JP4699016B2 - Voice recognition device - Google Patents

Description

  The present invention relates to a speech recognition apparatus that registers speech uttered by a human and outputs a recognition result when speech having a high similarity to the registered speech is input.

  The voice recognition device refers to a system in which a machine automatically recognizes a human voice, and has high practicality in operation of a machine by voice. Conventional speech recognition devices are described in Non-Patent Document 1 and Non-Patent Document 2. In such a conventional speech recognition device, if the registered speech is inappropriate, such as the number of syllables of the registered speech is small, the number of vowels is small, or noise is included, the speech other than the registered speech May be misrecognized, and if the recognition target word is uttered under conditions different from the conditions for registering speech, it may not be recognized correctly.

  Moreover, there exists a technique shown by patent document 1-patent document 5 as a technique relevant to this application. In Patent Document 1, the correctness / incorrectness of the syllable identification result obtained at the time of speech input is instructed, the appearance frequency and the error frequency of each syllable are obtained based on the syllable identification result and the correct / incorrect instruction, and each syllable obtained By determining the syllables to be registered or re-registered in relation to the appearance frequency and the error frequency, it becomes possible to efficiently find and re-register syllables that require re-registration with a smaller amount of processing. It is described.

  Patent Document 2 includes a syllable dictionary that prestores the characteristics of a speaker's single syllable, and a reading sequence created by combining single syllables of the syllable dictionary when the speaker's input speech is registered in the speech dictionary; Comparing the degree of similarity with the speaker's input voice, and registering the input voice in the voice dictionary when it is similar to a predetermined threshold value or less, ensures that the voice registration that has the most influence on recognition is performed. Is described.

  Further, in Patent Document 3, when a vocabulary to be registered is input, the number of syllables of the input vocabulary is equal to or less than a predetermined number of syllables, or the input vocabulary and the number of syllables are equal and common to vowel arrangements. It is described that when a certain registered vocabulary exists, a vocabulary dictionary with a high recognition rate is created by determining that the registration is not appropriate.

  Furthermore, Patent Document 4 has a part for collecting sound, a part for converting sound into a pattern, an acoustic reproduction unit, a device for listening to reproduced sound in both ears of a speaker, and a part for registering sound. In addition, it is described that voice registration is performed so that the recognition rate does not decrease even under noise by registering voice while letting a speaker hear a specific sound.

  Furthermore, in Patent Document 5, when a plurality of words are continuously recognized, voice input and echo back are performed in a unit of a plurality of words, and when a recognition error occurs in a recognized word, that is, a speaker is When a negative response is made to echo back, echo back is performed in units of words, and the speaker's confirmation response is input for each word. It is described that the recognition rate does not decrease even if the number of words increases, and that the recognition rate is improved from the first time when re-input.

L. RABINER, B.M. H. JUANG, translated by Sadaaki Furui, “Basics of Speech Recognition” (up and down), NTT Advanced Technology, November 1995, p. 1-291 (top), p. 1-322 (bottom) Sadaaki Furui, “Speech Information Processing”, Morikita Publishing Co., Ltd., June 1998, p. 79-132 JP 59-157699 (page 5, upper right column, line 18 to page 5, lower left column, line 7) JP-A-63-149698 (2nd page, upper right column, 16th to 2nd page, lower left column, 3rd line, 3rd page, upper right column, 4th to 3rd page, upper right column, 5th line) JP 63-292197 (page 3, upper left column, line 14 to page 3, upper right column, line 6) JP-A-2-154299 (page 2, upper left column, line 19 to page 2, upper right column, line 3, page 3, upper right column, line 18 to page 3, upper right column, line 19) JP-A-60-260095 (page 2, upper left column, line 10 to page 2, upper left column, line 16, page 3, lower left column, line 4 to page 3, lower left column, line 7)

  Since the conventional speech recognition device is configured as described above, if the registered speech is inappropriate, such as the number of syllables of the registered speech is small, the number of vowels is low, or noise is included, There was a problem that the recognition accuracy was lowered because a sound other than the registered voice was erroneously recognized. In addition, when a recognition target word is uttered under conditions different from the conditions for registering speech, there is a problem in that recognition is not possible and recognition accuracy is reduced.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a speech recognition device that can improve recognition accuracy by preventing inappropriate speech from being registered. . It is another object of the present invention to provide a speech recognition apparatus that can improve the recognition accuracy even when the conditions for speech registration and speech recognition are different.

The speech recognition apparatus according to the present invention includes a speech change degree calculating means for inputting a registered speech of two or more syllables and calculating a sound change degree over the whole speech section, and a sound change degree calculated by the sound change degree calculating means. A voice registration determination unit that determines whether or not to register the input registered voice by comparing with an average value of the voice change degree of a word of a predetermined syllable, and a determination result by the voice registration determination unit is unregisterable A registration voice change request means for outputting a registration voice change request; and a voice standard pattern generation means for generating a voice standard pattern from the input registered voice when the determination result by the voice registration determination means is registerable. It is a thing.

  According to the present invention, it is possible to reduce misrecognition caused by improper registered speech and to improve the recognition accuracy.

An embodiment of the present invention will be described below.
Embodiment 1 FIG.
1 is a block diagram showing a configuration of a speech recognition apparatus according to Embodiment 1 of the present invention. The speech recognition apparatus includes a registration unit 100 and a verification unit 300. The registration unit 100 includes a voice change degree calculation unit 111, a voice registration determination unit 112, a registered voice change request unit 113, a voice registration switch 114, a voice standard pattern generation unit 115, and a voice standard pattern storage unit 116. A similarity calculation unit 311 and a voice collation determination unit 312 are provided.

  In the registration means 100, the voice change degree calculation means 111 inputs the registered voice 11 uttered by the user and calculates the voice change degree. The voice registration determination unit 112 determines whether or not to register the input registered voice 11 based on the voice change degree calculated by the voice change degree calculation unit 111. The registered voice change request unit 113 outputs a registered voice change request 12 to the user when the result of determination by the voice registration determination unit 112 is unregisterable. The voice registration determination unit 112 connects the voice registration switch 114 when the determination result can be registered, and the voice standard pattern generation unit 115 inputs the registered voice 11 and generates a voice standard pattern to store the voice standard pattern storage unit. 116.

  In the matching unit 300, the speech similarity calculation unit 311 calculates the speech similarity by inputting the recognition target speech 31 uttered by the user and the speech standard pattern stored in the speech standard pattern storage unit 116. The voice collation determination unit 312 outputs the recognition result 33 when the voice similarity calculated by the voice similarity calculation unit 311 is equal to or greater than a preset threshold value 32.

  In the first embodiment, the voice change degree calculation means 111, the voice registration judgment means 112, the registered voice change request means 113, the voice standard pattern generation means 115, the voice similarity degree calculation means 311 and the voice collation judgment means 312 are hard-wired. However, it is also possible to create a speech recognition program that describes the processing contents of each means and allow the computer to execute the speech recognition program.

Next, the operation will be described.
FIG. 2 is a flowchart showing the flow of processing of the speech recognition apparatus according to Embodiment 1 of the present invention. First, at the time of voice registration, in step ST101, the voice change degree calculating means 111 inputs the registered voice 11 uttered by the user, and in step ST102, the voice change degree calculating means 111 calculates the voice change degree.

ここで、ｃ_t は時刻ｔのフレームのケプストラム、ｃ＾は音声区間のケプストラム平均、Ｎは音声区間のフレーム数であり、Ｔは転置行列を示している。なお、ケプストラムについては上記非特許文献２の２−２節に記載されている。 Here, the registered voice 11 is a voice of a word or sentence uttered by the user to register the voice. At the time of voice recognition, a recognition result 33 is output when an utterance having a high similarity to the registered voice is performed. The voice change degree indicates how much the voice has changed in one utterance, and the voice change degree becomes larger as the number of different syllables increases. For example, the voice change degree is larger for the voice of “Aiueo” than for the voice of “A”. As an extraction method, for example, the input signal of the registered voice 11 is analyzed for every 10 ms frame, and is obtained from the degree of change of the extracted spectral component. In speech recognition, a cepstrum is often used as an efficient representation of the spectral components of speech, and the degree of speech change d using this cepstrum is obtained by the following equation (1).

Here, c _t is the cepstrum of the frame at time t, c ^ is the cepstrum average of the speech segment, N is the number of frames in the speech segment, and T is the transposed matrix. The cepstrum is described in Section 2-2 of Non-Patent Document 2.

  In step ST103, the voice registration determination means 112 receives the voice change degree d calculated by the voice change degree calculation means 111, and determines whether or not the voice change degree d is equal to or greater than a predetermined battle value THd. It is determined whether or not the input registration voice 11 is registered. As a determination method, it is determined that the registered voice 11 can be registered when the voice change degree d is equal to or greater than a predetermined threshold THd. On the other hand, when the voice change degree d is smaller than the threshold value THd, it is determined that the registered voice 11 cannot be registered. To do. For example, if the threshold THd is set as an average of the voice change d of words having three or more syllables, registration is possible if the word has three or more syllables.

  If the result of determination by the voice registration determination unit 112 is not registered in step ST103, the registration voice change request unit 113 outputs a registration voice change request 12 to the user in step ST104. When the voice change degree d is small, it is determined that the number of syllables of the registered voice 11 is small. Therefore, when the voice standard pattern is generated using the registered voice 11 and the voice recognition is performed, the amount of information Misrecognition increases because there are few. Therefore, the registration voice change request 12 is output. As a request from the speech recognition apparatus, for example, a message “Please say a word with a large number of syllables” is transmitted to the user by a synthesized sound, and the process returns to step ST101 to repeat the registration process.

  On the other hand, in step ST103, if the determination result by the voice registration determination unit 112 can be registered, in step ST105, the voice registration determination unit 112 connects the voice registration switch 114, and the voice standard pattern generation unit 115 The registered voice 11 is input, a voice standard pattern is generated and stored in the voice standard pattern storage means 116. The generated speech standard pattern efficiently expresses the characteristics of the input registered speech 11. For example, the cepstrum obtained by analyzing every frame of 5 to 10 ms, the dynamic feature amount of the cepstrum, and It is a time series of logarithmic power dynamic features. Also, a speech standard pattern can be generated by learning a Hidden Markov Model (HMM) using a time series of cepstrum, cepstral dynamic features and logarithmic power dynamic features. Since the hidden Markov model is described in Section 5-4 of Non-Patent Document 2, the description thereof is omitted.

