JP2996019B2 - 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 method for recognizing an input voice including a pause in a voice recognition device used for a voice input device, an automatic interpreter, or the like.

[0002]

2. Description of the Related Art Voice is one of human-machine interfaces that are natural and easy for humans to use, and there is a strong demand for practical use of a question answering device with a computer by voice input and an automatic interpreter for voice input / output. In these devices, it is desired that a sentence in a natural language or a conversational sentence be input as naturally as possible. Conventionally, in order to recognize a voice in a signal input from a microphone, a telephone, or the like, these apparatuses have been disclosed in, for example, “Speech Recognition” by Yasunaga Niimi, Kyoritsu Shuppan Co., Ltd. From page 7
As shown on page 0, the start and end of a speech section to be recognized are determined using signal power information and the number of zero crossings, and recognition processing is performed on this speech section. When detecting the end of a voice section using power information, a certain period of time is needed to distinguish between a silent part such as a plosive in the voice or a short pause during utterance and a silent part after the voice ends. If the voice input has been continued for a longer time, it has been determined that the voice input has ended. In this case, by combining a silence pattern or a noise pattern before and after the standard pattern, a few silence sections or noise sections are included at both ends of the voice section so that correct matching can be performed in many cases.

[0003]

When a person naturally utters a sentence, a pause often occurs in the sentence, that is, the utterance is interrupted. For example, in the case of "Please give two tickets", you may think about the number after uttering "The ticket". In this case, it will be uttered like "Please give <pause> two tickets" . In the case of a long sentence, a pause may occur as a result of breathing in the middle of the sentence. In order to cope with this, in the related art, it is determined that the voice input has been completed only when a silent section or a noise section lasts for a certain time length L or more. Thus, even when a pause of less than or equal to the predetermined time L is included in the sentence, it is not erroneously determined that the voice input has ended, and the input voice including the utterance after the pause can be recognized.

However, in this method, the recognition result cannot be output until the above-mentioned predetermined time L has elapsed even if the voice input is completed. There is a disadvantage that the recognition result is not easily output after the input is completed. If the recognition result is required immediately after the voice input is completed, L
Cannot be set to a very large value, and in this case, there is a disadvantage that a long pause cannot be placed in the input voice when uttering, and as a result, a burden is imposed on the user of the recognition device, The device was difficult to use.

[0005] It is an object of the present invention to correctly recognize even if a long pause is placed in the middle of the input voice during voice input, and to output a recognition result immediately upon completion of voice input. It is another object of the present invention to provide a speech recognition apparatus capable of promptly rejecting a speech or a noise sound other than a recognition target when the speech or noise sound is input.

[0006]

According to a first aspect of the present invention, there is provided a speech recognition apparatus comprising: an analyzing unit for converting an input signal into a feature vector time series; A voice detection unit that detects the start and end of a voice section, compares and compares the feature vector time series with a pre-registered standard pattern, obtains a maximum similarity at each point of the input signal, and ends voice input. Sometimes, a comparison / matching unit that obtains a standard pattern that gives the maximum similarity as a recognition result, and at least one near the end of the voice section.
And a first input termination determination unit that outputs a voice input termination signal if the maximum similarity is greater than a first threshold at a certain point in time.

According to a second aspect of the present invention, there is provided a speech recognition apparatus for analyzing an input signal into a feature vector time series, and using a power information of the feature vector time series to start and end a speech section in the input signal. And a voice detection unit for detecting the maximum similarity at each time point of the input signal by comparing and comparing the feature vector time series with a pre-registered standard pattern, and giving the maximum similarity at the end of the voice input. A comparison / matching unit for obtaining a standard pattern as a recognition result; and a difference between a maximum similarity of the standard pattern and a maximum similarity of a partial pattern in the standard pattern at at least one point near the end of the voice section. Alternatively, a second input end determination unit that outputs a voice input end signal when the ratio is larger than the second threshold value.

In the speech recognition apparatus according to a third aspect of the present invention, in the first or second aspect, the maximum similarity of the partial pattern in the standard pattern is equal to the third similarity at the time of the input completion determination.
Output a reject signal if smaller than the threshold
Characterized in that it has a rejection section.

