combines the scores of the same vocabulary to be recognized into an integrated score. It is characterized by:

[0010] In the invention according to claim 5 , when the recognition result is not adopted in the determination means, the recognition means combines the scores of the synonyms to obtain an integrated score. It is characterized by.

[0011]

According to the first aspect of the invention, when the recognition result is not adopted and the restatement is made, the re-uttered utterance is
Change the recognition target vocabulary and recognize. At this time, the certification after the change
By making the target vocabulary the top candidate of the previous recognition result,
Reduce the number of vocabulary words to be recognized in the post-change recognition process
Can be.

[0012]

According to the second aspect of the present invention, by setting the recognition target vocabulary for the utterance of the rephrasing to the top candidate of the previous recognition result and its synonym, if the utterance cannot be recognized by the previous utterance, Even if the user re-utters in a different language, correct recognition can be performed.

According to the third and fourth aspects of the present invention, in the case where the recognition result is not adopted and restatement is performed, recognition of the reuttered utterance is performed by combining the previous score and the current score.

Further, by combining the previous score with the synonym as in the invention of claim 5 , higher recognition performance can be obtained.

[0016]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a speech recognition device according to the present invention. The voice recognition device of FIG. 1 collates a voice input unit 1 for inputting voice, a feature pattern of the input voice with various standard patterns registered in advance, and detects a feature pattern of the input voice among the standard patterns. The recognizing unit 2 selects similar ones as candidates (recognition results), and the one having the highest similarity among the selected candidates is basically selected as the final recognition result. And a determination unit 3 that determines whether to adopt this candidate as a final recognition result in consideration of the recognition reliability.

Various methods can be used for the recognition method in the recognition section 2. The recognition unit 2 prepares, for example, one word standard pattern for each vocabulary to be recognized, and
Using a matching technique, the distance between the feature pattern of the input speech and the standard pattern is obtained as a similarity value (score) of the vocabulary to be recognized, and candidates up to a predetermined rank are selected in the order of the magnitude of the similarity value. Has become.

Also, various methods can be used for determining the recognition reliability in the determination unit 3. For example, the similarity value R of the first candidate having the largest similarity value
The difference or ratio between (1) and the similarity value R (2) of the next second candidate is taken as the recognition reliability S of the first candidate.
Can be obtained as For example, the recognition reliability S is obtained from the difference between the similarity values R (1) and R (2) as in the following equation.

[0019]

S = R (2) -R (1)

The determination unit 3 determines whether or not to use the first candidate as a recognition result based on whether or not the recognition reliability S is larger than, for example, a threshold value TH. Thus, the user is urged to re-speak, that is, rephrase.

In the first embodiment of the present invention, when the determination unit 3 does not use the recognition result of the first candidate, prompts the user to rephrase, and when the user re-speaks,
The recognition unit 2 recognizes the next input voice by changing the recognition target vocabulary. That is, the recognition unit 2
Either set the top candidate of the previous recognition result in the changed recognition target vocabulary, or set the top candidate of the previous recognition result and the synonym of the top candidate in the changed recognition target vocabulary, and The input speech is recognized.

In the speech recognition apparatus according to the first embodiment having such a configuration, the recognition unit 2 performs recognition by setting the upper candidate of the previous recognition result in the changed recognition target vocabulary. Will be described. The user is currently
A voice is input, and the input voice is recognized by the recognition unit 2, and candidates from the first rank to a predetermined rank are obtained. As a result of the determination by the determination unit 3, the recognition reliability S of the first candidate becomes the threshold TH. If it is smaller, the current recognition result is, for example, as shown in FIG. 2, "A: (first word), B: (second word), C: (third word), "Please input with Iroha." Is displayed on the display or instructed by a synthesized voice to urge the user to restate.

As a result, when the user re-speaks and is sent to the speech recognition unit 2, the speech recognition unit 2 selects the top candidates (for example, from the first place to the third place, ie, a, b) of the above recognition result. , C) as the recognition target vocabulary, recognizes the rephrased voice input, and determines
The recognition reliability S of the first candidate is obtained from the recognition result, and when the recognition reliability S is larger than the threshold value TH, the first candidate is set as the final recognition result.

As described above, even when the final recognition result is not obtained by the first utterance of a specific word, the word is re-stated, and the input speech of this word is changed in the recognition target vocabulary. The first
, The probability that the recognition reliability S of the candidate becomes larger than the threshold value TH can be increased, and a correct recognition result can be easily obtained.

Next, the operation in the case where the speech recognition unit 2 performs recognition by setting the upper candidate of the previous recognition result and the synonym of the higher candidate in the recognition target vocabulary after the change. Will be described. The reason that the range of the vocabulary to be recognized after the change was extended to the synonyms of the top candidates is that the first uttered voice was not recognized because this word was not included in the vocabulary to be recognized. It takes into account that the user may think and have the same meaning but rephrase in different words.

