JPH05323990A - Talker recognizing method - Google Patents

Abstract

PURPOSE:To prevent a person from impersonating another person by reproducing the speech of the person recorded as that of another person by a vocalization content depending method. CONSTITUTION:Audio data for learning and a KANA (Japanese syllabary) representing the system of a vocalization content or phonetic symbol for each talker to be registered are inputted to a talker adaptation part 1 with teacher, and a phoneme/syllable model from a phoneme/syllable model accumulation part 2 for unspecified talker is adapted to the talker to be registered by using them, and it is accumulated in an accumulation part 3 with the ID of the talker. Talker recognition is performed by showing by generating a new sentence/work at every recognition from a sentence/work generating part 4 for recognition, vocalizing the sentence/work, and converting the speech of the sentence/work to a feature parameter system by a feature parameter extraction part 5. The talker recognition can be performed by connecting an adapted phoneme/syllable model in the accumulation part 3 according to the sentence/work generated from the generating part 4 by a speech model generating part 6, and calculating the similarity of input speech with the feature parameter system by a calculation part 7 based on the result of the calculation.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is used to recognize who a visitor is from a voice of an intercom, and to identify that the visitor is the same person who presents a personal identification number by input voice. For example, the input speech is converted into an expression format using a characteristic parameter, and the similarity between the input speech in the expression format and the feature of the voice of the speaker in the expression format previously generated corresponding to the speaker. The present invention relates to a speaker recognition method for recognizing a speaker who uttered an input voice, and a speaker recognition method for effectively normalizing the similarity when determining whether or not a designated sentence or word is correctly uttered. ..

[0002]

2. Description of the Related Art Conventional speaker recognition methods are classified into a voicing content limited type method in which voicing content is limited in advance and a voicing content independent type method in which a voice having any utterance content can be used. You can In the method of limited utterance content, usually, a sentence or a plurality of words to be uttered is determined in advance, and the voice uttered by the speaker is converted into a time series of characteristic parameters (eg, cepstrum, pitch, etc.), and each registered utterance is converted. A method is often used in which the degree of similarity of a person's same sentence or word to a time series of characteristic parameters is determined, and the value is used to determine who is the speaker.

According to this method, the similarity of the characteristic parameter time series of the voice of the same sentence or word is used.
It is relatively easy to recognize a speaker with high accuracy, but if you record the voice of a registered speaker, you can easily impersonate the registered speaker by playing the recorded voice. There was a big drawback. On the other hand, in the utterance content independent method, usually, the distribution of the characteristic parameters included in the voice such as the sentence uttered by the speaker is obtained using a method such as vector quantization, and the distribution of the characteristic parameters of the registered speakers is performed. A method of determining based on the degree of similarity is often used.

According to this method, the sentence or word uttered for speaker recognition can be changed each time, but any utterance content can be accepted. It was not possible to avoid the drawback of being able to impersonate another person.

[0005]

According to the invention of claim 1, a model of a phoneme or a syllable adapted to the voice of each speaker to be recognized is created and registered, and those models are connected and generated. The similarity between the model of the voice of the sentence or the word and the input voice converted into the expression form using the characteristic parameter is obtained, and the speaker who uttered the input voice is recognized.

According to the third or fourth aspect of the present invention, a model of a phoneme or a syllable group adapted to the voice of each speaker to be recognized is created and registered, and a sentence or word generated by connecting these models is generated. The similarity between the model of the voice and the input voice converted into the expression format using the characteristic parameter is obtained, and the speaker who uttered the input voice is recognized. According to the invention of claim 3, the model of the phoneme or syllable group approximates the distribution of the characteristic parameters in the learning voice to a plurality of Gaussian distributions, and weights the Gaussian distributions to add the characteristic parameters A phoneme or syllable group model with an approximate distribution is used. According to the invention of claim 4, a phoneme or syllable group model in which the distribution of the characteristic parameters in the learning speech is approximated by a codebook and the distribution of the characteristic parameters of each phoneme or syllable group is approximated by weighted addition of these codebook elements. And

According to the invention of claim 6, the input voice for recognizing the speaker is converted into an expression form using the characteristic parameter, and the input voice in the expression form and the utterance content, that is, the words in the expression form. The characteristics of each speaker's voice that does not depend on each speaker and the respective degrees of similarity with the words to be uttered are obtained, and the speakers that have uttered the input speech are recognized by combining these degrees of similarity.