In a fourth aspect of the present invention, in the first or second aspect of the present invention, when the input completion determination unit determines,
A second reject unit that outputs a reject signal when a difference or a ratio between the maximum similarity of the standard pattern and the maximum similarity of the partial patterns in the standard pattern is smaller than a fourth threshold value. .

According to a fifth aspect of the present invention, in the speech recognition apparatus according to the first to fourth aspects, when the input end determination unit outputs a voice input end signal, the same partial pattern as the standard pattern giving the maximum similarity is provided. Is provided, an end signal delay unit for outputting a voice input end signal again after a certain time has elapsed from that point.

According to a sixth aspect of the present invention, in the speech recognition apparatus according to the first to fifth aspects, a similarity between a pattern obtained by combining standard patterns of recognition units in a predetermined order and a feature vector time series of the speech section. A reference similarity calculating unit that determines a maximum value as a reference similarity, and a third rejecting unit that outputs a reject signal when the reference similarity is smaller than a fifth threshold at the time of determination by the input end determination unit. It is characterized by having.

According to a seventh aspect of the present invention, in the speech recognition apparatus according to the first to sixth aspects, an inter-vector similarity between a feature vector constituting the standard pattern and a feature vector in a feature vector time series of the speech section is provided. An inter-vector similarity calculating unit for obtaining an accumulated value; and a fourth rejecting unit for outputting a reject signal when the reference similarity is smaller than a sixth threshold at the time of determination by the input end determining unit. And

The speech recognition apparatus according to an eighth aspect of the present invention is the speech recognition apparatus according to the first to seventh aspects, further comprising: A fifth reject unit that outputs a reject signal when the noise similarity is greater than a seventh threshold value at the time of the determination by the input end determination unit.

According to a ninth aspect of the present invention, in the first to eighth aspects of the present invention, the first, second, third, fourth, fifth, and sixth aspects of the speech recognition apparatus according to an elapsed time from the end of the voice section. And a threshold calculator for changing the seventh threshold.

According to a tenth aspect of the present invention, in the speech recognition apparatus according to the first to ninth aspects, an elapsed time measuring section for measuring an elapsed time from the time of the judgment by the input end judging section; And a sixth reject unit that outputs a reject signal when the reject signal is output when the voice detection unit does not detect the start end of the next voice section.

[0016]

When a person naturally utters a sentence, there is a pause in the sentence, that is, the utterance is often interrupted. However, if a voice section is detected only by using the power information of the input signal, a pause in the sentence is obtained. Was mistaken for silence after the end of the utterance, and it was not possible to correctly recognize the entire utterance including the sound following the pause. The speech recognition device of the present invention detects the end point of utterance by simultaneously using not only the power information of the input signal but also the similarity between the standard pattern of the speech to be recognized and the input signal. It is. Thus, even when a pause is included in the input voice, the pause is not mistaken for silence after the utterance ends.

In the first invention, first, an input signal is converted into a feature vector time series by an analyzer. For this analysis, it is possible to use a method based on mel cepstrum, a method based on LPC analysis, etc. shown on pages 32 to 98 of “Digital Speech Processing” (hereinafter referred to as Reference 2) published by Tokai University Press. it can.

Next, the speech detection section detects the start and end of the speech section in the input signal using the power information in the feature vector time series obtained by the analysis section. For this purpose, a voice detection method shown on pages 68 to 70 of Document 1 can be used. The voice detection unit detects the start of a section in which the power of the input signal is greater than or equal to a certain threshold and continues for a predetermined time or more as the start of a voice section. Further, when the power has continued for a certain period of time with the power dropped below a certain threshold value, the time point when the power dropped below the threshold value is detected as the end of the voice section.

The comparison / comparison section compares and compares the feature vector time series of the input signal after the start end detected by the voice detection section with a standard pattern to be recognized which has been registered in advance. The maximum value of the similarity between the input signal and the input signal, that is, the maximum similarity is calculated. Also, a standard pattern that gives the maximum similarity at the end of the input voice is output as a recognition result. At this time, an arbitrary sentence can be recognized by using, as a standard pattern, a unit voice pattern of syllables, semisyllables, words, or the like connected according to a prepared grammar. For example, in Japanese Patent Application No. 54-104669, "Continuous Speech Recognition Apparatus" (hereinafter referred to as Document 3) describes a method of recognizing continuous speech by connecting word patterns in accordance with a grammar represented by a finite state automaton. Have been.

