JPH09258765A - Voice interval start end correcting method and device for voice recognition, and voice recognizing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To exclude unsteady noise, detected as a voice interval by mistake, from an object to be recognized by detecting unsteady noise, and correcting the start end of the voice interval. SOLUTION: A sound analysis result delay part 2 delays a sound analysis result 1A, sent from a sound analyzing part 1, by the specified time and sends its result 2A to a word collating part 7 and a noise detecting part 13. A partial sentence generating part 9 obtains each partial sentence and each collating score from a word collated result 7A sent from the word collating part 7 and sends its result 9A to a word predicting part 6, a recognized result output part 10 and the noise detecting part 13. The noise detecting part 13 detects unsteady noise on the basis of the collation score and sentence structure analysis result of the partial sentence, the interval length of an input signal judged to be coincident with a silent standard pattern, and the adaptation of the input signal to the partial sentence. A voice interval detecting switch 11 connects with the start end detecting part 3 side upon receiving a noise detection signal 13A from the noise detecting part 13.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for correcting the start of a voice section in order to exclude non-stationary noise erroneously detected as a voice section from a recognition target in voice section detection for voice recognition. A voice recognition method using this method is particularly useful for improving the accuracy of voice recognition in a man-machine interface by a voice language.

[0002]

2. Description of the Related Art In a conventional voice recognition apparatus, when a voice signal is input, the start and end of the voice section are detected based on the power of the input signal, and the word standard is obtained in the resulting voice section. The voice is recognized by continuously matching the pattern with the input signal based on the grammatical rules.

Grammar rules are based on the assumption that a sentence is uttered by a person, and various combinations of ordinary words appearing in the sentence and silences (pauses) occurring between and before and after the word are used in combination for recognition. . The word standard pattern includes a standard pattern corresponding to the original word and a standard pattern corresponding to silence (silent standard pattern). Even if it is silent, there is practically no sound at all, and there is stationary noise (hereinafter, stationary noise) such as white noise on the telephone line.
Normally, stationary noise in the environment to be recognized is treated as silence.

Therefore, in speech recognition, since silence is a recognition target and is considered to be a type of word, unless otherwise noted, a word includes silence, and a word standard pattern includes a silence standard pattern. Shall be included.

As shown in FIG. 10, a conventional voice recognition device of this type has an acoustic analysis unit 1 and an acoustic analysis result delay unit 2.
A starting point detecting section 3, an ending detecting section 4, a grammar rule storing section 5, a word predicting section 6, a word matching section 7, a word standard pattern storing section 8, a partial sentence generating section 9, and a recognition result. Output unit 10
A voice section detection switch 11 and a power calculation unit 12
And is mainly realized by a central processing unit (CPU) and software.

In FIG. 10, various grammatical rules used for recognition are stored in advance in the grammatical rule storage unit 5, and various standard word patterns are stored in the standard word pattern storage unit 8. The voice section detection switch 11 is connected in advance to the start end detection unit 3 side. Under such a premise, the voice recognition device of FIG. 10 performs voice recognition according to the following procedures (1) to (9).

(1) The power calculation unit 12 sequentially calculates the power of the input signal after the device is activated, and sends the result 12A to the start end detection unit 3 through the voice section detection switch 11.

Here, the voice section detecting switch 11 is initially connected to the starting end detecting section 3 as described above, and thereafter, when the starting end detecting signal 3A is received from the starting end detecting section 3, the ending detecting section 4 is detected. , And when the end detection signal 4A is received from the end detection unit 4, the connection is returned to the start end detection unit 3 side, and the power calculation result 12 is calculated for each connected side.
Send A.

(2) The leading edge detecting section 3 sequentially compares the power sent from the power calculating section 12 with a threshold TH1 for detecting the leading edge, as shown in FIG. At the time when the predetermined time t2 or more continues, the start end detection signal 3A is set to the voice section detection switch 11 and the word prediction unit 6
Send to

(3) On the other hand, the end detection unit 4 successively compares the power sent from the power calculation unit 12 with the end detection threshold TH2 as shown in FIG. At the time when the state to be taken continues for a predetermined time t3 or longer, the end detection signal 4A is set so as to perform the collation from the time when the power first becomes the threshold value TH2 or less to the time t4.
Is sent to the voice section detection switch 11, the word prediction section 6 and the recognition result output section 10.

(4) The acoustic analysis unit 1 sequentially performs acoustic analysis for extracting characteristic parameters of the input signal, and sends the result 1A to the acoustic analysis result delay unit 2.

(5) The acoustic analysis result delay unit 2 sets a margin time t1 (see FIG. 11) for the acoustic analysis result 1A sent from the acoustic analysis unit 1 so as to avoid a start end detection error.
And the result 2A is sent to the word matching unit 7.

(6) In the word prediction unit 6, the partial sentence generation unit 9
From the sub-sentence matching result (each sub-sentence and each matching score) 9A and the grammar rule 5A stored in the grammar rule storage unit 5, the word matching unit 7 matches the next word, that is, the matching. One or a plurality of words 6A are estimated and sent to the word matching unit 7.

The processing of the word predicting section 6 is performed by the starting point detecting section 3
Starting from the time when the leading edge detection signal 3A is received,
It continues until the end detection signal 4A is received from the end detection unit 4.

(7) The word collating unit 7 reads out the word standard pattern corresponding to the collating word 6A received from the word predicting unit 6, that is, the collating word standard pattern 8A from the word standard pattern storing unit 8 and obtains the collating word standard. The pattern 8A is collated with the delayed acoustic analysis result 2A, and the result, that is, the word collation result 7A is sent to the sub-sentence generation unit 9 for the word that has been collated up to the end of the word standard pattern.

As a well-known method of calculating the matching score in the word matching unit 7, DP matching, a hidden Markov model (HMM), or the like is used. (Refer to Seiichi Nakagawa, "Speech Recognition by Stochastic Model")

As word prediction and matching progress, part of the sentence obtained by arranging the matched words at each time according to grammatical rules, that is, a partial sentence, is obtained, and the matching score for each partial sentence is calculated. Can be calculated.

(8) Then, the sub-sentence generating unit 9 obtains each sub-sentence and its matching score from the word matching result 7A sent from the word matching unit 7, and the result 9A is recognized by the word predicting unit 6 and the recognition unit. It is sent to the result output unit 10.