According to the invention of claim 8, a voice model independent of the utterance content is created from the voice data for learning of the speaker, and between this model and the input voice converted into the expression format using the characteristic parameters. , And subtracting this from the similarity between the sentence or word speech model in which the phoneme or syllable (group) model is connected and the input speech converted into the expression format using the characteristic parameters, Normalize the variation in the similarity value due to the difference in the text or word content, recording time, transmission system, microphone and other test conditions.

A speech model independent of the utterance content of each speaker is
The invention of claim 1 is the same as the model of each phoneme or syllable, that is, the voice is represented by a time series of characteristic parameters, or the model is represented by a hidden Markov model. According to the invention of claim 3, the distribution of the characteristic parameters in the learning voice is approximated by a plurality of Gaussian distributions. According to the invention of claim 4, the distribution of the characteristic parameters in the learning voice is approximated by a codebook. According to the invention of claim 6, the distribution of the characteristic parameters in the learning speech is approximated by a codebook or a plurality of Gaussian distributions.

In any of the above-mentioned inventions, for each recognition, the speaker designates a sentence or word to be uttered, and the input voice is recognized.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, a detailed description will be given with reference to the drawings. FIG. 1 shows an embodiment of the invention of claims 1 and 2. First, in the step of registering a speaker, learning voice data is input to the speaker-accommodating speaker adaptation unit 1 for each speaker to be registered, and a phoneme or syllable model adapted to the speaker is created. In other words, the utterance content of the learning voice data is decided in advance,
The utterance content represented by a kana or phonetic symbol sequence is input to the teacher-equipped speaker adaptation unit 1 together with the voice. The phoneme / syllable model storage unit 2 for unspecified speakers stores models of each phoneme or syllable created in advance from the voices of many speakers. For each phoneme or syllable model, a voice is represented by a time series of characteristic parameters,
Alternatively, a hidden Markov model is used. As a method of adapting these phoneme / syllable models to a speaker to be registered, various methods that have already been established, for example, the literature “Jean-Luc Gauvain and Chin-Hu” are used.
i Lee: ″ Improved Acoustic Modeling With Bayesian
Learning, Proc. IEEE ICASSP 92, 1922 ”and the like can be used. The phoneme or syllable model adapted to the speaker is registered in the adapted phoneme / syllable model storage unit 3.

To identify a speaker, an identification number (personal identification number, hereinafter referred to as a speaker ID) peculiar to the speaker is input with a key or the like, and the ID is used as an adaptive phoneme of the speaker. It is stored in the storage unit 3 together with the syllable model. next,
At the stage of recognizing the speaker, a sentence or word to be uttered is determined in advance, or a new sentence or word is generated by the recognition sentence / word generation unit 4 and the new sentence or word is shown to the speaker. The sentence or word is uttered, and the voice is input to the characteristic parameter extraction unit 5. The characteristic parameter extraction unit 5 converts the input voice into an expression format using characteristic parameters such as cepstrum and pitch. On the other hand, in the sentence / word voice model generation unit 6,
According to a sentence or word generated by the recognition sentence / word generation unit 4, or a predetermined sentence or word, the phoneme or syllable model registered in the adapted phoneme / syllable model storage unit 3 is connected, Generate a phonetic model of the sentence or word. At that time, in the case of speaker verification, that is, when it is determined whether or not the voice is the voice of the person,
Ask the speaker to enter his or her ID with a key, etc.
The phoneme or syllable model corresponding to the registered speaker corresponding to D is input to the sentence / word speech model generation unit 6. In the case of speaker identification, that is, to determine who has a voice registered in advance, phoneme or syllable models corresponding to all registered speakers are sequentially input to the sentence / word voice model generation unit 6.