In the case where the speech recognition unit 2 has such a configuration, for example, when a candidate as shown in FIG. Recognition of the rephrased speech is shown in FIG.
For example, the top three candidates "kinou", "kinyou", and "kyo" of the candidates shown in (a) are set as recognition target vocabularies, and these synonyms are also set as recognition target vocabularies. For example, if today is Tuesday, the synonyms for "Kinou" are "
"Kyoubi" as a synonym of "Kyou" is added to the recognition target vocabulary, and the recognition target vocabulary is as shown in FIG. Thus, when the final recognition result is not obtained by the first utterance, a correct recognition result is easily obtained even if the user re-utters with the same meaning but different words.

The synonyms may be registered in advance, or may be dynamically changed according to the operation state of the apparatus. Further, if the recognition result is indicated as shown in FIG. 2, "i, ro, ha" may be respectively recognized as synonyms of "kinou, kinyou, kyo".

In the second embodiment of the present invention,
When the determination unit 3 does not adopt the recognition result of the first candidate and urges the user to rephrase, and the user makes a re-utterance, the recognition unit 2 determines the similarity of each candidate obtained by the current recognition. The combined similarity value (score) is obtained by combining the similarity value (score) of each candidate obtained by the next input voice with the similarity value (score). That is, the recognizing unit 2 obtains an integrated similarity value (score) by combining similarity values (scores) of the same recognition target vocabulary, or integrates by combining similarity values (scores) of synonyms. It is determined as a similarity value (score), and the next input speech is recognized.

In the speech recognition apparatus of the second embodiment having such a configuration, the recognition unit 2 obtains a combined similarity value (score) by combining similarity values (scores) of the same vocabulary to be recognized, and performs recognition. The operation in the case where the operation is performed will be described. The user currently inputs a voice, the input voice is recognized by the recognition unit 2, candidates from the first rank to a predetermined rank are obtained, and the determination unit 3 determines that the recognition reliability of the first candidate is high. When S is smaller than the threshold value TH, the current recognition result is held, for example, as shown in FIG. That is, in the example of FIG. 4A, the recognition reliability S of the first candidate “Kinou” is “2” according to Expression 1, and is smaller than the threshold TH if the threshold is set to “5”, for example. Therefore, for the candidates from the first place to the predetermined place, the distance between the standard pattern of the candidate and the feature pattern of the input voice, that is, the similarity value (score) is held together with the candidate name.

Next, as described in the first embodiment,
The user is encouraged to re-speak, that is, rephrasing, and when the user re-speaks, only the top candidate is recognized as the recognition target vocabulary for the input voice. As a result, for example, if a result as shown in FIG.
Then, for each of the top candidates (for example, the top three candidates), the similarity value (score) in the first utterance shown in FIG. 4A and the second utterance shown in FIG. And the similarity values (scores) are combined. For example, these are added. As a result, the integrated similarity value (score) of each candidate is as shown in FIG. 4C, and the determination unit 3 recognizes the first candidate based on the integrated similarity value (score). The reliability 2S is calculated, the threshold value TH is doubled, for example, and it is determined whether the first candidate is used as a recognition result based on whether the recognition reliability 2S is greater than the threshold value 2TH. In the example of FIG. 4C, the first candidate “Kinou”
The recognition reliability 2S is “17” in accordance with Equation 1, and if the threshold value 2TH is “10”, the first candidate “Kinou” is larger than the final recognition result. Is obtained as

As described above, even when the final recognition result is not obtained by the first utterance, the word is re-worded and the recognition process for the second utterance is performed, and the first similarity value is obtained. By performing recognition in combination with the (score) and the second similarity value (score), a correct recognition result is easily obtained, and the recognition accuracy is improved.

Next, a description will be given of an operation in a case where the recognizing unit 2 performs recognition by obtaining similarity values (scores) by combining similarity values (scores) of synonyms. For the first utterance, the current recognition result is assumed to be held, for example, as shown in FIG. Next, when the user re-speaks,
Recognition of the re-speech is performed on the input speech using the top candidates and their synonyms as recognition target vocabulary. For example,
For example, the top three candidates “Kinou”, “Kinyou”, and “Kyou” of the candidates shown in FIG. 4A are set as recognition target vocabularies, and if today is Tuesday, “Kinou”
Also, "getsoubi" as a synonym of "" and "kayoubi" as a synonym of "kyo" are added to the vocabulary to be recognized.
Assuming that a recognition result as shown in FIG. 5 is obtained for such a vocabulary to be recognized, the recognizing unit 2 further calculates
Similarity value (score) in the first utterance shown in FIG.
And the similarity value (score) of the second utterance shown in FIG. 5 is combined to calculate an integrated similarity value (score).