(9) The recognition result output unit 10 is the end detection unit 4
After receiving the end detection signal 4A from, the sub-sentence having the highest matching score among all the sub-sentence matching results 9A sent from the sub-sentence generating unit 9 is output as the recognition result 10A for the input signal.

[0020]

However, in such a conventional method, before the user of the voice recognition device (device user) utters the voice of the recognition target, a cough, inhalation, exhalation, and nose are generated. Soothing sound, lip noise, people's voice around, bell sound,
When non-stationary noise (hereinafter referred to as non-stationary noise) such as pulse sound, handset sound, door sound, footstep sound, paper rubbing sound, keyboard sound, etc. is input, if the power is sufficiently large, it is a voice section. The start edge of is detected.

Therefore, although these non-stationary noises are sounds that are not included in the word standard pattern and are not recognized, only these non-stationary noises are detected as speech to start speech recognition, or the non-stationary noises are detected. The recognition target voice may be detected as one voice together with the voice to be recognized, and voice recognition may be started. Since the voice recognition is performed, the voice recognition is started and the recognition performance is deteriorated.

Therefore, the present invention provides a method and apparatus for correcting the beginning of a voice section in order to exclude non-stationary noise erroneously detected as a voice section from a recognition target, and a voice recognition method using this method. To aim.

[0023]

According to the present invention, an input signal and a word standard pattern are recognized in speech recognition for continuously recognizing a voice by collating a word standard pattern with an input signal continuously based on grammatical rules. Matching score and parsing result of the partial sentence obtained from the matching result with, and the section length of the input signal determined to match the silent standard pattern,
Based on the matching degree between the input signal and the sub-sentence, non-stationary noise outside the recognition target is detected, the start end of the voice section is corrected, and then the recognition is started again.

That is, according to the method for detecting the beginning of a voice section of the present invention, in a device for recognizing a voice by continuously matching a standard word pattern with an input signal based on grammatical rules, the result of matching between the input signal and the standard word pattern is used. The matching score and the syntactic analysis result of the sub-sentence obtained, the section length of the input signal determined to match the silent standard pattern, and the non-stationary noise etc. are detected based on the matching degree between the input signal and the sub-sentence, It is characterized in that the start end of the voice section is corrected.

Further, according to another method of detecting the beginning of a voice section of the present invention, when recognizing a voice by matching a word standard pattern with an input signal based on a grammatical rule, it is obtained from a matching result of the input signal and the word standard pattern. From the partial sentences and their matching scores, the first condition that the matching score of the partial sentence ending in silence is the maximum is satisfied, and the section length of the input signal corresponding to the silence at the end of the partial sentence is equal to or longer than a predetermined time. It is determined whether or not there is a sub-sentence that satisfies the second condition that is, and at the time when it is determined that both the first and second conditions are satisfied, the matching score of the input signal and the separately determined reference pattern. Is a reference score, and the difference between the reference score and the matching score of the same partial sentence obtained from the matching result of the input signal and the word standard pattern in the previous section is within a predetermined range. Determining whether or not fulfilled,
When it is determined that the third condition is satisfied, it is determined that the input signal of the section corresponding to the sub-sentences that are determined to satisfy both the first and second conditions is non-stationary noise, and the start end of the voice section is corrected. Or, in addition, the first and second
And when the third condition is satisfied, the section length of the input signal corresponding to the sub-sentence determined to satisfy both the first and second conditions is obtained, and the section length is within a predetermined range. It is determined whether or not the fourth condition of conforming to the section length is satisfied, and when it is determined that the fourth condition is satisfied, it is determined that both the first, second and third conditions are satisfied. It is characterized in that the input signal of the section corresponding to the sub-sentence is determined to be non-stationary noise, and the start end of the speech section is corrected, or instead of the third condition, the first and second conditions are set. For the sub-sentences that are determined to satisfy both, the matching score in each state of the same sub-sentence is compared to obtain the maximum value, and the input signal corresponding to the same sub-sentence up to the state where the maximum value is obtained and the reference pattern separately defined. The reference score is the matching score with It is determined whether the fifth condition that the difference from the maximum value is within a predetermined range is satisfied, and when it is determined that the fifth condition is satisfied, both the first and second conditions are satisfied. It is characterized that the input signal of the section corresponding to the sub-sentence determined to be non-stationary noise is corrected and the start end of the voice section is corrected, or it is determined that both the first and second conditions are satisfied. A sixth condition that the section length of the input signal corresponding to the same sub-sentence in the state of taking the maximum value among the sub-sentences is obtained, and the section length conforms to the standard section length of the same sub-sentence within a predetermined range. When it is determined that the sixth condition is satisfied,
The input signal of the section corresponding to the sub-sentence determined to satisfy all of the first, second and fifth conditions is non-stationary noise, and the beginning of the speech section is corrected, or As the correction of the beginning of the voice section, the beginning of the voice section is corrected at a certain time before the time point when the input signal of the section corresponding to the partial sentence is determined to be non-stationary noise.

Further, the voice section start edge detecting apparatus of the present invention is
In a device for recognizing speech by matching a standard word pattern with an input signal based on grammatical rules, a partial sentence obtained from the matching result of the input signal and the standard word pattern and its matching score are used to match a partial sentence ending in silence. Whether there is a partial sentence that satisfies the first condition that the score is maximum and that satisfies the second condition that the section length of the input signal corresponding to silence at the end of the partial sentence is equal to or longer than a predetermined time. At the time when it is determined that both the first and the second conditions are satisfied by the first determining means for determining whether or not both the first signal and the second condition are satisfied, the matching score between the input signal and the separately determined reference pattern is set as the reference score, and the reference score in the previous section is set. Secondly determining whether or not a third condition that a difference between the matching score of the same partial sentence obtained from the matching result of the input signal and the standard word pattern and the reference score is within a predetermined range is satisfied When the determination means determines that the third condition is satisfied, the input signal in the section corresponding to the sub-sentence determined to satisfy both the first and second conditions is determined to be non-stationary noise, and the speech section And a third means for correcting the starting point of, or a portion determined to satisfy both the first and second conditions at the time when the first, second and third conditions are satisfied. Means for determining a section length of an input signal corresponding to a sentence, and determining whether or not a fourth condition that the section length conforms to the standard section length of the sub-sentence within a predetermined range is satisfied, The third means, when it is determined that the fourth condition is satisfied, that the input signal in the section corresponding to the sub-sentences that are determined to satisfy the first, second and third conditions is non-stationary noise. It is characterized in that the start end of the voice section is determined and the second end is corrected. Instead of the third condition, the determining means compares the matching scores in each state of the partial sentence with respect to the partial sentence determined to satisfy both the first and second conditions, and obtains the maximum value. A fifth is that a matching score between an input signal corresponding to the same sub-sentence seen in a state of taking a value and a separately defined reference pattern is used as a reference score, and a difference between the reference score and the maximum value is within a predetermined range. Section that corresponds to the sub-sentence that is determined to satisfy both the first and second conditions when the third means determines that the fifth condition is satisfied. The input signal is determined to be non-stationary noise and the start end of the voice section is corrected, or the maximum value is taken from the sub-sentences determined to satisfy both the first and second conditions. Obtain the section length of the input signal corresponding to the same sub-sentence seen in the state , And a means for determining whether or not a sixth condition that this section length conforms to the average section length of the same partial sentence within a predetermined range is satisfied, and the third means satisfies the sixth condition. When it is determined that the input signal of the section corresponding to the sub-sentence determined to satisfy all of the first, second and fifth conditions is non-stationary noise, the beginning end of the speech section is corrected. Is characterized by.