  At the time of speech recognition, in step ST106, the speech similarity calculation unit 311 inputs the recognition target speech 31 uttered by the user. In step ST107, the speech similarity calculation unit 311 is stored in the speech standard pattern storage unit 116. The voice similarity with respect to the recognition target voice 31 is calculated using the existing voice standard pattern. Here, the recognition target voice 31 is a voice uttered by the user for causing the voice recognition apparatus to perform voice recognition, and the degree of voice similarity is the standard voice pattern registered in the recognition target voice 31 and the voice recognition apparatus. It is an index indicating whether they are similar. If the speech standard pattern is a time series of cepstrum, dynamic features of cepstrum, and logarithmic power, the cepstrum, cepstrum dynamic features, and logarithmic power dynamic features of the recognition target speech 31 as well. The speech similarity can be calculated by, for example, dynamic programming. The details of the dynamic programming method are described in Section 5-3 of Non-Patent Document 2, and thus the description thereof is omitted.

  In step ST108, the voice collation determination unit 312 determines whether or not the voice similarity calculated by the voice similarity calculation unit 311 is equal to or greater than a predetermined threshold 32. If the voice similarity is greater than or equal to the predetermined threshold value 32 in step ST108, the voice collation determination unit 312 outputs the voice corresponding to the voice standard pattern stored in the voice standard pattern storage unit 116 in step ST109. It is determined that the user uttered, and the recognition result 33 is output. On the other hand, if the voice similarity is smaller than the threshold value 32 in step ST108, the voice collation determination unit 312 returns to step ST106 without outputting the recognition result 33.

  Here, the case of only one utterance has been described. However, at the time of voice registration, a voice standard pattern is generated for each of a plurality of different utterances, and at the time of voice recognition, a plurality of standard patterns are used for the recognition target voice 31. It is also possible to calculate the degree and output the utterance number indicating the standard pattern having the largest voice similarity with a voice similarity of 32 or more as a recognition result.

  As described above, according to the first embodiment, the voice change degree calculation means 111 calculates the voice change degree of the registered voice 11, and the voice registration determination means 112 determines the calculated voice change degree as a predetermined battle. By registering the input registered voice 11 when the value is greater than or equal to the value, it is possible to prevent registration of a voice having a small number of syllables, and misrecognition caused by an inappropriate registered voice 11, that is, the number of syllables. It is possible to reduce misrecognition caused by registering a small amount of voice, and to improve the recognition accuracy.

Embodiment 2. FIG.
FIG. 3 is a block diagram showing the configuration of a speech recognition apparatus according to Embodiment 2 of the present invention. The speech recognition apparatus includes a registration unit 100 and a verification unit 300. The registration unit 100 includes a syllable number extraction unit 121, a voice registration determination unit 122, a registered voice change request unit 113, a voice registration switch 114, a voice standard pattern generation unit 115, and a voice standard pattern storage unit 116. Degree calculation means 311 and voice collation judgment means 312 are provided.

  In the registration means 100, the syllable number detection means 121 inputs the registered voice 11 uttered by the user and extracts the number of syllables. The voice registration determination unit 122 determines whether or not to register the registered voice 11 input based on the syllable number extracted by the syllable number extraction unit 121. The registered voice change request unit 113 outputs the registered voice change request 12 to the user when the result of determination by the voice registration determination unit 122 is unregisterable. When the determination result can be registered, the voice registration determination unit 122 connects the voice registration switch 114, and the voice standard pattern generation unit 115 inputs the registered voice 11, generates a voice standard pattern, and generates a voice standard pattern. Store in the storage means 116.

  In the matching unit 300, the functions of the voice similarity calculation unit 311 and the voice matching determination unit 312 are the same as those shown in FIG.

  In the second embodiment, the syllable number extraction unit 121, the voice registration determination unit 122, the registered voice change request unit 113, the voice standard pattern generation unit 115, the voice similarity calculation unit 311 and the voice collation determination unit 312 are provided in hardware. However, it is also possible to create a voice recognition program that describes the processing contents of each means and cause the computer to execute the voice recognition program.

Next, the operation will be described.
FIG. 4 is a flowchart showing the flow of processing of the speech recognition apparatus according to Embodiment 2 of the present invention. First, at the time of voice registration, in step ST201, the syllable number extracting means 121 inputs the registered voice 11 uttered by the user, and in step ST202, the syllable number extracting means 121 extracts the syllable number. Here, the syllable is a speech unit corresponding to one Hiragana character in Japanese.

  The processing of the syllable number extracting unit 121 will be described using speech recognition in syllable units. In order to perform speech recognition in syllable units, a standard pattern in syllable units is prepared in advance. As a standard pattern, for example, a hidden Markov model (HMM) is used, and model parameters are learned using a large amount of speech.

  FIG. 5 is a diagram for explaining speech recognition in syllable units. When the hidden Markov model of the syllable “A” is Λa and the hidden Markov model of the syllable “I” is Λi, the speech recognition per syllable is performed as shown in FIG. A similarity calculation process is performed on the registered speech 11 in accordance with the connection network, and the combination of syllables having the highest similarity is output as a recognition result in syllable units. Then, the number of syllables is obtained by extracting the number of syllables of the output syllable recognition result. For example, if the syllable unit recognition result is “good morning”, the number of syllables is four. In FIG. 5, black circles indicate nodes, and arrows indicate transition directions.

  In step ST203, the speech registration determination means 122 determines whether or not to register the input registered speech 11 by determining whether or not the number of syllables extracted by the syllable number extraction means 121 is greater than or equal to a predetermined number. Judge whether. As a determination method, when the number of extracted syllables is equal to or larger than a predetermined number, it is determined that the registered speech 11 can be registered, and on the other hand, when the number of extracted syllables is smaller than a predetermined number. Determines that the registered voice 11 cannot be registered. For example, the predetermined number is 3, and it is determined that 3 or more syllables can be registered.

  In step ST203, if the determination result by the voice registration determination unit 122 cannot be registered, the registration voice change request unit 113 outputs the registration voice change request 12 to the user in step ST204. When the speech standard pattern is generated using the registered speech 11 having a small number of syllables and the speech recognition is performed, the amount of information is small, so that misrecognition increases. Therefore, the registration voice change request 12 is output. As a request from the speech recognition apparatus, for example, a message “Please say a word with a large number of syllables” is transmitted to the user by synthetic sound, and the process returns to step ST201 to repeat the registration process.

  On the other hand, if the determination result by the voice registration determination unit 122 can be registered in step ST203, the voice registration determination unit 122 connects the voice registration switch 114 and the voice standard pattern generation unit 115 registers in step ST205. A voice standard pattern is generated by inputting the voice 11 and stored in the voice standard pattern storage means 116.

  The processing of speech similarity calculation means 311 in steps ST206 and ST207 at the time of speech recognition and the processing of speech collation determination means 312 in steps ST208 and ST209 are the same as the processing shown in FIG. 2 of the first embodiment.

  Here, the case of only one utterance has been described. However, at the time of voice registration, a voice standard pattern is generated for each of a plurality of different utterances, and at the time of voice recognition, a plurality of standard patterns are used for the recognition target voice 31. It is also possible to calculate the degree and output the utterance number indicating the standard pattern having the largest voice similarity with a voice similarity of 32 or more as a recognition result.

  As described above, according to the second embodiment, the syllable number extracting unit 121 extracts the number of syllables of the registered speech 11, and the speech registration determining unit 122 determines that the number of extracted syllables is equal to or greater than a predetermined number. By registering the registered voice 11, it is possible to prevent registration of a voice having a small number of syllables of the registered voice 11, and a misrecognition caused by an inappropriate registered voice 11, that is, a voice having a small number of syllables is registered. Misrecognition caused by this can be reduced, and the recognition accuracy can be improved.

Embodiment 3 FIG.
FIG. 6 is a block diagram showing the configuration of a speech recognition apparatus according to Embodiment 3 of the present invention. The speech recognition apparatus includes a registration unit 100 and a verification unit 300. The registration unit 100 includes a vowel likelihood calculation unit 131, a voice registration determination unit 132, a registered voice change request unit 113, a voice registration switch 114, a voice standard pattern generation unit 115, and a voice standard pattern storage unit 116. A similarity calculation unit 311 and a voice collation determination unit 312 are provided.

  In the registration means 100, the vowel likelihood calculation means 131 inputs the registered speech 11 uttered by the user and calculates the vowel likelihood. The voice registration determination unit 132 determines whether or not to register the input registered voice 11 based on the vowel likelihood calculated by the vowel likelihood calculation unit 131. The registered voice change request unit 113 outputs the registered voice change request 12 to the user when the determination result of the voice registration determination unit 132 is not registerable. When the determination result can be registered, the voice registration determination unit 132 connects the voice registration switch 114, and the voice standard pattern generation unit 115 inputs the registered voice 11 and generates a voice standard pattern to generate a voice standard pattern. Store in the storage means 116.

  In the matching unit 300, the functions of the voice similarity calculation unit 311 and the voice matching determination unit 312 are the same as those shown in FIG.

  In the third embodiment, the vowel likelihood calculation means 131, the voice registration determination means 132, the registered voice change request means 113, the voice standard pattern generation means 115, the voice similarity calculation means 311 and the voice collation determination means 312 are hard-wired. However, it is also possible to create a speech recognition program that describes the processing contents of each means and allow the computer to execute the speech recognition program.

Next, the operation will be described.
FIG. 7 is a flowchart showing the flow of processing of the speech recognition apparatus according to Embodiment 3 of the present invention. First, at the time of voice registration, in step ST301, the vowel likelihood calculating means 131 inputs the registered voice 11 uttered by the user, and in step ST302, the vowel likelihood calculating means 131 calculates the vowel likelihood. Here, a vowel is a sustainable resonance sound that flows through the center of the mouth without making a roar due to interference in the mouth. “Introduction to phonetics” Koizumi Tadashi, September 1996, Shorin, the University of Tokyo P. 85. In Japanese, “A (/ a /)”, “I (/ i /)”, “U (/ u /)”, “E (/ e /)”, “O / o /” are equivalent to vowels. To do.