The first input end determining unit determines that the maximum similarity calculated by the comparison / matching unit is smaller than a threshold value at the end point of the voice section detected by the voice detection unit or at least one point in the vicinity thereof. If it is larger, it is determined that the voice input has ended, and a voice input end signal is output.

For example, as standard patterns to be recognized, "Please give two tickets", "Please give three tickets" and "Please give you tickets" are registered. It is assumed that the pattern has not been registered.
At this time, it is assumed that a pause is made at the time of inputting "the ticket" when the voice "Please input two tickets" is input. The voice detecting unit detects the end of the "the ticket" as the end of the voice section due to the presence of this pause. At this time, the comparison / collation unit outputs the maximum similarity Si as a result of the comparison and comparison between the feature vector time series of “the ticket” and the standard pattern. Then, the maximum similarity Si at this point is a relatively small value because it is the result of comparison and collation with a word string different from the standard pattern.

On the other hand, when the voice "Please two" is input following the pause, the voice detection unit detects the start and end of the voice section again. At the end of this time, the comparison and collation unit says, "Please give me two <pause>tickets."
Is compared and compared with the standard pattern, so that the maximum similarity Sj with the standard pattern of "Please give two tickets", which is the same word string as the input voice, is a relatively large value. Becomes

Accordingly, by setting an appropriate threshold value in advance so that Si and Sj can be classified, the input end determination unit outputs a voice input end signal when the "up to that ticket" is input. On the other hand, it is possible to immediately output a voice input end signal when up to "Please enter two <pause>tickets". As a result, even for an input voice including a pause, it is possible to output a recognition result immediately after the sentence is completely input without terminating the voice recognition process at the position of the pause.

In order to compare and match the feature vector time series of the section including the pose with the standard field edge, a method of comparing and matching the feature pattern time series with the pause section removed in advance and the standard pattern, Alternatively, a method of inserting a silence model for modeling a feature vector time series of a pause section into a standard pattern is known.

[0025] In the second invention, the second input end determining unit determines the maximum similarity of the standard pattern with respect to the input voice at the end of the voice section or at least one point in the vicinity thereof. If the difference or ratio of the partial pattern with the maximum similarity is greater than the threshold value, it is determined that the voice input has ended, and a voice input end signal is output.

For example, for the standard pattern "Please give two tickets", "that", "this ticket", and "two tickets" are determined in advance as partial patterns. At this time, it is assumed that a pause is made at the time of inputting "the ticket" when the voice "Please input two tickets" is input. At this point, the comparison / matching unit outputs the maximum similarity Si between the feature vector time series of “the ticket” and the standard pattern, and also outputs the maximum similarity Pi between the same feature vector time series and the partial pattern. In this case, in the case of comparison and matching with the standard pattern, comparison and matching with a word string different from the standard pattern is performed.
Is a relatively small value. On the other hand, in the case of comparison and collation with a partial pattern, a large maximum similarity Pi is obtained in comparison with the partial pattern "the ticket". As a result, the difference Si-Pi between these maximum similarities generally has a relatively small value (in this case, a negative value).

On the other hand, when the voice of "please give two sheets" is input following the above pause, at the end of the voice section, the comparison and collation unit will say "Please give the ticket <pause> two sheets."
The maximum similarity Sj of the standard pattern with respect to the feature vector time series and the maximum similarity Pj between the same feature vector time series and the partial pattern are output. In this case, the maximum similarity Sj of the standard pattern has a relatively large value, whereas the maximum similarity Pj with the partial pattern has a relatively small value. As a result, the difference Sj-Pj between these maximum similarities becomes a relatively large value (in this case, a positive value).

Therefore, (Si-Pi) and (Sj-Pj)
By setting an appropriate threshold value in advance so that it can be classified, the input end determination unit does not output the voice input end signal at the time when "the ticket" is input, while the "the ticket" When <2><pause> please input>, a voice input end signal can be output immediately. As a result, similarly to the first aspect, even for an input voice including a pause, the voice recognition processing is not terminated at the position of the pause, and the recognition result is immediately generated when the sentence is completely input. It becomes possible to output.

When the similarity is represented by a probability value, for example, it is better to calculate the ratio than to calculate the difference between the maximum similarities.