Further, according to the voice recognition method of the present invention, the start end of the voice section is detected by the power calculation of the input signal when the apparatus for recognizing the voice is started by collating the standard word pattern with the input signal based on the grammatical rules. After that, the detected start end is corrected by any one of the above voice section start end correction methods to perform voice recognition, or the voice is recognized by matching the standard word pattern with the input signal based on the grammatical rules. The speech recognition and the speech segment start end correction method according to any one of claims 1 to 6 are started at the time of starting the device without detecting the beginning of the speech segment by the power calculation of the input signal or the like. The voice is recognized while being corrected by the correction method, and the input signal is matched with the standard word pattern without detecting the end of the voice section by calculating the power of the input signal. The feature is that the end of the voice section is detected and the voice recognition is terminated based on the matching score and the syntactic analysis result of the sub-sentence obtained and the section length of the input signal determined to match the silent standard pattern. And

[0028]

BEST MODE FOR CARRYING OUT THE INVENTION When a grammatical rule is used to continuously match a standard word pattern with an input signal such as speech or noise, the sub-sentence and the input signal are separated at each time as word prediction and matching progress. A matching score is obtained. Regardless of whether the input signal is voice, noise, etc., in the silent section, the matching score of the partial sentence ending in silence among all the partial sentences is the maximum, and it is determined that it matches the standard pattern of silence (silent section) Continues for a while.

However, when the non-stationary noise, which is a non-recognition sound, is collated with a different pattern, which is the word standard pattern, the collation score is low. Also, the section length determined to match the sub-sentence is often significantly different from the originally expected sub-sentence length.

Therefore, using the matching score of the partial sentence and the syntactic analysis result obtained by matching the input signal and the standard pattern of the word, and the section length of the input signal determined to match the standard pattern of silence, the silence is used. It is determined whether or not it is a section, and further at this time, it is determined whether or not the input before that is non-stationary noise based on the fitness of the input signal and the sub-sentence,
If it is determined to be non-stationary noise, the beginning of the voice section is corrected in the silent section several frames before from that point,
Restart voice recognition.

By applying the above-described method or apparatus for correcting the beginning of the speech section to continuous speech recognition, the input signal and the word standard pattern can be collated in the section excluding the non-stationary noise which was conventionally erroneously detected. Therefore, the performance of voice recognition is improved.

[0032]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings with reference to the accompanying drawings.

FIG. 1 shows a voice recognition device according to an embodiment of the present invention. The speech recognition apparatus of FIG. 1 detects the beginning of a speech section using the power of an input signal, and uses the input signal of the section up to the end of a speech section that is also detected using the power of the input speech as grammatical rules. In a voice recognition device that continuously matches a word standard pattern based on the above, it is determined that the matching score and the syntactic analysis result of the partial sentence obtained from the matching result of the input signal and the word standard pattern match the silent standard pattern. This is a method in which non-stationary noise or the like is detected based on the section length of the input signal and the matching degree with the sub-sentence of the input signal, and the start end of the speech section is corrected.

That is, this speech recognition apparatus is provided with the acoustic analysis unit 1
An acoustic analysis result delay unit 2, a start end detection unit 3, an end detection unit 4, a grammar rule storage unit 5, a word prediction unit 6, a word collation unit 7, a word standard pattern storage unit 8, and a portion. It is composed of a sentence generation unit 9, a recognition result output unit 10, a voice section detection switch 11, a power calculation unit 12, and a noise detection unit 13, and is mainly realized by a central processing unit (CPU) and software.

The grammar rule storage unit 5 stores some grammar rules used for recognition in advance. For example, various grammatical rules representing “(silence) Yano-san (silence) of the general affairs department” as shown in FIG. 2 are prepared. In FIG. 2, the grammatical rules are shown using a network.

The word standard pattern storage unit 8 stores standard patterns of all the words appearing in the grammar rules (in the example of FIG. 2, "general affairs department", "no", "Yano", "san", "(silence)". ) Is prepared in advance.

It is assumed that the voice section detecting switch 11 is connected to the starting end detecting section 3 side at the beginning, that is, at the time of starting the apparatus, and after that, when the starting end detecting signal 3A is received from the starting end detecting section 3, the end detecting section 4 is detected. When the end detection signal 4A is received from the end detection unit 4 or the noise detection signal 13A is received from the noise detection unit 13, the start end detection unit 3 side is connected.

The functions of the respective parts of the speech recognition apparatus shown in FIG. 1 will be described below in accordance with the procedure of speech recognition.

The voice power calculation unit 12 sequentially calculates the power of the input signal, and sends the result 12A to the start end detection unit 3 through the voice section detection switch 11.

As described above, the voice section detection switch 11
Is first connected to the starting end detecting unit 3, and thereafter, when the starting end detecting signal 3A is received from the starting end detecting unit 3, it is connected to the ending end detecting unit 4 and the ending detection signal 4A is received from the ending end detecting unit 4. When the noise detection signal 13A is received from the noise detection unit 13, the connection is returned to the starting end detection unit 3 side, and the power calculation result 12A is sent to each connected side.