  A method for calculating the vowel likelihood will be described using an HMM. First, a large amount of HMM for each vowel “A (/ a /)”, “I (/ i /)”, “U (/ u /)”, “E (/ e /)”, “O / o /” Create using the voice.

  FIG. 8 is a diagram for explaining speech recognition in syllable units. When the hidden Markov model of syllable “A” is Λa and the hidden Markov model of syllable “I” is Λi, these vowel HMMs are connected as shown in FIG. Calculation is performed (likelihood calculation method is described in Chapter 5 of Non-Patent Document 2). A value P ′ obtained by dividing the obtained likelihood P by the number N of frames and normalized is set as a vowel likelihood calculated by the vowel likelihood calculating means 131.

  In step ST303, the speech registration determination unit 132 registers the input registered speech 11 by determining whether or not the vowel likelihood P ′ calculated by the vowel likelihood calculation unit 131 is equal to or greater than a predetermined threshold. It is determined whether or not. As a determination method, when the calculated vowel likelihood P ′ is equal to or greater than a predetermined threshold, it is determined that the registered speech 11 can be registered. On the other hand, the calculated vowel likelihood P ′ is greater than a predetermined threshold. If it is smaller, it is determined that the registered voice 11 cannot be registered.

  If the result of determination by the voice registration determination unit 132 is not registered in step ST303, the registration voice change request unit 113 outputs a registration voice change request 12 to the user in step ST304. When the vowel likelihood P ′ is low, it is determined that the percentage of vowels included in the registered speech 11 is low. Therefore, when speech recognition is performed by generating a speech standard pattern using such registered speech 11 In some cases, misrecognition increases because the amount of information is small. Therefore, the registration voice change request 12 is output. As a request from the speech recognition apparatus, for example, a message “Please say a word with a lot of vowels” is transmitted to the user by a synthesized sound, and the process returns to step ST301 to repeat the registration process.

  On the other hand, if the determination result by the voice registration determination unit 132 can be registered in step ST303, the voice registration determination unit 132 is connected to the voice registration switch 114 and the voice standard pattern generation unit 115 is set in step ST305. The registered voice 11 is input, a voice standard pattern is generated and stored in the voice standard pattern storage means 116.

  The processing of speech similarity calculation means 311 in steps ST306 and ST307 at the time of speech recognition, and the processing of speech collation determination means 312 in steps ST308 and ST309 are the same as those shown in FIG. 2 of the first embodiment.

  Here, the case of only one utterance has been described. However, at the time of voice registration, a voice standard pattern is generated for each of a plurality of different utterances, and at the time of voice recognition, a plurality of standard patterns are used for the recognition target voice 31. It is also possible to calculate the degree and output the utterance number indicating the standard pattern having the largest voice similarity with a voice similarity of 32 or more as a recognition result.

  As described above, according to the third embodiment, the vowel likelihood calculating unit 131 inputs the registered speech 11 to calculate the vowel likelihood, and the speech registration determining unit 132 is configured to set a threshold value with a predetermined vowel likelihood. By registering the above registered voice 11, it is possible to prevent registration of a voice with a low vowel ratio of the registered voice 11, and misrecognition caused by inappropriate registration voice 11, that is, a low vowel ratio. There is less misrecognition caused by the registration of voice, and the recognition accuracy can be improved.

Embodiment 4 FIG.
FIG. 9 is a block diagram showing a configuration of a speech recognition apparatus according to Embodiment 4 of the present invention. The speech recognition apparatus includes a registration unit 100 and a verification unit 300. The registration unit 100 includes an audio standard pattern generation unit 115, an audio standard pattern storage unit 116, an ambient sound similarity calculation unit 141, an audio registration determination unit 142, and a registered audio change request unit 113, and the collation unit 300 includes an audio similarity calculation unit. 311 and voice collation determination means 312.

  In the registration unit 100, the voice standard pattern generation unit 115 receives the registered voice 11 uttered by the user, generates a voice standard pattern, and stores it in the voice standard pattern storage unit 116. The ambient sound similarity calculation unit 141 calculates the ambient sound similarity with respect to the ambient sound 13 using the speech standard pattern stored in the speech standard pattern storage unit 116. The voice registration determination unit 142 determines whether or not to register the registered voice 11 by determining whether or not the ambient sound similarity calculated by the ambient sound similarity calculation unit 141 is greater than a predetermined threshold. . The registered voice change request unit 113 outputs the registered voice change request 12 to the user when the result of determination by the voice registration determination unit 142 is unregisterable.

  In the matching unit 300, the functions of the voice similarity calculation unit 311 and the voice matching determination unit 312 are the same as those shown in FIG.

  In the fourth embodiment, the voice standard pattern generation means 115, the ambient sound similarity calculation means 141, the voice registration determination means 142, the registered voice change request means 113, the voice similarity calculation means 311 and the voice collation determination means 312 are provided. Although it may be configured by hardware, a voice recognition program describing the processing contents of each means may be created, and the computer may execute the voice recognition program.

Next, the operation will be described.
FIG. 10 is a flowchart showing the flow of processing of the speech recognition apparatus according to Embodiment 4 of the present invention. First, at the time of voice registration, in step ST401, the voice registration pattern generation means 115 inputs the registered voice 11 uttered by the user. In step ST402, the voice registration pattern generation means 115 generates a voice standard pattern from the registration voice 11. And stored in the voice standard pattern storage means 116.

  In step ST403, the ambient sound similarity calculating unit 141 calculates the ambient sound similarity with respect to the ambient sound 13 using the speech standard pattern stored in the speech standard pattern storage unit 116. Here, the ambient sound is a sound at a place where the voice recognition device is installed. The processing executed by the ambient sound similarity calculation unit 141 is performed immediately after the registered voice 11 is uttered or at a constant time interval, for example, an hour interval.

  In step ST404, the voice registration determination unit 142 registers the input registered voice 11 by determining whether or not the ambient sound similarity calculated by the ambient sound similarity calculation unit 141 is larger than a predetermined threshold. It is determined whether or not. As a determination method, when the ambient sound similarity is equal to or lower than a predetermined threshold, it is determined that the registered sound 11 can be registered. On the other hand, when the ambient sound similarity is larger than the predetermined threshold, the registered sound 11 is determined. Is determined to be unregisterable.

  If the result of determination by the voice registration determination unit 142 is not registerable in step ST404, the registration voice change request unit 113 outputs a registration voice change request 12 to the user in step ST405. If the ambient sound similarity calculated by the ambient sound similarity calculating unit 141 in step ST404 is greater than the threshold value, the speech standard pattern generated by the registered speech 11 may erroneously recognize the ambient sound 13. Is expensive. Therefore, the registration voice change request 12 is output. The request from the speech recognition device is, for example, a message “Please say another word” is transmitted to the user by synthetic sound, and the process returns to step ST401 to repeat the registration process.

  On the other hand, in step ST404, if the determination result by the voice registration determination unit 142 can be registered, the registration ends.

  The processing of the speech similarity calculation means 311 in steps ST406 and ST407 at the time of speech recognition and the processing of the speech collation determination means 312 in steps ST408 and ST409 are the same as those shown in FIG. 2 of the first embodiment.

  Here, the case of only one utterance has been described. However, at the time of voice registration, a voice standard pattern is generated for each of a plurality of different utterances, and at the time of voice recognition, a plurality of standard patterns are used for the recognition target voice 31. It is also possible to calculate the degree and output the utterance number indicating the standard pattern having the largest voice similarity with a voice similarity of 32 or more as a recognition result.

  As described above, according to the fourth embodiment, the ambient sound similarity calculating unit 141 calculates the ambient sound similarity with respect to the ambient sound by using the speech standard pattern generated by the input registered speech 11, and the speech When the calculated ambient sound similarity is equal to or less than the battle value, the registration determining unit 142 can prevent registration of a voice standard pattern that easily misidentifies the ambient sound by registering the input registered speech 11. Misrecognition caused by the sound 11 being inappropriate, that is, misrecognition caused by registering a sound similar to the surrounding sound can be reduced, and the effect of improving the recognition accuracy can be obtained. .

Embodiment 5 FIG.
FIG. 11 is a block diagram showing the structure of a speech recognition apparatus according to Embodiment 5 of the present invention. The speech recognition apparatus includes a registration unit 100 and a verification unit 300. The registration unit 100 includes an audio standard pattern generation unit 115, an audio standard pattern storage unit 116, an ambient sound similarity calculation unit 141, an ambient sound reproduction determination unit 151, an ambient sound reproduction switch 152, an ambient sound reproduction unit 153, and a registered audio change request unit. 113, the collation unit 300 includes a speech similarity calculation unit 311 and a speech collation determination unit 312.

  In the registration unit 100, the voice standard pattern generation unit 115 receives the registered voice 11, generates a voice standard pattern, and stores it in the voice standard pattern storage unit 116. The ambient sound similarity calculation unit 141 calculates the ambient sound similarity with respect to the ambient sound 13 using the speech standard pattern stored in the speech standard pattern storage unit 116. The ambient sound reproduction determination unit 151 determines to reproduce the ambient sound when the ambient sound similarity calculated by the ambient sound similarity calculation unit 141 is equal to or greater than a predetermined threshold, and connects the ambient sound reproduction switch 152. The ambient sound reproducing means 153 reproduces the ambient sound 13 and outputs the reproduced sound 14 when the determination result by the ambient sound reproduction determining means 151 is reproduction. The registered voice change request means 113 inputs the noise confirmation 15 and outputs the registered voice change request 12 to the user when the user who has heard the reproduced sound determines that the sound is noise.

  In the matching unit 300, the functions of the voice similarity calculation unit 311 and the voice matching determination unit 312 are the same as those shown in FIG.

  In the fifth embodiment, the sound standard pattern generation means 115, the ambient sound similarity calculation means 141, the ambient sound reproduction determination means 151, the ambient sound reproduction means 153, the registered sound change request means 113, and the sound similarity calculation means 311. The voice collation determination means 312 may be configured by hardware, but a voice recognition program describing the processing contents of each means may be created and the computer may execute the voice recognition program.