The time series of characteristic parameters obtained by the characteristic parameter extraction unit 5 and the speech model of the sentence or word generated by the sentence / word speech model generation unit 6 are input to the similarity calculation unit 7, The degree of similarity between the two is calculated. As a concrete method of this, for example, the document “Seiji Nakagawa:
"Speech Recognition by Stochastic Model", Section 3.1.3, pp. 40-
50, Institute of Electronics, Information and Communication Engineers, 1988 ”, etc. can be used. The calculated similarity value is temporarily stored in the similarity storage unit 8 and then the speaker recognition determination unit 9
Sent to. In the speaker recognition determination unit 9, the speaker ID is used in the case of speaker verification, and the threshold value storage unit 10 reads a threshold value indicating a range of variation in the degree of similarity that can be regarded as the voice of the person, The speech model of a sentence or word connected to the phoneme or syllable model of the person calculated as described above is compared with the similarity between the recognition sentence or the input voice of the word, and the similarity value is read out. If it is larger than the threshold, it is determined to be the voice of the person, and if smaller than the threshold, it is determined to be the voice of another person. In the case of speaker identification, the similarity between the input speech and the speech model of a sentence or word in which phoneme or syllable models adapted to all the registered speakers are connected is compared to obtain the highest similarity. A speaker is selected, and it is determined that the speaker uttered.

Next, an embodiment of the invention of claim 3 will be described with reference to FIG. First, at the speaker registration stage, learning voice data is input to the characteristic parameter extraction unit 11 for each speaker to be registered. The utterance content of the learning voice data is determined in advance. The characteristic parameter extraction unit 11 converts the input voice into an expression format using characteristic parameters such as cepstrum and pitch. Next, a plurality of Gaussian distributions are used as a method of expressing the features of the speaker's voice. That is, the learning speech data converted into the time series of the characteristic parameters is input to the mixed Gaussian distribution creating unit 12, and the distribution of the characteristic parameters included in the learning speech data is a combination of a plurality of Gaussian distributions (hereinafter referred to as " It is expressed as a “mixed Gaussian distribution”).

Next, the learning voice data, the text in which the utterance content thereof is represented by a kana or phonetic symbol sequence, and the Gaussian mixture distribution are input to the phoneme group model creating unit 13.
The phoneme group model creation unit 13 creates a phoneme group model in which the distribution of the feature parameters is expressed by weighted addition of elements of the mixed Gaussian distribution for each phoneme group in which several phonemes having similar feature parameter values are collected. Then, when these phoneme group models are temporally connected according to the text of the learning voice data, the weight of each element is determined so that the time series of the characteristic parameters of the learning voice data can be expressed most accurately. .. This method is described, for example, in the document "Yasuhiro Minami, Tatsuo Matsuoka, Kiyohiro Kano:" Evaluation of connected learning HMM using unspecified speaker continuous speech database ", Institute of Electronics, Information and Communication Engineers, Speech study group material, SP91-113, 1992". It is possible to use a known method or the like. The phoneme group model thus created estimates the speaker ID allowable range indicating the speaker who uttered the learning voice, and uses that value as the threshold value for similarity, together with the speaker ID, and accumulates the threshold value. Store in part 8.

Next, in the step of recognizing the speaker, a sentence or word to be uttered is determined in advance, or a new sentence or word is generated by the recognition sentence / word generation unit 4 and the new sentence or word is given to the speaker. , The speaker utters the sentence or word, and inputs the voice to the characteristic parameter extraction unit 5. The characteristic parameter extraction unit 5 converts the input voice into the same expression format as the characteristic parameter extraction unit 11. At the same time, in the sentence / word speech model generation unit 6, according to the sentence or word generated by the recognition sentence / word generation unit 4, or according to a predetermined sentence or word, the phoneme group model storage unit 3 stores the same. The phoneme group model is connected to generate a speech model of the sentence or word. At that time, in the case of speaker verification, that is, in the case of determining whether or not the voice is the voice of the person in question, the speaker is asked to input his own ID with a key or the like, and the phoneme group adapted to the registered speaker corresponding to the ID The model is input to the sentence / word voice model generation unit 11. In the case of speaker identification, that is, in determining who the voice is registered in advance, phoneme or syllable models corresponding to all registered speakers are sequentially input to the sentence / word voice model generation unit 11.