The leading edge detecting section 3 compares the power sent from the power calculating section 12 with a threshold value TH1 as shown in FIG. 10, and when the state of exceeding the threshold value TH1 continues for a predetermined time t2 or more. , The leading edge detection signal 3A is sent to the voice section detection switch 11, the word prediction section 6, and the noise detection section 13.

On the other hand, the end detection unit 4 determines the power sent from the power calculation unit 12 as a threshold TH2 as shown in FIG.
In comparison with the threshold value TH2, when the state in which the value is equal to or less than the threshold value TH2 continues for a predetermined time t3 or more, the collation is performed from the time when the power first becomes the threshold value to the time t4.
The end detection signal 4A is sent to the voice section detection switch 11, the word prediction unit 6, the noise detection unit 13, and the recognition result output unit 10.

The acoustic analysis unit 1 sequentially performs acoustic analysis for extracting characteristic parameters of the input signal, and sends the result 1A to the acoustic analysis result delay unit 2.

The acoustic analysis result delay unit 2 delays the acoustic analysis result 1A sent from the acoustic analysis unit 1 by a margin time t1 which is set in order to avoid a start end detection error, and the result 2A is delayed by the word collation unit 7 and the word collation unit 7. It is sent to the noise detection unit 13.

Further, the acoustic analysis result delay unit 2 stores the acoustic analysis result for a certain period of time, and when the noise detection signal 13A is received from the noise detection unit 13, the acoustic analysis result before a certain time is re-worded. It is sent to the matching unit 7 and the noise detection unit 13. It is desirable that the fixed time before is set so as to correspond to the section length of the input signal corresponding to “silence at the end of a partial sentence”, which is a condition for detecting non-stationary noise by the noise detection unit 13.

In the word predicting unit 6, the partial sentence matching result (partial sentence and its matching score) sent from the partial sentence generating unit 9.
9A and the grammar rules 5A stored in the grammar rule storage unit 5, the word matching unit 7 next finds one or more words to be matched, that is, the matching words 6A, and sends them to the word matching unit 7.

The processing of the word predicting section 6 is started at the time point when the start edge detecting signal 3A is received from the start edge detecting section 3, and the end detecting section 4 receives the end detecting signal 4A or the noise detecting section 13 receives the noise detecting signal 13A. Continue until you receive it.
When the noise detection signal 13A is received, the process is newly started at the time when the next start edge detection signal 3A is received.

The word collating unit 7 reads out the word standard pattern corresponding to the collating word 6A received from the word predicting unit 6, that is, the collating word standard pattern 8A from the word standard pattern storing unit 8, and the delayed acoustic analysis result 2A, The obtained matching word standard pattern 8A is matched, and the word matching result 7A is sent to the sub-sentence generating unit 9 for the matched word up to the end of the word standard pattern.

As a method of calculating the matching score used in the matching in the word matching unit 2, there are the DP matching and the method using the hidden Markov model as described above.

When the hidden Markov model is used to calculate the matching score, the matching score is calculated by the following equation (1).

[0051]

[Equation 1]

The sub-sentence generating unit 9 obtains each sub-sentence and each collation score from the word collation result 7A sent from the word collating unit 7, and outputs the result 9A to the word predicting unit 6 and the recognition result output unit 10. send.

Speaking of the grammatical rule illustrated in FIG. 2, as sub-sentences, “(silence)”, “(silence) general affairs department”, “(silence) general affairs department”, and “(silence) general affairs department (Silence) ”,“ (Silence) General Affairs Department (silence) Yano ”,“ General Affairs Department (silence) Yano-san ”,“ (Silence) General Affairs Department (silence) Yano-san (silence) ” And a matching score is obtained for each.

Here, the generation of sub-sentences and the calculation of the matching score thereof will be described by taking the case of using an HMM (Hidden Markov Model) as an example. First, as for the matching score, for example, when "no" is predicted from "(silence) general affairs department",
The word HMM "no" in the HMM "(silence) general affairs department" of the partial sentence
HMM of "(silence) general affairs department"
And the calculation of Equation 1 is continued, the collation score of the partial sentence “(silence) of the general affairs department” is obtained. Also, "General Affairs Department", "No", "(silence)", "Yano",
By storing the word string “san”, the partial sentence “(silence) Yano san (silence) of the general affairs department” can be generated.

In the equation (1) of the equation 1, the above-mentioned calculation process of the matching score is described for each word. In FIG. 3 (a), for example, instead of making an HMM of a partial sentence “general affairs department's” by connecting word HMMs, writing with two word HMMs as shown in FIG. Then, for example, between the words "general affairs department" and "no", the score at the time of transition to the end state A of the word "general affairs department" and the beginning state B of the word "no" to the same state B The collation score at the time of self transition is compared, and the larger one is set as the collation score of state B.

However, when the matching of the word "no" is requested at the beginning, since the leading state B has not been matched yet, the score of the ending state A of the word "general affairs department" is unconditionally set to the leading state B. Use as a matching score. The scores of the other states C, D, and E of the word "no" are set to -∞ (minus infinity) for initialization.

By calculating the matching score in this way, the matching score of the partial sentence "(silence) general affairs department" becomes the matching score in the end state A of the word "general affairs department", and the partial sentence "(silence) general affairs department" is obtained. The collation score of "no" is the collation score of the word "no" in the end state E. The sub-sentence generating unit 9 associates the sub-sentence thus obtained with the matching score.

However, the partial sentence does not necessarily have to be held as an actual word string, but can be held in a format suitable for use in word prediction. As a method used for word prediction, "Early method", "LR method" (reference: Hirogo Nomura, basic technology of natural language processing, edited by the Institute of Electronics, Information and Communication Engineers) are widely used.

The noise detection unit 13 obtains the matching score of the partial sentence and the syntactic analysis result obtained by matching the input signal with the word standard pattern, and the section length of the input signal determined to correspond to the silent standard pattern. By using the goodness of fit between the input signal and the sub-sentence and the section length of the input signal that is determined to correspond to the standard pattern of the sub-sentence, the input speech matched up to a certain time is a non-stationary noise (non-recognition target). If it is determined that the noise is non-stationary noise, the start end of the voice section is corrected before a fixed time, and the matching process is performed again. That is, when the start edge is detected by the power, the recognition processing is performed again from the start edge detection by the power each time the non-stationary noise is detected.