Next, the operation will be described.
FIG. 12 is a flowchart showing the processing contents of the speech recognition apparatus according to Embodiment 5 of the present invention. First, at the time of voice registration, in step ST501, the voice standard pattern generation means 115 inputs the user's registered voice 11, and in step ST502, the voice standard pattern generation means 115 generates a voice standard pattern from the registered voice 11 and generates a voice. Stored in the standard pattern storage means 116.

  In step ST503, the ambient sound similarity calculating unit 141 calculates the ambient sound similarity with respect to the ambient sound 13 using the speech standard pattern stored in the speech standard pattern storage unit 116.

  In step ST504, the ambient sound reproduction determination unit 151 reproduces the ambient sound by determining whether or not the ambient sound similarity calculated by the ambient sound similarity calculation unit 141 is equal to or greater than a predetermined threshold value. Determine whether or not to ask. As a determination method, when the ambient sound similarity is equal to or higher than a predetermined threshold, it is determined to reproduce the ambient sound.

  If it is determined in step ST504 that the ambient sound is to be reproduced, in step ST505, the ambient sound reproduction determination means 151 is connected to the ambient sound reproduction switch 152, and the ambient sound reproduction means 153 inputs the ambient sound 13 and plays the reproduced sound. By outputting 14, the user hears the reproduced sound 14.

  In step ST506, the user listens to the reproduced sound 14, determines that it is noise if it differs from the registered utterance, and notifies the speech recognition apparatus of the determination result of noise confirmation 15, and in step ST507, the registered voice change request means 113 inputs the noise confirmation 15 and outputs a registered voice change request 12 to the user. The request from the voice recognition device is, for example, a message “Please say another word” is transmitted to the user by synthesized sound, and the process returns to step ST501 to repeat the registration process. In this way, since the registered voice change request 12 is output only when the user hears the reproduced sound 14 and judges that it is noise, the ambient sound similarity calculating means 141 happens to be inadvertent when the ambient sound similarity is calculated. When the same voice as the registered voice 11 is uttered, the registered voice change request 12 is output, and the voice standard pattern is not changed by mistake.

  In step ST504, it is determined that the determination result of the ambient sound reproduction determination unit 151 is not required to be reproduced, or in step ST506, it is determined that the reproduction sound 14 output from the ambient sound reproduction unit 153 is not noise. In the case, the voice registration process is terminated.

  The processing of speech similarity calculation means 311 in steps ST508 and ST509 at the time of speech recognition and the processing of speech collation determination means 312 in steps ST510 and ST511 are the same as those shown in FIG. 2 of the first embodiment.

  Here, the case of only one utterance has been described. However, at the time of voice registration, a voice standard pattern is generated for each of a plurality of different utterances, and at the time of voice recognition, a plurality of standard patterns are used for the recognition target voice 31. It is also possible to calculate the degree and output the utterance number indicating the standard pattern having the largest voice similarity with a voice similarity of 32 or more as a recognition result.

  As described above, according to the fifth embodiment, the ambient sound similarity calculation unit 141 calculates the ambient sound similarity with respect to the ambient sound using the speech standard pattern generated by the input registered speech 11, and The sound reproduction determination unit 151 reproduces the ambient sound when the calculated ambient sound similarity is equal to or greater than the threshold, and inputs the changed registered sound 11 from the user who confirmed that the ambient sound is noise. Therefore, it is possible to prevent registration of a voice standard pattern that easily misrecognizes ambient sounds, and misrecognition caused by inappropriateness of the registered speech 11, that is, misrecognition caused by registration of sounds similar to ambient sounds. Can be reduced, and the recognition accuracy can be improved.

  Further, according to the fifth embodiment, since the registered voice change request 12 is output only when the user hears the reproduced sound 14 and judges that it is noise, the ambient sound similarity calculating unit 141 calculates the ambient sound similarity. At this time, when the user accidentally utters the same voice as the registered voice 11, the registered voice change request 12 is output, and there is an effect that the voice standard pattern is not erroneously changed.

Embodiment 6 FIG.
FIG. 13 is a block diagram showing the structure of a speech recognition apparatus according to Embodiment 6 of the present invention. The speech recognition apparatus includes a registration unit 100 and a verification unit 300. The registration unit 100 includes an audio standard pattern generation unit 115, an audio standard pattern storage unit 116, an ambient sound similarity calculation unit 141, an ambient sound reproduction determination unit 151, an ambient sound reproduction switch 152, an ambient sound reproduction unit 153, and a noise standard pattern generation unit. 161 and a noise standard pattern storage unit 162, and the collation unit 300 includes a voice similarity calculation unit 311, a noise similarity calculation unit 361, and a voice collation determination unit 362.

  In the registration means 100, the voice standard pattern generation means 115, the voice standard pattern storage means 116, the ambient sound similarity calculation means 141, the ambient sound reproduction determination means 151, the ambient sound reproduction switch 152, and the ambient sound reproduction means 153 are described in the embodiment. 5 has the same function as that shown in FIG. The noise standard pattern generation means 161 inputs the noise confirmation 15 by the user, inputs the ambient sound 13, generates a noise standard pattern, and stores it in the noise standard pattern storage means 162.

  Further, in the collating means 300, the voice similarity calculating means 311 inputs the recognition target voice 31 uttered by the user and the voice standard pattern stored in the voice standard pattern storage means 116, and calculates the voice similarity. The noise similarity calculation means 361 inputs the recognition target speech 31 uttered by the user and the noise standard pattern stored in the noise standard pattern storage means 162, and calculates the noise similarity. The voice collation determination unit 362 receives the voice similarity calculated by the voice similarity calculation unit 311 and the noise similarity calculated by the noise similarity calculation unit 361, and the voice similarity is greater than or equal to the voice similarity threshold 34. If the noise similarity is less than or equal to the noise similarity threshold 35, the recognition result 33 is output.

  In the sixth embodiment, the voice standard pattern generation unit 115, the ambient sound similarity calculation unit 141, the ambient sound reproduction determination unit 151, the ambient sound reproduction unit 153, the noise standard pattern generation unit 161, and the voice similarity calculation unit 311. The noise similarity calculation unit 361 and the voice collation determination unit 362 may be configured by hardware. However, a voice recognition program describing the processing contents of each unit is created, and the computer executes the voice recognition program. May be.

Next, the operation will be described.
FIG. 14 is a flowchart showing the flow of processing of the speech recognition apparatus according to Embodiment 6 of the present invention. First, at the time of voice registration, in step ST601, the voice standard pattern generation means 115 inputs the registered voice 11 uttered by the user. In step ST602, the voice standard pattern generation means 115 generates a voice standard pattern from the registered voice 11. And stored in the voice standard pattern storage means 116.

  In step ST <b> 603, the ambient sound similarity calculation unit 141 calculates the ambient sound similarity with respect to the ambient sound 13 using the speech standard pattern stored in the speech standard pattern storage unit 116.

  In step ST604, the ambient sound reproduction determination unit 151 reproduces and uses the ambient sound 13 by determining whether or not the ambient sound similarity calculated by the ambient sound similarity calculation unit 141 is equal to or greater than a predetermined threshold. Judge whether to let the person know. As a determination method, when the ambient sound similarity is equal to or higher than a predetermined threshold, it is determined to reproduce the ambient sound.

  If it is determined in step ST604 that ambient sound is to be reproduced, in step ST605, the ambient sound reproduction determining means 151 is connected to the ambient sound reproduction switch 152, and the ambient sound reproducing means 153 is input with the ambient sound 13. The reproduction sound 14 is output, and the user hears the reproduction sound 14.

  In step ST606, the user hears the reproduced sound 14 and determines that it is noise if it differs from the registered utterance. In step ST607, the user is notified of the determination result of noise confirmation 15 to the sound recognition pattern generation means. 161 receives the ambient sound 13 based on the noise confirmation 15, generates a noise standard pattern, and stores it in the noise standard pattern storage means 162. The noise standard pattern can be generated by using, for example, an HMM and learning the parameters of the HMM from the ambient sound 13. In this way, it is possible to determine whether noise has been misrecognized by registering noise that is easily misrecognized as a voice standard pattern as a noise standard pattern and calculating the noise similarity at the time of recognition. Further, since the noise standard pattern is generated only when the user hears the reproduced sound 14 and judges that it is noise, the user accidentally registers the registered voice when the ambient sound similarity calculating unit 141 calculates the ambient sound similarity. 11, the noise standard pattern is not generated by mistake using the user's voice.

  When the determination result by the ambient sound reproduction determination unit 151 does not need to be reproduced in step ST604, or when it is determined in step ST606 that the reproduced sound 14 that is the output of the ambient sound reproduction unit 153 is not noise, The registration process ends.

  At the time of speech recognition, the speech similarity calculation unit 311 inputs the recognition target speech 31 in step ST608, and in step ST609, the speech similarity calculation unit 311 uses the speech standard pattern stored in the speech standard pattern storage unit 116. Using this, the speech similarity to the recognition target speech 31 is calculated.

  In step ST610, the noise similarity calculation unit 361 inputs the recognition target speech 31, and calculates the noise similarity with respect to the recognition target speech 31 using the noise standard pattern stored in the noise standard pattern storage unit 162.

  In step ST611, the speech collation determining unit 362 receives the speech similarity calculated by the speech similarity calculating unit 311 and the noise similarity calculated by the noise similarity calculating unit 361, and the speech similarity is the speech similarity. It is determined whether or not the noise similarity is not less than the threshold 34 and the noise similarity threshold is not more than 35. In step ST611, when the voice similarity is equal to or higher than the voice similarity threshold 34 and the noise similarity is equal to or lower than the noise similarity threshold 35, in step ST612, the voice collation determination unit 362 stores the voice standard pattern storage unit 116. It is determined that the user has uttered voice corresponding to the stored voice standard pattern, and a recognition result 33 is output. On the other hand, if the speech similarity is smaller than the speech similarity threshold 34 or the noise similarity is larger than the noise similarity threshold 35 in step ST611, the speech matching determination unit 362 does not output the recognition result 33, and the step Return to ST608.

  Here, the case of only one utterance has been described, but at the time of voice registration, a voice standard pattern is generated for each of a plurality of different utterances, and at the time of voice recognition, a voice using a plurality of standard patterns for the recognition target voice 31 is used. It is also possible to calculate the similarity and output the utterance number indicating the standard pattern having the highest speech similarity with the speech similarity being the speech similarity threshold 34 or more and the noise similarity being the noise similarity threshold 35 or less as the recognition result. is there.