After that, the time series of the characteristic parameters obtained by the characteristic parameter extraction unit 5 and the voice model of the sentence or word generated by the sentence / word voice model generation unit 6 are input to the similarity calculation unit 7, The degree of similarity is calculated, and the speaker recognition determining section 9 identifies the speaker or recognizes the speaker in the same manner as in FIG. FIG. 3 shows an embodiment of the invention of claim 4.
In the embodiment of FIG. 2, the distribution of the characteristic parameters included in the learning voice data is expressed by the Gaussian mixture, but in the embodiment of FIG. 3, the time series of the characteristic parameters of the learning voice data is stored in the codebook creating unit 14. The distribution of the characteristic parameters included in the input speech data is represented by a vector quantization codebook. This codebook is input to the phoneme group model creating unit 13 instead of the Gaussian mixture distribution, and a phoneme model expressing the distribution of feature parameters is created by weighted addition of the elements of the codebook instead of the Gaussian mixture distribution. Others are Figure 2
It is similar to the case of. As a method for creating a mixed Gaussian distribution and a codebook, for example, a document “Tomoko Matsui, Sadahiro Furui:“ VQ, comparative study of text-independent speaker recognition methods using discrete / continuous HMM ”, and the Institute of Electronics, Information and Communication Engineers Speech The method described in "Study Group Material, SP91-89, 1991" can be used.

In the embodiments shown in FIGS. 2 and 3, a syllable group model may be used instead of the phoneme group model. Next, an embodiment of the invention of claim 6 will be described with reference to FIG. At the speaker registration stage, for each speaker to be registered, learning voice data is input to the characteristic parameter extraction unit 11 and converted into a time series of characteristic parameters (eg, cepstrum, pitch, etc.), and then the speakers are registered. Input to the feature calculator 15 and independent of the speaking content of the speaker (not dependent on the speaking content)
Extract voice features. As a method of extracting a feature of a voice independent of the utterance content, for example, a document “Tomoko Matsui, Sadahiro Furui:
"Comparative study of text-independent speaker recognition methods using VQ and discrete / continuous HMM", The Institute of Electronics, Information and Communication Engineers, Speech Study Group materials,
SP91-89, 1991 ", that is, the distribution of feature parameters is represented by a vector quantization codebook or a plurality of Gaussian distributions.
The extracted feature of the voice is stored in the speaker feature storage unit 16 together with the ID of the speaker.

On the other hand, the voice data of a large number of speakers is input to the characteristic parameter extraction unit 17 and converted into a time series of characteristic parameters, and then input to the phoneme / syllable model generation unit 18 to be shared by a large number of speakers. , I.e., create a speaker-independent phoneme / syllable model. This method is described in the literature, "Yasuhiro Minami, Tatsuo Matsuoka, Yasuhiro Kano:" Evaluation of connected learning HMM using unspecified speaker continuous speech database ", Institute of Electronics, Information and Communication Engineers, Speech study group material, SP91-113, 1992". The method described can be used. That is, the utterance content of the learning voice data is determined in advance, and the utterance content represented by a kana or phonetic symbol sequence is input to the phoneme / syllable model creating unit 18 together with the voice. As a model of each phoneme or syllable, a speech represented by a time series of characteristic parameters, or a representation of it by a hidden Markov model is used. Created phonemes /
The syllable model is stored in the phoneme / syllable model storage unit 2.