Specifically, it is constantly checked whether or not the following conditions 1 to 3 are satisfied, and when it is satisfied, it is determined that the input voice is non-stationary noise, and the noise detection signal 13A is set to the word prediction unit 6. , The voice section detection switch 11 and the acoustic analysis result delay unit 2, respectively.

Condition 1: Partial sentence ending in silence, eg, FIG.
In the grammar example, the matching score of "(silence)", "(silence) General Affairs Department (silence)" or "(silence) General Affairs Department (silence) Yano-san (silence)" matches all sub-sentences. Be the highest of the scores.

Condition 2: The section length of the input signal judged to match the silent standard pattern is equal to or longer than a predetermined time.

Condition 3: The matching degree between the input signal and the sub-sentence does not satisfy the standard.

Here, as the matching degree between the input signal and the sub-sentence, a method of using the collation score of the sub-sentence, or the collation score of the sub-sentence and other standard patterns (reference patterns) used as a reference and the input signal There is a method of comparing with the matching score of, for example, using the difference.

In addition to the goodness of fit using these matching scores, a comparison between the section length determined to match the sub-sentence and the originally expected section length of the sub-sentence, for example, a ratio can be used. When such a degree of conformity based on the section length is used together, the accuracy of determining non-stationary noise is further improved.

A concrete example of the noise detecting section 13 is illustrated in FIG. The noise detecting unit 13 includes a silent section length calculating unit 16, a noise determining unit 17, a partial sentence matching score comparing unit 18, a syntax analyzing unit 19, a state-based matching score comparing unit 20, a reference pattern matching unit 21, and a fitness determining unit. 22 and a silence determination section 23.

In the noise detection unit 13 illustrated in FIG. 4, the partial sentence matching score comparison unit 18 first outputs the partial sentence matching result 9
The matching scores of A are compared with each other, the sub-sentence having the maximum matching score is searched, and the result 18A is sent to the parsing unit 19.

Subsequently, the syntactic analysis unit 19 determines whether or not the last word (end of the partial sentence) of the partial sentence having the maximum matching score searched by the partial sentence matching score comparison unit 18 is "(silence)". And the result 19A is determined as the silence determination unit 23.
Send to

The sub-sentence matching score comparing section 18 and the syntax analyzing section 19 determine whether or not Condition 1 described above is satisfied.

The silent section length calculation unit 16 determines from the partial sentence matching score comparison result 18A whether Condition 2 is satisfied, that is, an input signal determined to match the silent standard pattern (example of FIG. 10). Then, the section length 16A of the signal corresponding to “(silence)” at the end of the sentence is calculated and the silence determination unit 2
Send to 3.

Here, how to obtain the section length of the input signal that matches the silent standard pattern will be described.

When the matching score is calculated using the HMM in the word matching unit 7 (see the expression (1) of the expression 1), the expressions (2.1) to (2.5) of the following expression 2 and the expression (3) of the expression 3 are used. ) Calculates the section length of the input signal that matches the standard pattern of words (including silence).

[0073]

[Equation 2]

[0074]

[Equation 3] [Continuation time of input voice that matches silent standard pattern] = [current time] -begin (current time, end state of silent standard pattern) Equation (3)

Here, with reference to FIG. 5, an example of calculating the section length of the input signal which is determined to match the word when the HMM is used as the word standard pattern will be described.

In FIG. 5, assuming that the score collated on the route 200 at time t + 5 is the highest, in this case, the time at which collation up to state 4 of the word standard pattern at time t + 5 is started is obtained.

First, in the first state 1 of the word, if the score of the transition A from the previous word is higher than the score of the transition B from itself, the time t + 1 is stored. That is, be
gin (t + 1, state 1) = t + 1.

For words other than the first state 1 of the word, the time held at the state transition source giving the highest matching score is inherited. For example, begin (t + 3, state 2) = be
It becomes gin (t + 2, state 2).

By repeating this process, when the matching is performed up to the state 4 of the word standard pattern at each time, the time when the matching with the word standard pattern on the route 200 giving the highest matching score is started is obtained. To be

The voice section length that matches the standard word pattern at time t + 5 is (t + 5) -begin (t +
5, state 4).

By this calculation, the section length of the input signal that matches the silence is found by finding the section length of the sentence end of the partial sentence obtained by the partial sentence matching score comparison unit 18 that matches the word standard pattern corresponding to silence. Is required.

In the silence judging section 23, when the syntax analyzing section 19 judges that the last word of the partial sentence having the maximum matching score obtained by the matching score comparing section 18 is “(silent)”, It is determined whether or not the silent section length 16A obtained by the silent section length calculation unit 16 for the partial sentence is longer than or equal to a predetermined time, and when it is longer than or equal to the predetermined time, the silent section detection signal 23A is determined as noise. Send to section 17.

In the state-based collation score comparison unit 20,
The word matching unit 7 compares the matching scores in each state (matching score by state) obtained at the time of matching the input signal and the word standard pattern with each other to obtain the maximum value of the matching score by state,
The duration of the voice section (input signal) corresponding to the sub-sentence seen in the state of taking the maximum value is obtained, and the result 20A is sent to the fitness determining unit 22.

Here, how to obtain the section length of the input signal which is determined to match the sub-sentence will be described.

The section length of the input signal that matches up to a certain state of the HMM which is the standard pattern of the sub-sentence is generally expressed by the following equation 4.
Expression (4), Expression 5 (5.1), (5.2), Expression 6 (6.1) ~
(6.5), which is calculated by the equations (7.1) and (7.2) of Equation 7.