  As described above, according to the sixth embodiment, the noise standard pattern generation unit 161 generates and registers a noise standard pattern based on ambient sounds that are easily recognized as noise standard patterns and is determined as noise. The noise similarity calculation unit 361 calculates the noise similarity, and the voice collation determination unit 362 performs voice collation using the calculated noise similarity, thereby making it possible to determine whether or not the noise is erroneously recognized. The effect that the recognition accuracy can be improved is obtained.

  Further, according to the sixth embodiment, since the noise standard pattern is generated only when the user hears the reproduced sound 14 and judges that it is noise, the ambient sound similarity calculating unit 141 calculates the ambient sound similarity. If the user accidentally makes the same utterance as the registered voice 11, no noise standard pattern is erroneously generated using the user's voice, so that the recognition accuracy can be improved. It is done.

Embodiment 7 FIG.
FIG. 15 is a block diagram showing the configuration of a speech recognition apparatus according to Embodiment 7 of the present invention. The speech recognition apparatus includes a registration unit 100 and a verification unit 300. The registration unit 100 includes a noise standard pattern storage unit 171, a noise similarity calculation unit 172, a voice registration determination unit 173, a registered voice change request unit 113, a voice registration switch 114, a voice standard pattern generation unit 115, and a voice standard pattern storage unit 116. The collation unit 300 includes a speech similarity calculation unit 311 and a speech collation determination unit 312.

  In the registration unit 100, the noise standard pattern storage unit 171 stores a noise standard pattern prepared in advance. The noise similarity calculation means 172 receives the registered voice 11 and calculates the noise similarity for the registered voice 11 using the noise standard pattern stored in the noise standard pattern storage means 171. The voice registration determination unit 173 determines whether or not to register the registered voice 11 by determining whether or not the noise similarity calculated by the noise similarity calculation unit 172 is equal to or greater than a predetermined threshold. When the determination result by the voice registration determination unit 173 is not registrationable, the registration voice change request unit 113 outputs a registration voice change request 12 to the user. The voice registration determination unit 173 connects the voice registration switch 114 when the determination result can be registered, and the voice standard pattern generation unit 115 inputs the registered voice 11 and generates a voice standard pattern to generate a voice standard pattern storage unit. 116.

  In the matching unit 300, the functions of the voice similarity calculation unit 311 and the voice matching determination unit 312 are the same as those shown in FIG.

  In the seventh embodiment, the noise similarity calculation unit 172, the voice registration determination unit 173, the registered voice change request unit 113, the voice standard pattern generation unit 115, the voice similarity calculation unit 311, and the voice collation determination unit 312 are hardened. However, it is also possible to create a speech recognition program that describes the processing contents of each means and allow the computer to execute the speech recognition program.

Next, the operation will be described.
FIG. 16 is a flowchart showing the flow of processing of the speech recognition apparatus according to Embodiment 7 of the present invention. First, at the time of voice registration, in step ST701, the noise similarity calculation means 172 inputs the registered voice 11 uttered by the user, and in step ST702, the noise similarity calculation means 172 is stored in the noise standard pattern storage means 171. The noise similarity with respect to the registered voice 11 is calculated using the noise standard pattern. Here, the noise standard pattern is created so as to output a high degree of similarity when noise other than voice is input. For example, a mixed distribution type HMM in one state is used, and noise is generated as learning data to perform HMM learning. As for the calculation of the noise similarity, if the standard pattern is a mixed distribution type HMM with one state, the similarity is calculated by the method described in Chapter 5 of Non-Patent Document 2.

  In step 703, the voice registration determination unit 173 determines whether or not to register the registered voice 11 by determining whether or not the noise similarity calculated by the noise similarity calculation unit 172 is equal to or greater than a predetermined threshold. To do. As a determination method, when the noise similarity is equal to or greater than a predetermined threshold, it is determined that voice registration is impossible, and when the noise similarity is smaller than the threshold, it is determined that voice registration is possible.

  If the determination result by the voice registration determination unit 173 is not registered in step 703, the registration voice change request unit 113 outputs the registration voice change request 12 to the user in step ST704. When the noise similarity calculated by the noise similarity calculation means 172 is equal to or higher than the threshold value, the voice standard pattern generated by the registered voice 11 is highly likely to erroneously recognize noise. Therefore, the registration voice change required 12 is output. As a request from the speech recognition apparatus, for example, a message “Please say another word” is transmitted to the user by synthetic sound, and the process returns to step ST701 to repeat the registration process.

  On the other hand, if the determination result by the voice registration determination unit 173 can be registered in step 703, the voice registration determination unit 173 connects the voice registration switch 114 and the voice standard pattern generation unit 115 in step ST705. The registered voice 11 is input, a voice standard pattern is generated and stored in the voice standard pattern storage means 116.

  The processing of speech similarity calculation means 311 in steps ST706 and ST707 during speech recognition and the processing of speech collation determination means 312 in steps ST708 and ST709 are the same as those shown in FIG. 2 of the first embodiment.

  Here, the case of only one utterance has been described. However, at the time of voice registration, a voice standard pattern is generated for each of a plurality of different utterances. It is also possible to calculate the degree and output the utterance number indicating the standard pattern having the largest voice similarity with a voice similarity of 32 or more as a recognition result.

  As described above, according to the seventh embodiment, the noise similarity calculating unit 172 calculates the noise similarity for the registered speech 11 using the noise standard pattern, and the speech registration determining unit 173 calculates the calculated noise similarity. , It is possible to prevent registration of a voice standard pattern that is likely to misrecognize noise, and misrecognize the cause that the registered voice 11 is inappropriate. That is, it is possible to reduce misrecognition caused by registering a voice standard pattern similar to noise, and to improve the recognition accuracy.

Embodiment 8 FIG.
FIG. 17 is a block diagram showing the configuration of a speech recognition apparatus according to Embodiment 8 of the present invention. The speech recognition apparatus includes a registration unit 100 and a verification unit 300. The registration unit 100 includes a speech standard pattern generation unit 115, a speech standard pattern storage unit 116, a separate use state speech standard pattern generation unit 181 and a separate use state speech standard pattern storage unit 182, and the collation unit 300 calculates speech similarity. Means 311, another usage state voice similarity calculation means 381 and voice collation determination means 382 are provided.

  In the registration unit 100, the voice standard pattern generation unit 115 receives the registered voice 11, generates a voice standard pattern, and stores it in the voice standard pattern storage unit 116. The separate usage state voice standard pattern generation unit 181 receives the voice standard pattern stored in the voice standard pattern storage unit 116, generates another use state voice standard pattern, and generates another use state voice standard pattern storage unit 182. To store.

  In the matching unit 300, the speech similarity calculation unit 311 inputs the recognition target speech 11 uttered by the user and the speech standard pattern stored in the speech standard pattern storage unit 116, and calculates speech similarity. The different usage state voice similarity calculation means 381 inputs the recognition target voice 11 uttered by the user and the different usage state voice standard pattern storage means 182 and inputs the different usage state voice standard pattern. Calculate voice similarity. The voice collation determination unit 382 inputs the voice similarity calculated by the voice similarity calculation unit 311 and the different use state voice similarity calculated by the different use state voice similarity calculation unit 381, and uses the voice similarity or another use. If the state voice similarity is greater than or equal to the threshold 32, a recognition result 33 is output.

  In the eighth embodiment, the voice standard pattern generation means 115, the separate use state voice standard pattern generation means 181, the voice similarity calculation means 311, the separate use state voice similarity calculation means 381, and the voice collation determination means 382 are provided. Although it may be configured by hardware, a voice recognition program describing the processing contents of each means may be created, and the computer may execute the voice recognition program.

Next, the operation will be described.
FIG. 18 is a flowchart showing the flow of processing of the speech recognition apparatus according to Embodiment 8 of the present invention. First, at the time of voice registration, in step ST801, the voice registration pattern generation means 115 inputs the registered voice 11 uttered by the user. In step ST802, the voice registration pattern generation means 115 generates a voice standard pattern from the registered voice 11. And stored in the voice standard pattern storage means 116.

  In step ST803, the separate use state sound standard pattern generation unit 181 inputs the sound standard pattern stored in the sound standard pattern storage unit 116, generates another use state sound standard pattern, and uses the other use state sound standard pattern. The data is stored in the standard pattern storage means 182. Here, the separate usage state is a case where the recognition target speech 31 is input in a state different from the usage state in which the registered speech 11 is input.

  FIG. 19 is a diagram showing the appearance of a camera-equipped mobile phone. For example, in a device in which a voice recognition device is applied to a camera-equipped mobile phone as shown in FIG. 19 and the camera shutter is released by voice, the registered voice 11 generally faces the microphone hole facing the front of the mobile phone. However, in the case of a self-portrait at the time of shooting, the recognition target voice 31 is input from the opposite direction of the microphone hole directly facing the back of the mobile phone, and is in another use state.

Here, a method of generating another use state voice standard pattern will be described. The different use state voice standard pattern is generated in advance from the voice standard pattern by previously obtaining the difference between the input frequency characteristic when the registered voice 11 is input and the input frequency characteristic in the different use state. .
FIG. 20 is a diagram illustrating the frequency characteristics of voice input. 20A shows the frequency characteristics when the registered voice 11 is input, FIG. 20B shows the frequency characteristics of the input in a different use state, and FIG. 20C shows the frequency characteristics of FIGS. 20A and 20B. This is the difference in frequency characteristics. In the example of this figure, by subtracting the frequency characteristics of FIG. 20C from the voice standard pattern of FIG. 20A, the voice standard pattern for another use state of FIG. 20B is generated.

ここで、ｃ_t は音声標準パタンのフレーム時刻ｔのケプストラム、バーｃは登録音声１１が入力される周波数特性と別使用状態の入力の周波数特性の差のケプストラム、ｃ’_t は別使用状態用音声標準パタンのフレーム時刻ｔのケプストラムである。 If there is a cepstrum analyzed for each frame as a constituent of the voice standard pattern, a cepstrum of a voice standard pattern for another use state is generated by the following equation (2).