In the third invention, when the input rejection unit determines whether or not the input is completed in the first reject unit, the maximum similarity between the input voice and the partial pattern in the standard pattern is determined by the threshold value. If it is smaller than the above, a reject signal is generated.

That is, when a voice other than the recognition target is input, the maximum similarity between the voice and the standard pattern has a small value, and therefore, in many cases, the input end determination unit outputs a voice input end signal. And waits for voice input. Therefore, according to the third aspect, when a speech to be recognized is input, the maximum similarity between the input speech and the partial pattern in the pause in the middle becomes a relatively large value, whereas the maximum similarity between the input speech and the partial pattern is relatively large. When a voice is input, the maximum similarity between the input voice and the standard pattern generally has a small value. Therefore, by setting an appropriate threshold, a reject signal can be output immediately.

In the fourth invention, the maximum resemblance of the standard pattern with respect to the input voice and the partial pattern in the standard pattern are determined when the input end determining unit determines whether or not the input is completed in the second reject unit. A reject signal is generated when the difference or ratio between the maximum similarity and the ratio is smaller than a threshold value.

That is, when a voice other than the recognition target is correctly input, the similarity between the input voice and the partial pattern becomes relatively large in the middle pause, so that the maximum similarity Si of the standard pattern and the partial pattern Is a relatively small value or a negative value as described above, whereas if a speech other than the recognition target is input, the partial pattern is compared with the input pattern. Is not particularly large, and Si-Pi does not have a very small value. Therefore, by appropriately setting the threshold value, it is possible to immediately output a reject signal when a voice other than the recognition target is input.

In the fifth invention, when the input end determining section outputs a speech input end signal, if there is a partial pattern such as a word string or a syllable string which is the same as the standard pattern giving the maximum similarity at that time, At that time, the output of the recognition result is temporarily postponed, and the recognition result is output by outputting a voice input signal again after a certain time has elapsed.

For example, “Yes, please give me cash”, “Yes, please give me 2 pieces of cash”, and “Yes, cash” are registered as standard patterns, and “Yes, please give me cash” as partial patterns. Is registered. At this time, assuming that a voice of "Yes, please give me two <pause> with cash" is input, and when "Yes, with cash" is input, a standard pattern "Yes, with cash" Gives a relatively large value of maximum similarity.
However, if the recognition process is terminated at this point and the recognition result is output, the "2 sheets please" input after this will be input.
Cannot be recognized, and an incorrect recognition result is output. Therefore, if there is a partial pattern that is the same as the standard pattern that gives the maximum similarity, the output of the recognition result is postponed until a certain time elapses. As a result, the entire input can be correctly recognized even when "Please two" is input after the pose after "Yes, with cash".
And when the input voice is only "Yes, cash",
The recognition result can be output after a certain time has elapsed.

In the sixth aspect, the similarity between a pattern in which recognition units such as words, syllables, and syllables are combined in a predetermined order, and the speech section of the input signal are obtained as a reference similarity. When it is determined whether the input has been completed or not, if the reference similarity is smaller than the threshold, a reject signal is output.

When an unexpected sound such as a noise sound other than a sound is input, the similarity with the standard pattern to be recognized becomes relatively small at the end of the sound. The end determination unit cannot determine that the voice input has ended, and waits for the input of the next voice as it is. On the other hand, the similarity between a pattern in which syllables or semisyllables are combined in an order that allows an arbitrary syllable string and a noise sound other than speech has a relatively small value. Therefore, by setting an appropriate threshold value, a reject signal can be output immediately when a noise sound is input.

The calculation of the reference similarity can be performed by the same method as the calculation of the similarity in the comparison / collation unit.

In the seventh invention, the cumulative value of the inter-vector similarity between the feature vector constituting the standard pattern and the feature vector of the input signal is obtained, and when the input end determining unit determines whether or not the input has been completed, And outputs a reject signal when the inter-vector similarity cumulative value is smaller than the threshold value.

That is, since the feature vectors constituting the standard pattern are generally feature vectors constituting a human voice, if noise noise other than voice is input, the inter-vector similarity cumulative value is a relatively small value. become. Therefore,
As in the sixth aspect, by setting an appropriate threshold value, a reject signal can be output immediately when a noise sound is input.