[0086]

## EQU00004 ## (1) For the head state of the word HMM at the beginning of the sentence at the time t _S of recognition start: length (t _S , j) = 0 ... Formula (4)

[0087]

[Equation 5] (2) For other than at the start of recognition: For other than the leading state of the word standard pattern, length (t + 1, j) = length (t, j) ... Equation (5.1) However, for all states m When α (t, j) + O (k (t), j → j) ≧ α (t, m) + O (k (t), m → j) length (t + 1, j) = length (t, m) Formula (5.2) However, for a certain state m, α (t, j) + O (k (t), j → j) <α (t, m)
When + O (k (t), m → j)

[0088]

[Equation 6] Regarding the start state of the word standard pattern (a) When the word matching is requested for the first time, length (t + 1, j) = length (t, F) Equation (6.1) However, the preceding word is silent. Time length (t + 1, j) = t-begin (t, F) + length (t, F) Equation (6.2) However, when the previous word is other than silent (b) after that length (t + 1, j) = length ( t, j) Equation (6.3) where α (t, j) + O (k (t), j → j) ≧ α (t
+ 1, F) length (t + 1, j) = length (t, F) Equation (6.4) where α (t, j) + O (k (t), j → j) <α (t
+ 1, F) and the previous word is silent length (t + 1, j) = t-begin (t, F) + length (t, F) Equation (6.5) However, when the previous word is other than silent

Here, the meanings of the symbols in the equations 4 to 6 are as follows. length (t, j): length of the section to the end state of the preceding word when matching progresses to the state j of the standard word pattern at time t m: j from the previous state that can transit to state j Excluding. F: End state of the word standard pattern preceding this word

[0090]

## EQU00007 ## <Section Length of Input Signal Determined to Match Partial Sentence at Time t, State j> When word at end of partial sentence is silent = length (t, j) Equation (7.1) When the word is a word other than silence = t-begin (t, j) + length (t, j) ... Formula (7.2)

Here, with reference to FIG. 6, an example of calculating the section length of the input signal which is determined to match the sub-sentence when the HMM is used as the word standard pattern will be described.

Since the speech section length that matches the word standard pattern at each time is obtained by the above equations (2.1) to (2.5) and equation (3) of equation 3, it corresponds to the preceding word. The section length that matches the sub-sentence is obtained by taking over the section length of the input signal.

First, at the start of recognition t _S , the partial sentence length 0 is stored in the head state 1 of the word at the beginning of the sentence. That is, length (t
_S , state 1) = 0.

Then, in the first state 1 of the word, when the score of the transition A from the previous word is higher than the score of the transition B from itself, the section length to the end state of the previous word is stored.

That is, if the previous word is silent, len
If th (t, state 3 ′) is a word other than silence, then length (t, state 3 ′) + (t) −be
Store gin (t, state 3 ′) in length (t + 1, state 1).

For the states other than the first state 1, the time held at the state transition source giving the highest matching score is taken over. For example, length (t + 3, state 2) = length
th (t + 2, state 2).

By repeating this process, it is possible to refer to the section length of the input signal that matches the partial sentence up to the previous word in each state of the word standard pattern at each time.

From this, for example, the section length of the input signal that matches the word standard pattern seen in state 4 at time t + 5 becomes length (t + 5, state 4) if there is no sound, and if it is a word other than silence, (T + 5) -b
egin (t + 5, state 4) + length (t + 5,
State 4) is reached.

By such a calculation, the section length matching the partial sentence in the state of the word standard pattern having the highest matching score is obtained, and the result and the matching score are sent to the matching degree judging unit 22.

In the reference pattern matching unit 21, from the time when the start edge detection signal 3A is received to the time when the end edge detection signal 13A is received, the input signal and the reference pattern (details will be described later) are calculated according to the following equation (8). Are sequentially performed, and the matching score is sent to the fitness determining unit 22.

[0101]

(Equation 8)

As the reference pattern used here, a model in which the noise standard patterns shown in FIG. 7 are combined, a standard pattern for syllable recognition shown in FIG. 8, or a pattern shown in FIG. 9 in which these are combined (hybrid) is used. Conceivable. FIGS. 7, 8 and 9 show the case where the HMM is used as the reference pattern. Here, the non-stationary noise is a model in which the noise standard patterns shown in FIG. 7 are combined, a syllable recognition standard pattern shown in FIG. 8, or a hybrid thereof.
It is a good match with the pattern shown in Fig. 4 and its matching score is relatively high, but the speech to be recognized (words and silence) does not have a good match with these reference patterns and has a relatively low matching score. The difference between is used.

The conformity determination unit 22 uses the following expression 9 which is a condition based on the difference between the matching score of the reference pattern and the matching score by state.
(9) and the following equation (10), which is a condition related to the ratio of the duration of the voice section of the sub-sentence and the originally expected duration of the sub-sentence, are not satisfied,
The nonconforming signal 22A is sent to the noise determination unit 17.

[0104]

(9) (L (s, t) −L _p (t)) ÷ D (s, t) ≧ θ _L (Equation (9)

Here, s (t): the state in which the matching score is maximum at a certain time t L (s, t): the matching score in the state s (t) L _p (t): the matching score of the reference pattern D (S, t): Section length of the input signal that matches the sub-sentence obtained from the matching result θ _L : A parameter that determines the upper limit of the difference between the matching score by state and the matching score of the reference pattern. D (s,
t) is used for normalization.

[0106]

[Equation 10] M (s, t) × θ _min ≦ D (s, t) ≦ M (s, t) × θ _max Equation (10)

Here, s (t): the state in which the matching score is maximum at a certain time t D (s, t): the section length M (s, t) of the input signal that matches the partial sentence obtained from the matching result ): Expected section length of sub-sentence θ _min : Parameter defining lower limit of section length θ _max : Parameter defining upper limit of section length.

It should be noted that, as the reference pattern, the non-stationary noise as shown in FIGS. 7 to 9 is relatively favorably matched and the matching score thereof is relatively high, but the speech to be recognized (word and silence).
It is also possible to independently use a plurality of N types of patterns having the characteristic that the matching with is not so favorable and the matching score is relatively low. In that case, the reference pattern matching unit 21 finds the matching scores of the respective reference patterns in parallel by the equation (8) of the equation (8).
Assuming that the matching score of the n-th reference pattern is L _pn (t), N pieces of upper limit parameters θ _Ln corresponding to each reference pattern n are prepared in the conformity determination unit 22, and the equation (11) below is used. ), If any of the N conditions represented by) is not satisfied, as in the case where the equation (9) of the equation 9 is not satisfied,
The nonconforming signal 22A is sent to the noise determination unit 17.

[0109]

[Equation 11] (L (s, t) -L _pn (t)) ÷ D (s, t) ≧ θ _Ln (where n = 1 to N) Equation (9)

When the noise determining section 17 receives both the silent section detection signal 23A sent from the silence determining section 23 and the nonconforming signal 22A sent from the fitness determining section 22, the input before that is non-stationary noise. Therefore, the noise detection signal 13A is output.

After receiving the end detection signal 4A from the end detection unit 4, the recognition result output unit 10 has the highest matching score among all the partial sentence matching results 9A sent from the partial sentence generation unit 9. Is output as the recognition result 10A.