Next, at the stage of recognizing the speaker, a sentence or word to be uttered is determined in advance, or a new sentence or word is generated by the recognition sentence / word generation unit 4 and the new sentence or word is given to the speaker. The speaker utters the sentence or word, and inputs the voice to the characteristic parameter extraction unit 5. The characteristic parameter extraction unit 5 converts the input voice into an expression format using the same characteristic parameters as the characteristic parameter extraction unit 11. On the other hand, in the sentence / word speech model generation unit 6, according to the sentence or word generated by the recognition sentence / word generation unit 4, or according to a predetermined sentence or word, it is registered in the phoneme / syllable model storage unit 2. Connect the phoneme or syllable models that exist to generate a phonetic model of the sentence or word.

The time series of characteristic parameters obtained by the characteristic parameter extraction unit 5 and the speech model of the sentence or word generated by the sentence / word speech model generation unit 6 are input to the utterance content similarity calculation unit 19. , The degree of similarity between the two is calculated. As a concrete method, for example, “Seiichi Nakagawa:“ Speech recognition by probabilistic model ”, Section 3.1.3, p.
p. 40-50, The Institute of Electronics, Information and Communication Engineers, 1988 ”can be used. The calculated similarity value is temporarily stored in the utterance content similarity storage unit 21.

On the other hand, the time series of the characteristic parameters obtained by the characteristic parameter extracting section 5 and the characteristics of each speaker's voice stored in the speaker characteristic accumulating section 16 are calculated by the speaker similarity calculating section 2
Sent to 2. In the speaker similarity calculation unit 22, in the case of speaker verification, the speaker ID is used to calculate the similarity with the feature of the speaker's voice. In the case of speaker identification, the speaker feature storage unit 1
The similarity with the voice features of all speakers stored in 6 is calculated. The calculated similarity value is temporarily stored in the speaker similarity storage unit 23.

Next, the similarity between the utterance contents stored in the utterance content similarity storage unit 21 and the similarity of the voice of the speaker stored in the speaker similarity storage unit 23 are similar. The degree of similarity is input to the degree combination unit 24 and the total degree of similarity is calculated.
There are various possible methods for calculating the overall similarity. For example, (1) a weighted average value of the two similarities is calculated, (2) a threshold is set in advance for the utterance content similarity, and when the similarity is smaller than the threshold, the overall similarity is calculated. It is conceivable that a very small value such as 0 is given to and the value of the speaker similarity is used as the overall similarity only when the similarity exceeds the threshold value. The calculated overall similarity is input to the speaker recognition determination unit 25, and the speaker recognition is determined. When performing speaker verification, a threshold value indicating the range of variation in the degree of similarity that can be regarded as the voice of the person is read from the threshold value storage unit 10 according to the speaker ID,
Compared with the overall similarity calculated as described above, if the value of the overall similarity is larger than the read threshold value, it is determined that the voice is the person's voice, and it is smaller than the threshold value. If so, it is determined to be the voice of another person. In the case of speaker identification, all the similarities between the input voice and all the registered (stored) speakers are compared, the speaker with the highest overall similarity is selected, and that speaker is selected. Is judged to have been uttered.

Next, an embodiment of the invention relating to claim 8 will be described with reference to FIG. First, at the speaker registration stage, learning voice data is input to the characteristic parameter extraction unit 11 for each speaker to be registered. The utterance content of the learning voice data is determined in advance. Feature parameter extraction unit 11
Then, the input voice is converted into an expression format using characteristic parameters such as cepstrum and pitch. Next, the learning voice data converted into the time series of the characteristic parameters is input to the speaker-adaptive unit 1 with a teacher, and a model of a phoneme or a syllable adapted to the speaker is created. The phoneme / syllable model storage unit 2 for unspecified speakers stores models of each phoneme or syllable created in advance from the voices of many speakers. As a model of each phoneme or syllable, it is possible to employ various representation formats such as a representation of speech in time series of characteristic parameters, a representation of this in hidden Markov model, and the like. As a method for adapting these phoneme / syllable models to a speaker to be registered, various methods that have been established, for example, the document “Jean-Luc Gauvain and Chin-Hui Lee:“ Improved Aco
ustic Modeling With Bayesian Learing, Proc. IEEE I
CASSP 92,1992 ”. The phoneme or syllable model adapted to the speaker is registered in the adapted phoneme / syllable model storage unit 3. In addition, adapted phonemes /
From the error between the connected syllable model and the time series of the characteristic parameters of the training voice data, the allowable range of fluctuation of the voice is estimated, and this value is used as the threshold value for similarity and the threshold value is accumulated together with the speaker ID. Store in part 10.