Here, c _t is a cepstrum at the frame time t of the voice standard pattern, bar c is a cepstrum of the difference between the frequency characteristic to which the registered voice 11 is inputted and the frequency characteristic of the input in a different use state, and c ′ _t is for another use state. It is a cepstrum at the frame time t of the voice standard pattern.

  During speech recognition, the speech similarity calculation means 311 inputs the recognition target speech 11 in step ST804 of FIG. 18, and the speech similarity calculation means 311 stores the speech standard stored in the speech standard pattern storage means 116 in step ST805. Using the pattern, the speech similarity to the recognition target speech 11 is calculated.

  In step ST806, the different usage state voice similarity calculation means 381 inputs the recognition target voice 31, and uses the different usage state voice standard pattern storage means 182 to recognize the recognition target voice. The different use state voice similarity to the voice 11 is calculated.

  In step ST807, the voice collation determination unit 382 receives the voice similarity calculated by the voice similarity calculation unit 311 and the different usage state voice similarity calculated by the different usage state voice similarity calculation unit 381, It is determined whether one or both of the similarity and the different use state audio similarity is equal to or greater than a predetermined threshold value 32. In step ST807, if one or both of the voice similarity and the different use state voice similarity is equal to or greater than the predetermined threshold 32, the voice collation determination unit 382 outputs the recognition result 33 in step ST808. On the other hand, when both the voice similarity and the different use state voice similarity are smaller than the threshold value 32 in step ST807, the voice collation determination unit 382 does not output the recognition result 33 and returns to step ST804.

  Here, the case of only one utterance has been described. However, at the time of voice registration, a voice standard pattern and a different usage state voice standard pattern are generated for a plurality of different utterances, and at the time of voice recognition, a plurality of recognition target voices 31 are generated. The voice similarity and the different use state voice similarity are calculated using the standard pattern and the different use state voice standard pattern, and the voice similarity or the different use state voice similarity is greater than or equal to the threshold 32, and the voice similarity or the other use It is also possible to output the utterance number indicating the standard pattern having the highest state voice similarity as the recognition result.

  As described above, according to the eighth embodiment, the separate usage state voice standard pattern generation unit 181 generates the separate usage state voice standard pattern that is different from the state in which the registered voice 11 is input. The different usage state voice similarity calculation means 381 calculates another usage state voice similarity for the recognition target voice 31 using the different usage state voice standard pattern, and the voice matching determination means 382 also considers the different usage state voice similarity. Then, by outputting the recognition result 33, there is an effect that the recognition accuracy can be improved by reducing the number of misrecognitions caused by the similarity being lowered due to the difference in the frequency characteristics depending on the different use states.

Embodiment 9 FIG.
FIG. 21 is a block diagram showing the structure of a speech recognition apparatus according to Embodiment 9 of the present invention. The speech recognition apparatus includes a registration unit 100 and a verification unit 300. The registration unit 100 includes a face standard pattern generation unit 191, a face standard pattern storage unit 192, a normal use state audio standard pattern generation unit 193, a normal use state audio standard pattern storage unit 194, and a separate use state audio standard pattern generation unit 195. And another use state voice standard pattern storage means 196. The collation means 300 includes a face similarity calculation means 391, a face determination means 392, a voice standard pattern selection means 393, a voice similarity calculation means 311 and a voice collation determination means 312. I have.

  In the registration unit 100, the face standard pattern generation unit 191 receives the registered face image 16 of the user, generates a face standard pattern, and stores it in the face standard pattern storage unit 192. The normal use state voice standard pattern generation unit 193 receives the normal use state registration voice 17 uttered by the user, generates a normal use state sound standard pattern, and stores it in the normal use state sound standard pattern storage unit 194. Store. The separate usage state voice standard pattern generation means 195 receives the different usage state registration voice 18 uttered by the user, generates another usage state voice standard pattern, and stores it in the separate usage state voice standard pattern storage means 196. Store.

  In the matching unit 300, the face similarity calculation unit 391 inputs the recognition target image 36 and the face standard pattern stored in the face standard pattern storage unit 192 and calculates the face similarity. The face determination unit 392 determines whether the registered user is shown by determining whether the face similarity calculated by the face similarity calculation unit 391 is greater than or equal to a predetermined threshold. The voice standard pattern selection unit 393 receives the determination result of the face determination unit 392, and is used for the normal use state voice standard pattern stored in the normal use state voice standard pattern storage unit 194, or for another use state voice standard pattern. The voice standard pattern for another use state stored in the storage unit 196 is selected as the voice standard pattern. The voice similarity calculation means 311 inputs the recognition target voice 31 uttered by the user and the voice standard pattern selected by the voice standard pattern selection means 393, and calculates the voice similarity. The voice collation determination unit 312 receives the voice similarity calculated by the voice similarity calculation unit 311, and outputs a recognition result 33 when the voice similarity is equal to or greater than the threshold 32.

  In the ninth embodiment, the face standard pattern generation means 191, the normal use state sound standard pattern generation means 193, the separate use state sound standard pattern generation means 195, the face similarity calculation means 391, the face determination means 392, The voice standard pattern selection unit 393, the voice similarity calculation unit 3111, and the voice collation determination unit 312 may be configured by hardware. However, a voice recognition program that describes the processing contents of each unit is created, and the computer performs the voice recognition. The program may be executed.

Next, the operation will be described.
FIG. 22 is a flowchart showing the flow of processing of the speech recognition apparatus according to Embodiment 9 of the present invention. Here, a case where a voice recognition device is mounted on the camera-equipped mobile phone shown in FIG. 19 will be described as an example. First, at the time of voice registration, the face standard pattern generation means 191 inputs the user's registered face image 16 in step ST901, and the face standard pattern generation means 191 generates a face standard pattern from the registered face image 16 in step ST902. And stored in the face standard pattern storage means 192. Here, the registered face image 16 is obtained by inputting a user's face using a camera. The face standard pattern can be generated by a method described in, for example, Japanese Patent Application Laid-Open No. 2000-99722, a human face recognition device and a human face recognition method.

  In step ST903, the normal use state voice standard pattern generation means 193 inputs the normal use state registered voice 17 uttered by the user. In step ST904, the normal use state sound standard pattern generation means 193 is for the normal use state. A normal use state sound standard pattern is generated from the registered sound 17 and stored in the normal use state sound standard pattern storage means 194. Here, for example, when the voice recognition device is mounted on a camera-equipped mobile phone as shown in FIG. 19, the normal use state registration voice 17 is uttered facing the surface (front surface) having the microphone hole of the mobile phone. Registered voice. Therefore, the utterance is when the user is not shown on the camera. At the time of voice recognition, the voice standard pattern for normal use is the voice standard pattern used when facing the surface (front surface) where the microphone hole of the mobile phone is located.

  In step ST905, the separate use state voice standard pattern generation means 195 inputs the separate use state registration voice 18 uttered by the user. A different voice condition standard pattern for use state is generated from the registered voice 18 and stored in the voice standard pattern storage means for different use condition 196. Here, the separate use state registration voice 18 is a registration voice when the positional relationship between the user and the microphone is different from the normal use state registration voice 17. For example, when a voice recognition device is mounted on a camera-equipped mobile phone as shown in FIG. Therefore, the utterance is in a state where the user is reflected in the camera. At the time of voice recognition, the voice standard pattern for another use state is the voice standard pattern used when facing directly to the surface (back surface) without the microphone hole of the mobile phone.

  During speech recognition, the face similarity calculation unit 391 inputs the recognition target image 36 in step ST907, and the face similarity calculation unit 391 uses the face standard pattern stored in the face standard pattern storage unit 192 in step ST908. The face similarity with respect to the recognition target image 36 is calculated. Here, the recognition target image 36 is an image input from the camera at the time of voice recognition. The face similarity is calculated, for example, by the method described in Japanese Patent Laid-Open No. 2000-99722.

  In step ST909, the face determination unit 392 determines whether the user is in the camera by determining whether the face similarity calculated by the face similarity calculation unit 391 is greater than or equal to a predetermined threshold. To do. In step ST909, if the face similarity is greater than or equal to a predetermined threshold, it is determined that the user is in the camera. On the other hand, if the face similarity is less than the threshold, the user is in the camera. Judge that it is not.

  If it is determined in step ST909 that the face similarity is smaller than a predetermined threshold and the user is not in the camera, the voice standard pattern selection unit 393 stores the voice standard pattern for normal use in step ST910. The normal use state voice standard pattern stored in the means 194 is selected as the voice standard pattern. On the other hand, if it is determined in step ST909 that the face similarity is equal to or greater than a predetermined threshold value and the user is in the camera, the audio standard pattern selection means 393 determines in step ST911 that the audio standard for different use states. The voice standard pattern for another use state stored in the pattern storage unit 196 is selected as the voice standard pattern.

  In step ST912, the speech similarity calculation unit 311 inputs the recognition target speech 31, and in step ST913, the speech similarity calculation unit 311 uses the speech standard pattern selected by the speech standard pattern selection unit 393 to recognize the target speech. The voice similarity to the voice 31 is calculated.

  In step ST914, the speech collation determining unit 312 determines whether the speech similarity calculated by the speech similarity calculating unit 311 is equal to or greater than a predetermined threshold 32. If the voice similarity is greater than or equal to the threshold value 32 in step ST914, in step ST915, the voice collation determination unit 312 stores the normal standard for use state voice standard stored in the normal use state voice standard pattern storage unit 194. It is determined that the user has uttered a voice corresponding to the voice standard pattern for different usage status stored in the pattern or the voice standard pattern storage status for different usage status 196, and the recognition result 33 is output. On the other hand, if the voice similarity is smaller than the threshold value 32 in step ST914, the voice collation determination unit 312 does not output the recognition result 33 and returns to step ST907.

  Here, the case of only one utterance has been described. However, at the time of voice registration, a normal use state voice standard pattern and a separate use state voice standard pattern are respectively generated for a plurality of different utterances. The voice similarity is calculated using a plurality of normal use state voice standard patterns or different use state voice standard patterns, and the voice similarity is equal to or greater than the threshold 32 and indicates the standard pattern having the highest voice similarity. It is also possible to output the utterance number as a recognition result.