In the above-described embodiment, the start end and the end of the voice section are detected by the start end detection section 3 and the end detection section 4 using the power calculation result 12A of the input signal. Both detections can be omitted. An example of the configuration is shown in FIG. Compared to FIG. 1, FIG. 12 does not include the voice detection switch 11, the voice power calculation unit 12, the start end detection unit 3, and the end detection unit 4, but instead includes the end detection unit 34. At the end detection unit 34, when both the condition that the matching score is the maximum among all the sub-sentences that are grammatically accepted and the condition that the silent section at the end of the sentence is a certain time or more are both satisfied, Detects the end of the voice section. Specifically, the voice recognition device starts voice recognition at the same time as starting, and the noise detection unit 13 also starts detecting non-stationary noise at the same time as starting the voice recognition device. At the time of detection, the noise detection signal 13A is acoustically analyzed. It is given to the result delay unit 2 and the word prediction unit 6. The word prediction unit 6 starts the processing at the same time as it is activated. Based on the grammar rule 5A and the partial sentence matching result 9A, the end detection unit 34 determines that the matching score is the maximum among all the partial sentences that are grammatically accepted, and the section corresponding to silence at the end of the sentence. The end of the voice section is detected at the time when both the conditions of being equal to or longer than a certain time are satisfied, and the end detection signal 34A is given to the word prediction unit 6, the recognition result output unit 9, and the noise detection unit 13. Others are the same as the embodiment of FIG. As a result, it is possible to avoid the deterioration of the recognition performance due to the detection error of the start end and the end which has conventionally occurred when the level of the input signal is low. In this case,
The acoustic analysis result delay unit 2 does not need to have a function of delaying the acoustic analysis result 1A sent from the acoustic analysis unit 1 as shown in FIG. 2 by a margin time t1 placed in order to avoid a start end detection error. As described above, the acoustic analysis result is stored for a certain period of time, and the noise detection unit 1 outputs the noise detection signal 1
When 3A is received, it may have a function of sending again to the word matching unit 7 and the noise detection unit 13 based on the acoustic analysis result before a certain time. Further, the reference pattern matching unit 21 newly starts matching between the reference pattern and the input signal at the time of receiving the end detection signal 34A.

Next, the noise detecting section 1 of the embodiment shown in FIG.
3, the syntactic analysis unit 19 determines whether or not there is a silent section at the end of the sentence with respect to the partial sentence having the maximum matching score obtained by the partial sentence matching score comparison unit 18.
As shown in FIG. 3, the syntactic analysis unit 19 may search for a partial sentence having a silent section at the end of the sentence and the partial sentence matching score comparison unit 18 may find the partial sentence having the maximum matching score. In addition, the calculation of the silent section length is performed by the syntax analysis unit 1
For the partial sentence whose sub-sentence end is silence obtained by 9, the sub-sentence matching score comparison unit 18 may be performed in parallel with the sub-sentence section length calculation unit 16, or a sub-sentence end may have a silence section and maximum. The silent section length may be calculated only for the partial sentence having the matching score of.

Further, as shown in FIG. 14, as shown in FIG.
It is also possible to omit the state-based collation score comparison unit 20 from the configuration of FIG. In this case, the maximum value of the matching score by state,
Instead of the duration of the input signal corresponding to the sub-sentence seen in the state of taking the maximum value, the maximum value of the matching score of each sub-sentence and the input signal of the section corresponding to the sub-sentence taking the maximum value These may be used by obtaining the duration.

[0115]

As described above in detail with reference to the embodiments, the present invention provides a voice section excluding non-stationary noise,
Since the input signal can be matched with the word standard pattern, the performance of voice recognition is improved.

[Brief description of drawings]

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

FIG. 2 is a diagram showing an example of grammar rules.

FIG. 3 is an explanatory diagram of partial sentences and matching score calculation.

FIG. 4 is a block diagram showing an embodiment of a noise detection unit.

FIG. 5 is an explanatory diagram of a section length calculation that matches a word standard pattern.

FIG. 6 is an explanatory diagram of section length calculation that matches a partial sentence.

FIG. 7 is a diagram showing an example of a reference pattern (noise standard pattern).

FIG. 8 is a diagram showing an example of a reference pattern (standard pattern for syllable recognition).

FIG. 9 is a diagram showing an example of a reference pattern (hybrid configuration of noise standard pattern and syllable recognition standard pattern).

FIG. 10 is a block diagram of a conventional example.

FIG. 11 is a diagram showing an example of voice section detection using power.

FIG. 12 is a block diagram showing another embodiment of the present invention.

FIG. 13 is a block diagram showing another embodiment of the noise detection unit.

FIG. 14 is a block diagram showing still another embodiment of the noise detection unit.

[Explanation of symbols]

 1 acoustic analysis unit 2 acoustic analysis result delay unit 3 start edge detection unit 4 end detection unit 5 grammar rule storage unit 6 word prediction unit 7 word matching unit 8 word standard pattern storage unit 9 partial sentence generation unit 10 recognition result output unit 11 speech section Detection switch 12 Speech power calculation unit 13 Noise detection unit 16 Silence section length calculation unit 17 Noise determination unit 18 Partial sentence matching score comparison unit 19 Syntax analysis unit 20 State-based matching score comparison unit 21 Reference pattern matching unit 22 Fitness level determination Part 23 Silence determination part 34 End detection part

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical display location G10L 3/00 561 G10L 3/00 561G 5/06 5/06 A

Claims (12)