In the eighth invention, when the input end determining section determines whether or not the input has been completed, the similarity between the pattern of the noise sound prepared in advance and the input signal is obtained, and the similarity is smaller than the threshold value. Outputs a reject signal if larger.

As a result, a sound that is not assumed, such as a noise sound other than a sound, is input. At the end of the sound, the similarity with the standard pattern to be recognized becomes a relatively small value. The unit can immediately output a reject signal even when it is not possible to determine that the voice input has ended.

According to the ninth aspect, the threshold value for determining the end of the voice input in the input end determination unit and the threshold value for determining the rejection in the reject unit are determined by the elapsed time from the end point of the voice section to the determination point, that is, It is changed according to the duration of the pause up to the determination time. In this case, the input end determination and the rejection determination are performed at a plurality of times after the end of the voice section.

That is, if the duration of the pause is short, there is a high possibility that a voice will be continuously input after the pause, so that it is relatively difficult to determine that the voice input has ended as the threshold value for the voice input end determination. The threshold for reject determination is changed to a value that makes it relatively difficult to reject. On the other hand, if the duration of the pause is long, the possibility of continuous voice input is somewhat lower, so the threshold value for voice input end determination is set so that it is relatively easy to determine that voice input has ended. The threshold value for reject determination is changed to a value that is relatively easy to reject. As a result, if a pause is set during voice input and a voice to be recognized is input, the next voice can be accepted without rejecting even if a long pause is set. On the other hand, when a voice or noise sound other than the recognition target is input, it is possible to quickly reject even a short pause or end the recognition processing.

According to the tenth aspect, a reject signal is output when the next voice section does not start within a predetermined elapsed time from the time when the input completion determination section determines whether or not the input has been completed. Thus, when the voice input is interrupted on the way, the reject signal can be output without waiting for the next voice input.

[0046]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described in detail with reference to the drawings. FIG. 1 shows an embodiment of the present invention. The operation of the embodiment of FIG. 1 will be described.

The input signal is input to the analysis unit 1 and converted into a feature vector time series by feature analysis. The analysis here includes, for example, pages 60 to 64 of Document 2 and 1
A method based on LPC analysis as shown on page 55 can be used.

The feature vector time series is input to the voice detection unit 2, and the start and end of the voice section in the input signal are detected using the power information. To do this,
For example, the voice detection method shown on pages 68 to 70 of Document 1 can be used.

The detected start position of the speech section is input to the comparison / matching unit 3 together with the feature vector time series, and the feature vector time series after the start end is compared with a plurality of standard patterns registered in advance for recognition. Is done. As a result, the maximum similarity Si of the standard pattern at each time point of the input signal is obtained. As a method of calculating the similarity, a method shown in Reference 3 can be used. The comparison / collation unit 3 performs the comparison / collation until a voice input end signal is input from the input end determination unit.

When calculating the similarity between the standard pattern and the input signal, for example, as described in Japanese Patent Application No. 3-60786, "Speech Recognition Apparatus," By using a method of correcting or normalizing the similarity so as to prevent the similarity from fluctuating due to the influence of the above, a more accurate similarity can be obtained.

Next, when the voice detection section 2 detects the end of the voice section, the end position is input to the input end determination section 4. The input end determination unit 4 compares the maximum similarity Si of the standard pattern at the end point with the comparison / matching unit 3.
And compares the maximum similarity Si with the threshold value T1. As a result, if Si> T1, the voice input signal is output to the comparison / matching unit 3. If Si> T1, nothing is output to the comparison / collation unit 3.

When the voice input end signal is input to the comparison / comparison section 3, the comparison / comparison section 3 outputs a standard pattern giving the maximum similarity at that time as a recognition result.

As described above, according to the first aspect of the present invention, even for an input voice including a pause, the voice recognition processing is not terminated at the position of the pause, and immediately after the sentence is completely input. It is possible to output a recognition result.

According to the second aspect of the present invention, the comparison / matching section 3 obtains the maximum similarity Si of the standard pattern and the maximum similarity Pi of the partial pattern in the standard pattern at each time point of the input signal.

On the other hand, the input end determination section 4
, When the end of the voice section is detected, the maximum similarity Si of the standard pattern and the maximum similarity Pi of the partial pattern at the end point are received from the comparison / matching unit 3;
The difference between the difference Si-Pi and the threshold value T2 is compared.
As a result, if Si-Pi> T2, a voice input end signal is output to the comparison / collation unit 3. Otherwise, the comparison and collation unit 3
Does not output anything.