  As described above, according to the ninth embodiment, the face similarity calculation unit 391 inputs a camera image at the time of recognition and calculates the face similarity, and the face determination unit 392 uses the face similarity to determine whether the user is in the camera. The voice standard pattern selection unit 393 determines whether the user is captured as a result of the determination by the face determination unit 392, or the voice standard pattern for the normal use state or the voice standard for another use state. The pattern is selected as the voice standard pattern, and the voice similarity calculation means 311 calculates the voice similarity with respect to the recognition target voice 31 using the selected voice standard pattern. Therefore, it is possible to reduce the erroneous recognition caused by the low voice similarity and to improve the recognition accuracy.

Embodiment 10 FIG.
FIG. 23 is a block diagram showing the structure of a speech recognition apparatus according to Embodiment 10 of the present invention. The speech recognition apparatus includes a registration unit 100 and a verification unit 300. The registration unit 100 includes a normal use state audio standard pattern generation unit 193, a normal use state audio standard pattern storage unit 194, another use state audio standard pattern generation unit 195, and another use state audio standard pattern storage unit 196. The collation unit 300 includes a rotation detection unit 401, a voice standard pattern selection unit 402, a voice similarity calculation unit 311, and a voice collation determination unit 312.

  In the registration unit 100, the normal use state voice standard pattern generation unit 193 receives the normal use state registration voice 17 uttered by the user, generates a normal use state sound standard pattern, and generates a normal use state sound standard. Stored in the pattern storage means 194. The separate usage state voice standard pattern generation means 195 receives the different usage state registration voice 18 uttered by the user, generates another usage state voice standard pattern, and stores it in the separate usage state voice standard pattern storage means 196. Store.

  Further, in the collating means 300, the rotation detecting means 401 inputs a signal from the acceleration sensor 37 installed in the camera-equipped mobile phone, and detects whether or not the voice recognition device has rotated. The voice standard pattern selection unit 402 inputs the result of whether or not the voice recognition device determined by the rotation detection unit 401 has rotated, and is used for the normal use state stored in the normal use state voice standard pattern storage unit 194. The voice standard pattern or the voice standard pattern for different usage status stored in the voice standard pattern storage means 196 for different usage status is selected as the voice standard pattern. The voice similarity calculation means 311 inputs the recognition target voice 31 uttered by the user and the voice standard pattern selected by the voice standard pattern selection means 402 and calculates the voice similarity. The voice collation determination unit 312 outputs a recognition result 33 when the voice similarity calculated by the voice similarity calculation unit 311 is equal to or greater than the threshold value 32.

  In the tenth embodiment, normal use state voice standard pattern generation means 193, separate use state voice standard pattern generation means 195, rotation detection means 401, voice standard pattern selection means 402, voice similarity calculation means 311 and The voice collation determination unit 312 may be configured by hardware, but a voice recognition program describing the processing contents of each unit may be created and the computer may execute the voice recognition program.

Next, the operation will be described.
FIG. 24 is a flowchart showing a process flow of the speech recognition apparatus according to the tenth embodiment of the present invention. Here, a case where a voice recognition device is mounted on the camera-equipped mobile phone shown in FIG. 19 will be described as an example. First, at the time of voice registration, in step ST1001, the normal use state voice standard pattern generation means 193 inputs the normal use state registration voice 17 uttered by the user, and in step ST1002, the normal use state voice standard pattern generation means. In step 193, a normal use state sound standard pattern is generated by the normal use state registration sound 17 and stored in the normal use state sound standard pattern storage unit 194. For example, when the voice recognition device is mounted on a mobile phone as shown in FIG. 19, the normal use state registration voice 17 is a registered voice uttered directly on the surface (front surface) having the microphone hole of the mobile phone. is there. Therefore, at the time of voice recognition, the voice standard pattern for normal use is the voice standard pattern used when facing the surface (front surface) where the microphone hole of the mobile phone is located.

  In step ST1003, the separate use state voice standard pattern generation unit 195 inputs the separate use state use registered voice 18 uttered by the user. In step ST1004, the separate use state sound standard pattern generation unit 195 outputs another use state. The voice standard pattern for another use state is generated from the registered voice 18 and stored in the voice standard pattern storage unit 196 for another use state. Here, the separate usage state registration voice 18 is a registration voice when the positional relationship between the user and the microphone is different from the normal usage state registration voice 17. For example, when a voice recognition device is mounted on a camera-equipped mobile phone as shown in FIG. 19, it is a registered voice uttered directly on the surface (back surface) of the mobile phone that does not have a microphone hole. Therefore, at the time of voice recognition, the voice standard pattern for another use state is the voice standard pattern used when facing the surface (back side) without the microphone hole of the mobile phone.

  At the time of voice recognition, in step ST1005, the rotation detection unit 401 inputs an angle change amount from the acceleration sensor 37, and in step ST1006, the rotation detection unit 401 detects whether or not the voice recognition device has rotated. The determination as to whether or not the rotation has been made is to determine whether or not the voice recognition device has rotated a predetermined angle or more from the initial state where the voice recognition device has been activated. For example, when a voice recognition device is mounted on a camera-equipped mobile phone as shown in FIG. 19 and the voice recognition device is activated in a state of facing the surface, the rotation angle is 90 degrees or more. If so, it is determined that it has rotated because it faces the back surface. The angle change amount can be detected by the acceleration sensor 37 by, for example, a method described in Japanese Patent No. 3076124.

  If it is determined in step ST1006 that it does not rotate, in step ST1007, the voice standard pattern selection unit 402 uses the normal use state voice standard pattern stored in the normal use state voice standard pattern storage unit 194 as a voice standard. Select as a pattern. On the other hand, if it is determined in step ST1006 that the rotation has been made, in step ST1008, the audio standard pattern selection unit 402 determines that the audio standard pattern for different usage status stored in the voice standard pattern storage unit 196 for different usage status. Is selected as the voice standard pattern.

  In step ST1009, the speech similarity calculation unit 311 inputs the recognition target speech 31, and in step ST1010, the speech similarity calculation unit 311 recognizes using the speech standard pattern selected by the speech standard pattern selection unit 402. The voice similarity with respect to the target voice 31 is calculated.

  In step ST1011, the voice collation determining unit 312 determines whether or not the voice similarity calculated by the voice similarity calculating unit 311 is equal to or greater than a predetermined threshold 32. If it is determined in step ST1011 that the voice similarity is greater than or equal to the threshold value 32, in step ST1012, the voice collation determining unit 312 is for the normal use state stored in the normal use state voice standard pattern storage unit 194. It is determined that the user has uttered the voice corresponding to the voice standard pattern or the voice standard pattern for different usage status stored in the voice standard pattern storage unit 196 for different usage status, and the recognition result 33 is output. On the other hand, when it is determined in step ST1011 that the voice similarity is smaller than the threshold value 32, the voice collation determination unit 312 does not output the recognition result 33 and returns to step ST1005.

  Here, the case of only one utterance has been described. However, at the time of voice registration, a normal use state voice standard pattern and a separate use state voice standard pattern are respectively generated for a plurality of different utterances. The voice similarity is calculated using a plurality of normal use state sound standard patterns or a plurality of separate use state sound standard patterns, and the standard pattern having the highest sound similarity with a sound similarity of 32 or more is calculated. It is also possible to output the indicated utterance number as a recognition result.

  As described above, according to the tenth embodiment, the speech standard pattern selection unit 402 is based on the presence or absence of rotation of the speech recognition device detected by the rotation detection unit 401, or the normal use state speech standard pattern or another use. The state voice standard pattern is selected as the voice standard pattern, and the voice similarity calculation unit 311 calculates the voice similarity with respect to the recognition target voice 31 using the selected voice standard pattern. Due to the difference in frequency characteristics, it is possible to reduce misrecognition caused by a decrease in speech similarity and to improve the recognition accuracy.

Embodiment 11 FIG.
FIG. 25 is a block diagram showing the structure of a speech recognition apparatus according to Embodiment 11 of the present invention. The speech recognition apparatus includes a registration unit 100 and a verification unit 300. The registration unit 100 includes a face standard pattern generation unit 191, a face standard pattern storage unit 192, a voice standard pattern generation unit 115, and a voice standard pattern storage unit 116. The collation unit 300 includes a face similarity calculation unit 391, a face determination unit 392, A microphone directivity setting unit 411, a voice similarity calculation unit 311, and a voice collation determination unit 312 are provided.

  In the registration unit 100, the face standard pattern generation unit 191 receives the registered face image 16 of the user, generates a face standard pattern, and stores it in the face standard pattern storage unit 192. The voice standard pattern generation means 115 receives the registered voice 11 uttered by the user, generates a voice standard pattern, and stores it in the voice standard pattern storage means 116.

  In the matching unit 300, the face similarity calculation unit 391 inputs the recognition target image 36 and the face standard pattern stored in the face standard pattern storage unit 192 and calculates the face similarity. The face determination unit 392 determines whether or not the user is in the camera by determining whether or not the face similarity calculated by the face similarity calculation unit 391 is greater than or equal to a predetermined threshold. The microphone directivity setting means 411 inputs the determination result by the face determination means 392, sets the directivity of the microphone, and outputs a setting signal 38 for driving the microphone so that the set directivity is obtained. The voice similarity calculation means 311 inputs the recognition target voice 31 uttered by the user through a microphone set with directivity, and inputs the voice standard pattern stored in the voice standard pattern storage means 116 to input voice. Calculate similarity. The voice collation determination unit 312 outputs the recognition result 33 when the voice similarity calculated by the voice similarity calculation unit 311 is equal to or greater than the threshold value 32.

  In the eleventh embodiment, face standard pattern generation means 191, voice standard pattern generation means 115, face similarity calculation means 391, face determination means 392, microphone directivity setting means 411, sound similarity calculation means 311 and sound. Although the collation determination unit 312 may be configured by hardware, a voice recognition program describing the processing contents of each unit may be created and the computer may execute the voice recognition program.

Next, the operation will be described.
FIG. 26 is a flowchart showing the flow of processing of the speech recognition apparatus according to Embodiment 11 of the present invention. Here, a case where a voice recognition device is mounted on the camera-equipped mobile phone shown in FIG. 19 will be described as an example. First, at the time of voice registration, the face standard pattern generation means 191 inputs the user's registered face image 16 in step ST1101, and the face standard pattern generation means 191 generates a face standard pattern from the registered face image 16 in step ST1102. And stored in the face standard pattern storage means 192. Here, the registered face image 16 is obtained by inputting a user's face using a camera.