[Claims] 1. An apparatus for recognizing a voice by continuously matching a standard word pattern with an input signal based on a grammatical rule, which is obtained from a matching result between the input signal and the standard word pattern.
Matching score of sub-sentence and syntactic analysis result, section length of input signal judged to match with silent standard pattern, and non-stationary noise etc. are detected based on conformity of input signal and sub-sentence, A method for correcting the beginning of a voice section, characterized by correcting the beginning. 2. When recognizing a voice by matching a standard word pattern with an input signal based on a grammatical rule, a partial sentence obtained from the matching result of the input signal and the standard word pattern and its matching score end silently. The first condition that the matching score of the sub-sentence is maximum is satisfied, and
It is determined whether or not there is a sub-sentence that satisfies the second condition that the section length of the input signal corresponding to silence at the end of the sub-sentence is equal to or longer than a predetermined time, and both the first and second conditions are determined. When it is determined that they satisfy, the matching score of the input signal and the separately defined reference pattern is used as the reference score, and the matching score of the same sub-sentence obtained from the matching result of the input signal and the word standard pattern in the previous section. And whether the third condition that the difference between the reference score and the reference score is within a predetermined range is satisfied. When it is determined that the third condition is satisfied, the first condition and the second condition are determined. A method for correcting the beginning of a voice section, characterized in that the input signal in the section corresponding to a sub-sentence that is determined to satisfy both is determined to be non-stationary noise, and the beginning of the voice section is corrected. 3. When the first, second, and third conditions are satisfied, the section length of the input signal corresponding to the sub-sentence determined to satisfy both the first and second conditions is obtained, and Determining whether or not a fourth condition that the section length conforms to the standard section length of the sub-sentence within a predetermined range is satisfied; and when it is determined that the fourth condition is satisfied, the first and the first Two
3. The voice section according to claim 2, wherein the input signal of the section corresponding to the sub-sentence determined to satisfy both the first and third conditions is non-stationary noise, and the start end of the voice section is corrected. Starting point correction method. 4. For a sub-sentence that is determined to satisfy both the first and second conditions in place of the third condition, collation scores in each state of the sub-sentence are compared to obtain a maximum value, The reference score is the matching score between the input signal corresponding to the same sub-sentence and the reference pattern defined separately when the maximum value is taken,
It is determined whether a fifth condition that the difference between the reference score and the maximum value is within a predetermined range is satisfied. When it is determined that the fifth condition is satisfied, the first and second 3. The method of correcting the beginning of a voice section according to claim 2, wherein the input signal of the section corresponding to the sub-sentences that are determined to satisfy both conditions is non-stationary noise, and the beginning of the voice section is corrected. . 5. A section length of an input signal corresponding to a sub-sentence found in a state where the maximum value is obtained among sub-sentences determined to satisfy both the first and second conditions, and the section length is Determining whether or not a sixth condition that the average section length of the same partial sentence is satisfied within a predetermined range is satisfied; and when determining that the sixth condition is satisfied, the first and second conditions are satisfied.
5. The voice section according to claim 4, wherein the input signal of the section corresponding to the sub-sentence determined to satisfy both the first and fifth conditions is non-stationary noise, and the beginning of the voice section is corrected. Starting point correction method. 6. The correction of the start of the voice section, the start of the voice section is corrected at a certain time before a time point when the input signal of the section corresponding to the partial sentence is determined to be non-stationary noise. The method for correcting a voice section start end according to any one of claims 1 to 5. 7. An apparatus for recognizing a voice by matching a standard word pattern with an input signal based on a grammatical rule, in a silent state from a partial sentence obtained from the matching result of the input signal and the standard word pattern and its matching score. The first condition that the matching score of the ending sub-sentence is maximum is satisfied, and
First determining means for determining whether or not there is a partial sentence satisfying the second condition that the section length of the input signal corresponding to silence at the end of the partial sentence is equal to or longer than a predetermined time; When it is determined that both conditions of 2 are satisfied, the matching score of the input signal and the separately defined reference pattern is set as the reference score, and the matching score obtained from the matching result of the input signal and the word standard pattern in the section before that. Second determination means for determining whether or not a third condition that the difference between the matching score of the partial sentence and the reference score is within a predetermined range is satisfied, and it is determined that the third condition is satisfied. At this time, a third means is provided for determining that the input signal of the section corresponding to the sub-sentence determined to satisfy both the first and second conditions is non-stationary noise, and correcting the beginning of the voice section. The voice section start correction device. 8. When the first, second and third conditions are satisfied, the section length of the input signal corresponding to the sub-sentence determined to satisfy both the first and second conditions is calculated, and The third means has means for judging whether or not a fourth condition that the section length conforms to the standard section length of the sub-sentence in a predetermined range is satisfied, and the third means, when the fourth condition is satisfied. When the determination is made, it is determined that the input signal of the section corresponding to the sub-sentence determined to satisfy all the first, second and third conditions is non-stationary noise, and the start end of the speech section is corrected. The voice section start end correction device according to claim 7. 9. The second determination means compares the matching scores in each state of the same sub-sentence with respect to the sub-sentence determined to satisfy both the first and second conditions instead of the third condition. Then, the maximum value is obtained, and the matching score between the input signal corresponding to the same sub-sentence and the separately determined reference pattern seen in the state of taking the maximum value is set as the reference score, and the difference between this reference score and the maximum value is Determining whether or not a fifth condition of being within a predetermined range is satisfied, and when the third means determines that the fifth condition is satisfied, both of the first and second conditions are satisfied. It is determined that the input signal in the section corresponding to the sub-sentence determined to be non-stationary noise,
The voice section start end correction device according to claim 7, wherein the start point of the voice section is corrected. 10. A section length of an input signal corresponding to the same sub-sentence found in a state where the maximum value is obtained among sub-sentences determined to satisfy both the first and second conditions, and the section length is No. 6 that fits the average section length of the same sub-sentence within a predetermined range
And a means for determining whether or not the condition is satisfied, and when the third means determines that the sixth condition is satisfied, it is determined that both the first, second and fifth conditions are satisfied. 10. The voice section start edge correction device according to claim 9, wherein the input signal in the section corresponding to the partial sentence is determined to be non-stationary noise, and the start edge of the voice section is corrected. 11. When starting a device for recognizing a voice by matching a standard word pattern with an input signal based on a grammar rule,
A voice which is characterized in that after detecting the start of a voice section by calculating the power of an input signal or the like, the detected start is corrected by the voice section start correction method according to any one of claims 1 to 6 to perform voice recognition. Recognition method. 12. When starting a device for recognizing a voice by matching a standard word pattern with an input signal based on grammatical rules,
The voice recognition and the voice section start end correction method according to any one of claims 1 to 6 are started without detecting the start point of the voice section based on the power calculation of the input signal, and the correction is performed by this voice section start end correction method. While performing voice recognition, without detecting the end of the voice section by calculating the power of the input signal, the matching score and parsing result of the partial sentence, which is obtained from the matching result of the input signal and the word standard pattern, and the silent standard. A voice recognition method, characterized in that the end of a voice section is detected based on the section length of an input signal determined to match a pattern and the voice recognition is ended.