At this time, synonyms such as “Kinou”,
In “getsou”, one having the smaller similarity value (score), that is, the smaller distance is adopted and integrated. Specifically, in FIG. 5, between “Kinou” and “Getuyou”, “Kinou” has a smaller similarity value (score) than “Kinou”, so the similarity value (score) of “Kinou”, ie, “ 64 ", which is added to the similarity value (score) of" Kinou "in FIG. 4A, that is," 31 "and integrated, and the integrated similarity value (score) for" Kinou, Getuyou "is added. ) As "95". Similarly, for “Kinyou”, “86” is obtained as an integrated similarity value (score), and for “Kyou, Kayou”,
"64" is obtained. In this case, the first candidate is the one with the smallest integrated similarity value, that is, “Kyou, Kayou”, and the recognition reliability 2S of “Kyou, Kayou” is:
According to Equation 1, it becomes “22” and the threshold value 2TH is “10”.
If so, the first candidate “Kyo, Kayou” is obtained as the final recognition result.

As described above, the first similarity value (score)
By further recognizing the same similarity value (score) with the second similarity value (score), the correct recognition result can be more easily obtained, and the recognition accuracy can be further improved.

[0035]

As described above, according to the first aspect of the present invention, in the case where the recognition result is not adopted and the restatement is performed, the reworded utterance is recognized by changing the recognition target vocabulary. Therefore, a correct recognition result can be easily obtained, and a correct recognition result can be obtained at an early stage. In particular,
The recognition target vocabulary after the change is set as the top candidate of the previous recognition result.
In the recognition process after the change, the number of words to be recognized is
Less, which makes it easier to get correct recognition results
Become.

[0036]

According to the second aspect of the present invention, since the recognition target vocabulary for the utterance of the rephrasing is set to the top candidate of the previous recognition result and its synonym, the recognition is not possible by the previous utterance. Even if the user re-speaks in a different language, correct recognition can be performed.

According to the third and fourth aspects of the present invention, when a restatement is performed without using a recognition result, recognition of the reuttered utterance is recognized by combining the previous score and the current score. It is easy to get the correct recognition result,
A correct recognition result can be obtained early.

According to the fifth aspect of the present invention, when rephrasing is performed, the re-uttered utterance is recognized using the top candidate of the previous recognition result and its synonym as the recognition target vocabulary. Therefore, if the user cannot recognize by the previous utterance, even if the user re-utters with a different word, the user can correctly recognize the word, and furthermore, by combining the previous score with the synonym, And higher recognition performance can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech recognition device according to the present invention.

FIG. 2 is a diagram illustrating an example of a message urging a user to rephrase.

FIGS. 3A and 3B are diagrams for explaining processing in the first embodiment of the present invention.

FIGS. 4A, 4B, and 4C are diagrams for explaining processing according to a second embodiment of the present invention.

FIG. 5 is a diagram for explaining processing in a second embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Voice input part 2 Recognition part 3 Judgment part

──────────────────────────────────────────────────続 き Continuation of front page (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 15/00-17/00

Claims (5)

(57) [Claims] A voice input unit for inputting voice, a recognition unit for recognizing the input voice, and a determination unit for determining whether or not to use a recognition result obtained by the recognition unit; If the recognition result in the determination means is not adopted, the recognition means is adapted to recognize and modify a vocabulary to be recognized for the next input speech, this time, the
The recognition means uses the changed recognition target vocabulary as a result of the previous recognition result.
A speech recognition device characterized as being a top candidate . 2. A voice input means for inputting voice,
And a recognition means for recognizing the received speech.
Determining means for determining whether or not to employ the obtained recognition result.
The recognition result is not adopted by the determination means
In the case where the recognition target word is
In this case, the recognition means sets the recognition target vocabulary after the change to a top candidate of the previous recognition result and a synonym of the top candidate. <br/> be Ruoto voice recognition device. 3. A voice input device for inputting voice, a recognition device for recognizing the input voice, and a determination device for determining whether or not to use the recognition result, wherein the determination result Is not adopted, the recognizing means obtains an integrated score by combining the score of the recognition result of the next input voice with the score of the previous recognition result. . Wherein said recognition means, claim recognition result in the determination means when not adopted, characterized in that adapted to a score that integrates a combination of scores for the same recognition target vocabulary 3 The speech recognition device according to the above. 5. The speech recognition apparatus according to claim 3 , wherein said recognition means combines the scores of the synonyms to obtain an integrated score.