  In step ST1103, the voice standard pattern generation means 115 inputs the registered voice 11 uttered by the user. In step ST1104, the voice standard pattern generation means 115 generates a voice standard pattern from the registered voice 11 and stores it in the voice standard pattern storage means 116.

  At the time of speech recognition, the face similarity calculation unit 391 inputs the recognition target image 36 in step ST1105, and in step ST1106, the face similarity calculation unit 391 uses the face standard pattern stored in the face standard pattern storage unit 192. The face similarity with respect to the recognition target image 36 is calculated.

  In step ST1107, the face determination unit 392 determines whether the user is in the camera by determining whether the face similarity calculated by the face similarity calculation unit 391 is greater than or equal to a predetermined threshold. To do. Here, if the face similarity is equal to or greater than a predetermined threshold, it is determined that the user is in the camera. On the other hand, if the face similarity is smaller than the threshold, it is determined that the user is not in the camera. .

  If it is determined in step ST1107 that the face similarity is equal to or greater than the threshold and the user is in the camera, in step ST1108, the microphone directivity setting unit 411 is set in the same direction as the camera, that is, on the back surface. A microphone directivity is set, and a setting signal 38 that outputs the set directivity is output. On the other hand, if it is determined in step ST1107 that the face similarity is smaller than the threshold value and the user is not in the camera, in step ST1109, the microphone directivity setting unit 411 is in the opposite direction to the camera, that is, A microphone directivity is set on the surface, and a setting signal 38 that outputs the set directivity is output.

  Here, the microphone directivity is a high sensitivity to sound in a certain direction. Therefore, if the recognition target speech 31 is input from a direction with strong sensitivity, power is input larger than the surrounding noise, and a high recognition rate is obtained because the signal-to-noise ratio is large. For example, in the case of a camera-equipped mobile phone as shown in FIG. 19, when the determination result by the face determination unit 392 determines that the user is captured, the microphone is directed in the direction in which the camera is facing, that is, the back surface. Set gender. On the other hand, when the determination result by the face determination unit 392 determines that the user is not captured, the microphone directivity is set in the direction in which the camera is not facing, that is, on the surface.

  In step ST1110, the speech similarity calculation unit 311 inputs the recognition target speech 31 via the microphone in which directivity is set. In ST1111, the speech similarity calculation unit 311 is stored in the speech standard pattern storage unit 116. The voice similarity with respect to the recognition target voice 31 is calculated using the existing voice standard pattern.

  In step ST1112, the speech collation determining unit 312 determines whether or not the speech similarity calculated by the speech similarity calculating unit 311 is equal to or greater than a predetermined threshold value 32. If it is determined in step ST1112 that the voice similarity is greater than or equal to the threshold value 32, in step ST1113, the voice collation determination unit 312 selects a voice corresponding to the voice standard pattern stored in the voice standard pattern storage unit 116. It is determined that the user uttered, and the recognition result 33 is output. On the other hand, when it is determined in step ST1112 that the voice similarity is smaller than the threshold value 32, the voice collation determination unit 312 does not output the recognition result 33 and returns to step ST1105.

  Here, the case of only one utterance has been described, but at the time of voice registration, a voice standard pattern is generated for each of a plurality of different utterances, and at the time of voice recognition, a plurality of voice standard patterns are used for the recognition target voice 31. It is also possible to calculate the similarity and output the utterance number indicating the standard pattern having the highest voice similarity with the voice similarity of 32 or more as the recognition result.

  As described above, according to the eleventh embodiment, the face similarity calculation unit 391 inputs the recognition target image 36 and calculates the face similarity, and the face determination unit 392 uses the face similarity to determine whether the user is in the camera. The microphone directivity setting means 411 sets the directivity of the microphone depending on whether or not the user is photographed, and the sound similarity calculation means 311 passes through the microphone with the directivity set. By inputting the recognition target speech 31, it is possible to reduce misrecognition caused by a decrease in speech similarity due to a decrease in signal-to-noise ratio and a difference in frequency characteristics caused by different usage states. The effect that accuracy can be improved is obtained.

Embodiment 12 FIG.
FIG. 27 is a block diagram showing the structure of a speech recognition apparatus according to Embodiment 12 of the present invention. This speech recognition apparatus includes a registration unit 100 and a verification unit 300, and is different from the speech recognition apparatus shown in FIG. 1 of the first embodiment in that the verification unit 300 includes a delay processing unit 421. Other configurations are the same. Here, the delay processing unit 421 receives the recognition result 33 output from the voice collation determination unit 312 and outputs a processing start signal 39 after a predetermined time.

  In the twelfth embodiment, the voice change calculation means 111, the voice registration determination means 112, the registered voice change request means 113, the voice standard pattern generation means 115, the voice similarity calculation means 311, the voice collation determination means 312 and the delay. The processing unit 421 may be configured by hardware, but a voice recognition program describing the processing contents of each unit may be created and the computer may execute the voice recognition program.

Next, the operation will be described.
FIG. 28 is a flowchart showing the flow of processing of the speech recognition apparatus according to Embodiment 12 of the present invention. Here, a case where a voice recognition device is mounted on the camera-equipped mobile phone shown in FIG. 19 will be described as an example. The processing from step ST1201 to step ST1205 at the time of voice registration and the processing from step ST1206 to step ST1209 at the time of voice recognition are the same as the processing from step ST101 to step ST109 shown in FIG. 2 of the first embodiment. .

  In step ST1210, the delay processing unit 421 inputs the recognition result 33 by the voice collation determination unit 312 and outputs a processing start signal 39 after a predetermined time. By such processing, it is possible to prevent the transition to the next processing before the utterance is finished. Therefore, it is possible to prevent the face being uttered from being captured when photographing the subject with a device that releases the shutter of the camera by voice recognition.

  Here, the case of only one utterance has been described. However, at the time of voice registration, a voice standard pattern is generated for each of a plurality of different utterances, and at the time of voice recognition, a plurality of standard patterns are used for the recognition target voice 31. It is also possible to calculate the degree and output the utterance number indicating the standard pattern having the highest speech similarity as the recognition result.

  As described above, according to the twelfth embodiment, the same effect as in the first embodiment can be obtained, and the delay processing unit 421 receives the recognition result 33 by the voice collation determining unit 312 and receives a predetermined time. By outputting the processing start signal 39 after a while, there is an effect that it is possible to prevent a case where a trouble occurs if the process proceeds to the next processing during the utterance.

  In the twelfth embodiment, the delay processing means 421 is added to the configuration of the first embodiment, but the same applies even if the delay processing means 421 is added to the configurations of the second to eleventh embodiments. The effect is obtained.

It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 1 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 1 of this invention. It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 2 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 2 of this invention. It is a figure explaining the speech recognition of the syllable unit in the speech recognition apparatus by Embodiment 2 of this invention. It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 3 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 3 of this invention. It is a figure explaining the speech recognition of the syllable unit in the speech recognition apparatus by Embodiment 3 of this invention. It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 4 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 4 of this invention. It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 5 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 5 of this invention. It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 6 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 6 of this invention. It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 7 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 7 of this invention. It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 8 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 8 of this invention. It is a figure which shows the external appearance of the mobile phone with a camera. It is a figure explaining the frequency characteristic of a voice input. It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 9 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 9 of this invention. It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 10 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 10 of this invention. It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 11 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 11 of this invention. It is a block diagram which shows the structure of the speech recognition apparatus by Embodiment 12 of this invention. It is a flowchart which shows the flow of a process of the speech recognition apparatus by Embodiment 12 of this invention.

Explanation of symbols

  11 Registration voice, 12 Registration voice change request, 13 Ambient sound, 14 Playback sound, 15 Noise confirmation, 16 Registration face image, 17 Normal use state registration voice, 18 Separate use state registration voice, 31 Recognition target voice, 32 Fight Value, 33 recognition result, 34 voice similarity threshold, 35 noise similarity threshold, 36 recognition target image, 37 acceleration sensor, 38 setting signal, 39 processing start signal, 100 registration means, 111 voice change degree calculation means, 112 Voice registration determination means, 113 registration voice change request means, 114 voice registration switch, 115 voice standard pattern generation means, 116 voice standard pattern storage means, 121 syllable number extraction means, 122 voice registration judgment means, 131 vowel likelihood calculation means, 132 voice registration determination means, 141 ambient sound similarity calculation means, 142 voice registration determination means, 151 Ambient sound reproduction determination means, 152 Ambient sound reproduction switch, 153 Ambient sound reproduction means, 161 Noise standard pattern generation means, 162 Noise standard pattern storage means, 171 Noise standard pattern storage means, 172 Noise similarity calculation means, 173 Voice registration determination Means, 181 Voice standard pattern generation means for different use states, 182 Voice standard pattern storage means for different use states, 191 Face standard pattern generation means, 192 Face standard pattern storage means, 193 Voice standard pattern generation means for normal use states, 194 Normal use state speech standard pattern storage means, 195 Separate use state speech standard pattern generation means, 196 Separate use state speech standard pattern storage means, 300 collation means, 311 Speech similarity calculation means, 312 Speech collation determination means, 361 Noise similarity calculation means, 362 voice collation determination means, 38 Separate use state voice similarity calculation means, 382 voice collation determination means, 391 face similarity calculation means, 392 face determination means, 393 voice standard pattern selection means, 401 rotation detection means, 402 voice standard pattern selection means, 411 microphone directivity Setting means, 421 Delay processing means.

Claims (2)

A voice change degree calculating means for inputting a registered voice of two or more syllables and calculating a voice change degree over the whole voice section ;
A voice registration determination means for comparing the voice change degree calculated by the voice change degree calculation means with an average value of the voice change degrees of words of a predetermined syllable and determining whether or not to register the input registered voice;
A registration voice change requesting means for outputting a registration voice change request when the judgment result by the voice registration judgment means is not registerable;
A speech recognition apparatus comprising speech standard pattern generation means for generating a speech standard pattern from the input registered speech when the determination result by the speech registration determination means can be registered. 2. The speech recognition apparatus according to claim 1, further comprising delay processing means for inputting a recognition result at the time of speech recognition and outputting an operation start signal for performing the next processing after elapse of a predetermined time .

Images