Here, the learning voice data converted into the time series of the characteristic parameters is the utterance content independent model creating unit 26.
Is also input, and a speech model independent of the utterance content is generated. Then, these models are stored in the utterance content independent model storage unit 27. Next, an error between the connected adaptive phoneme / syllable model sent from the adaptive phoneme / syllable model storage unit 3 and the characteristic parameter time series of the learning voice data, and the utterance content independent model storage unit The allowable variation range of the similarity due to the difference in the utterance content is estimated from the error between the voice model independent of the utterance content and the time series of the characteristic parameters of the learning voice data, which is output from 27, and this value is determined. It is stored together with the speaker ID in the utterance content determination threshold storage unit 28 as a use threshold.

Next, at the stage of recognizing the speaker, a sentence or word to be uttered is determined in advance, or a new sentence or word is generated by the recognition sentence / word generation unit 4 and the new sentence or word is given to the speaker. The speaker utters the sentence or word, and inputs the voice to the characteristic parameter extraction unit 5. The characteristic parameter extraction unit 5 converts the input voice into the same expression format as the characteristic parameter extraction unit 11.
On the other hand, in the sentence / word speech model creating unit 6, the sentence / word is created by the recognition sentence / word creating unit 4, or is registered in the adaptive phoneme / syllable model accumulating unit 3 according to a predetermined sentence or word. Connect the phoneme or syllable models that have been created and create a phonetic model of the sentence or word. At that time, in the case of speaker verification, that is, when it is determined whether or not the voice is the voice of the person, the speaker inputs the ID of the speaker himself with a key or the like.
The phoneme or syllable model corresponding to the registered speaker corresponding to is input to the sentence / word speech model generation unit 6. In the case of speaker identification, that is, in determining who the voice is registered in advance, phoneme or syllable models corresponding to all registered speakers are input to the sentence / word voice model generation unit 6.

The time series of characteristic parameters obtained by the characteristic parameter extraction unit 5 and the speech model of the sentence or word generated by the sentence / word speech model generation unit 6 are input to the similarity calculation unit 7, The degree of similarity between the two is calculated.
As a concrete method of this, for example, the document “Seiichi Nakagawa:“ Speech Recognition by Stochastic Model ”, Section 3.1.3, pp. 40-50,
The method described in "The Institute of Electronics, Information and Communication Engineers, 1988" can be used. The calculated similarity value is sent to the similarity normalization unit 30 and the speaker recognition determination unit 9. At the same time, the time series of the feature parameters obtained by the feature parameter extraction unit 5 is input to the normalization similarity calculation unit 29 together with the voice model independent of the voice content stored in the voice content independent model storage unit 27. , The normalization similarity is calculated, and this value is input to the similarity normalization unit 30. The similarity normalization unit 30 subtracts the normalization similarity value sent from the normalization similarity calculation unit 29 from the similarity value sent from the similarity calculation unit 7 to normalize the similarity. Make a change. The normalized similarity value is input to the speech content determination unit 31. The utterance content determination unit 31 compares the normalized similarity value with the utterance content determination threshold value sent from the utterance content determination threshold value accumulating unit 28, and the normalized similarity value. Is greater than the threshold for utterance content determination, it is determined to be the correct utterance content, and if smaller than Threshold is determined to be the incorrect utterance content. The determination result is sent to the speaker recognition determination unit 9. The speaker recognition determination unit 9 uses the speaker ID in the case of speaker verification based on the value of the similarity sent from the similarity calculation unit 7, and uses the speaker ID from the threshold storage unit 10 to identify the person. The threshold value indicating the range of variation in the degree of similarity that can be regarded as the voice of the person is read, and the speech model of the sentence or word connected to the person's adapted phoneme or syllable model calculated as described above and the recognition sentence or If the similarity value is larger than the read threshold value, it is judged that it is the person's voice, and if it is smaller than the threshold value, it is the other person's voice. Judge that there is. In the case of speaker identification, the similarity between the input speech and the speech model of the sentence or word in which the respective adapted phonemes or syllable models of all the registered speakers are connected is compared to find the most similar. A speaker with a high frequency is selected and it is determined that the speaker has uttered. Further, the speaker recognition determination is comprehensively performed by combining the determination result and the determination result sent from the utterance content determination unit 31.