In this way, as in the first invention, even for an input voice including a pause, the sentence is completely input at the pause position without terminating the speech recognition processing. It becomes possible to output a recognition result immediately at a point in time.

According to the third aspect of the present invention, when the input end determination unit 4 determines whether or not the input is completed in the first reject unit 7, the maximum of the standard pattern obtained by the comparison / matching unit 3 is obtained. Comparing the magnitude of the similarity Si with the threshold T3,
If Si <T3, a reject signal is generated.

According to the fourth aspect, in the second rejection unit 8, when the input end determination unit 4 determines whether or not the input has ended, the maximum of the standard pattern obtained by the comparison / matching unit 3 is determined. Maximum similarity P between similarity Si and partial pattern
i, the difference Si-Pi is compared with a threshold value T4, and the Si-Pi
If <T4, a reject signal is generated.

According to the fifth aspect, the voice input end signal is output from the input end determination unit 4 to the end signal delay unit 6 instead of the comparison / comparison 3 temporarily. Upon receiving the voice input end signal, the end signal delay unit 6 inputs a standard pattern giving the maximum similarity from the comparison / matching unit 3. The end signal delay unit 6 checks whether a partial pattern of the same syllable string as the standard pattern exists, and if it exists, outputs a speech input end signal to the comparison / matching unit 3 after a predetermined time has elapsed. I do. If it does not exist, an audio input end signal is immediately output to the comparison and collation unit 3.

According to the sixth aspect, when the input end determining section 4 determines whether or not the input is completed, the reference similarity calculating section 9 sets the recognition unit such as a word, a syllable, a semi-syllable in advance. A plurality of patterns combined in a predetermined order are compared with a voice section of an input signal, and a reference similarity Ri is output. Next, the third reject unit 10 compares the magnitude of the reference similarity Ri with the threshold value T5, and generates a reject signal if Ri <T5.

According to the seventh aspect, when the input end determining unit 4 determines whether or not the input has been completed, the inter-vector similarity calculating unit 11 determines whether the feature vector constituting the standard pattern to be recognized is The cumulative value Di of the similarity between the input signal and the feature vector is output. Next, the fourth reject unit 12 compares the magnitude of the inter-vector similarity cumulative value Di with the threshold value T6, and generates a reject signal if Di <T6.

According to the eighth aspect, when the input end determining unit 4 determines whether or not the input has been completed, the noise similarity calculating unit 13 determines whether the input signal has a noise sound pattern prepared in advance. The similarity Ni is obtained. Next, the fifth reject unit 14 compares the magnitude of the similarity Ni with the threshold T7,
If Ni> T7, a reject signal is generated.

According to the ninth aspect, when the input end determining section 4 determines whether or not the input has been completed, the threshold value calculating section 5 obtains the time from the end of the voice section to the time of the determination. Thresholds T1 and T1 used in the determination unit 4, the first reject unit 7, the second reject unit 8, the third reject unit 10, the fourth reject unit 12, and the fifth reject unit 14.
Each of 2, T3, T4, T5, T6, and T7 is changed to a value corresponding to the elapsed time.

According to the tenth aspect, the sixth rejection unit 15 determines that the voice detection unit 2 has been turned on within a predetermined time after the input end determination unit 4 has determined whether or not the input has ended. If the beginning of the next voice segment is not detected,
Outputs a reject signal.

[0065]

As described above in detail, according to the present invention, the input voice can be correctly recognized even when a long pause is put in the middle of the input voice, and the voice input is completed. Recognition results can be output promptly at a point in time, and rejection can be made promptly when voice or noise other than the recognition target is input.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing one embodiment of the present invention.

[Explanation of symbols]

 Reference Signs List 1 analysis unit 2 voice detection unit 3 comparison and collation unit 4 input end determination unit 5 threshold calculation unit 6 end signal delay unit 7 first reject unit 8 second reject unit 9 reference similarity calculation unit 10 third reject unit 11 Inter-vector similarity calculation unit 12 Fourth reject unit 13 Noise similarity calculation unit 14 Fifth reject unit 15 Sixth reject unit

Continuation of the front page (56) References JP-A-5-66787 (JP, A) JP-A-4-184396 (JP, A) JP-A-3-167600 (JP, A) JP-A-62-111698 (JP) JP-A-61-282898 (JP, A) JP-A-3-71200 (JP, A) JP-A-1-185599 (JP, A) JP-A-63-5394 (JP, A) 63-97999 (JP, A) JP-A-2-178699 (JP, A) JP-A-3-122697 (JP, A) JP-A-3-1222698 (JP, A) (58) Fields investigated (Int. Cl. ^6, DB name) G10L 3/00 531 G10L 3/00 513 G10L 3/00 561 JICST file (JOIS)

Claims (10)

(57) [Claims] An analysis unit configured to convert an input signal into a feature vector time series; a speech detection unit configured to detect a beginning and an end of a speech section in the input signal using power information of the feature vector time series; The feature vector time series is compared with a pre-registered standard pattern to determine the maximum similarity at each point of the input signal, and to obtain a standard pattern that gives the maximum similarity as a recognition result at the end of voice input. A matching unit, wherein at least one point near the end of the voice section has the maximum similarity of the first
And a first input end determination unit that outputs a voice input end signal when the input end end is greater than a threshold value. 2. An analysis unit for converting an input signal into a feature vector time series, a speech detection unit for detecting the start and end of a speech section in the input signal using power information of the feature vector time series, The feature vector time series is compared with a pre-registered standard pattern to determine the maximum similarity at each point of the input signal, and to obtain a standard pattern that gives the maximum similarity as a recognition result at the end of voice input. A comparison unit, and at least one point in time near the end of the voice section, wherein a difference or ratio between a maximum similarity of the standard pattern and a maximum similarity of partial patterns in the standard pattern is smaller than a second threshold. A second input end determination unit that outputs a voice input end signal if the input is larger. 3. The method according to claim 1, wherein the maximum similarity of the partial pattern in the standard pattern is a third similarity at the time of determination by the input end determination unit.
Output a reject signal if smaller than the threshold
The voice recognition device according to claim 1, further comprising a reject unit. 4. When the difference or ratio between the maximum similarity of the standard pattern and the maximum similarity of a partial pattern in the standard pattern is smaller than a fourth threshold value at the time of determination by the input end determination unit, rejection is performed. The speech recognition device according to claim 1, further comprising a second reject unit that outputs a signal. 5. When the input end determination unit outputs a voice input end signal, if there is a partial pattern that is the same as the standard pattern that gives the maximum similarity, after a certain time has elapsed from that point in time. 5. The speech recognition device according to claim 1, further comprising an end signal delay unit for outputting a speech input end signal again. 6. A reference similarity calculating unit for obtaining, as a reference similarity, a maximum similarity between a pattern obtained by combining standard patterns of recognition units in a predetermined order and a feature vector time series of the speech section; The speech recognition device according to claim 1, further comprising a third reject unit that outputs a reject signal when the reference similarity is smaller than a fifth threshold value at the time of the determination by the end determination unit. 7. An inter-vector similarity calculation unit for obtaining an accumulated value of inter-vector similarity between a feature vector constituting the standard pattern and a feature vector in a feature vector time series of the voice section, and an input end determination unit. 7. The speech recognition device according to claim 1, further comprising: a fourth reject unit configured to output a reject signal when the reference similarity is smaller than a sixth threshold value in the determination. 8. A noise similarity calculating section for calculating a similarity between a pattern of a noise sound and a feature vector time series after the beginning of the voice section, and the noise similarity is determined to be 7th in the input end determining section. The speech recognition apparatus according to claim 1, further comprising: a fifth rejection unit that outputs a rejection signal when the value is larger than the threshold value. 9. The method according to claim 1, wherein the first, second, third, fourth, fifth, sixth, and seventh elapses from the end of the voice section.
10. A threshold calculator for changing a threshold value of the threshold value.
A speech recognition device according to claim 1. 10. An elapsed time measuring section for measuring an elapsed time from the time of determination by the input end determining section, and a case where the voice detecting section does not detect a start end of a next voice section within a predetermined elapsed time. The speech recognition apparatus according to any one of claims 1 to 9, further comprising a sixth reject unit that outputs a reject signal.