Next, an embodiment of the present invention will be described with reference to FIG. First, at the speaker registration stage, learning voice data is input to the feature parameter extraction unit 11 for each speaker to be registered. The utterance content of the learning voice data is determined in advance. The characteristic parameter extraction unit 11 converts the input voice into an expression format using characteristic parameters such as cepstrum, pitch, and the like. Next, as a method of expressing the features of the speaker's voice, a plurality of Gaussian distributions are used in the invention of claim 3, and a vector quantization codebook is used in the invention of claim 4. That is, the learning speech data converted into the time series of the characteristic parameters is input to the Gaussian mixture / codebook creating unit 32, and the distribution of the characteristic parameters included in the learning speech data is a combination of a plurality of Gaussian distributions or a vector. It is expressed by a quantization codebook (these are referred to as “mixed Gaussian distribution / codebook” hereinafter). As a method of combining a plurality of mixed Gaussian distributions or creating a codebook, for example, a document “Tomoko Matsui, Sadahi Furui:“ VQ, comparative study of text-independent speaker recognition method using discrete / continuous HMM ””, electronic information Communication Society of Japan Voice Study Material, SP91-89, 19
The method described in "91" can be used. The mixed Gaussian distribution / codebook created by the mixed Gaussian distribution / codebook creating unit 32 is stored in the utterance content independent model storage unit 27 as a speech model independent of the utterance content.

Next, the learning voice data, the text in which the utterance content is represented by a kana or phonetic symbol series, and the mixed Gaussian distribution / codebook are input to the phoneme group model generation unit 13. In the phoneme group model generation unit 13, for each phoneme group in which several phonemes having similar feature parameter values are grouped, the distribution of the feature parameters is represented by a mixed Gaussian distribution / weighted addition of elements of the codebook. When a model is created and these phoneme group models are temporally connected according to the text of the training voice data, the weight of each element is set so that the time series of the characteristic parameters of the training voice data can be expressed most accurately. To decide. For this method,
For example, "Yasuhiro Minami, Tatsuo Matsuoka, Kiyohiro Shikano:" Evaluation of Connected Learning HMMs by Unspecified Speaker Continuous Speech Database ", IEICE Speech Society Material, SP91-113, 19
The method described in "92" can be used. The phoneme group model thus created is stored in the phoneme group model storage unit 3 together with the speaker ID indicating the speaker who uttered the learning voice. Furthermore, the allowable variation range of the voice is estimated from the error between the connected phoneme group model and the time series of the characteristic parameters of the learning voice data, and this value is used as the threshold value for the similarity with the speaker ID. Threshold storage unit 1
Store to 0. Further, the error between the connected phoneme group model and the time series of the characteristic parameters of the learning voice data, and the speech model independent of the speech content sent from the speech content independent model accumulating unit 27 and the learning speech data. The allowable variation range of the similarity due to the difference in utterance content is estimated from the error with the characteristic parameter time series, and this value is used as the utterance content determination threshold value and the utterance content determination threshold value is stored together with the speaker ID. Store in part 28.

Next, at the stage of recognizing the speaker, the sentence or word to be uttered is determined in advance, or a new sentence or word is generated by the recognition sentence / word generation unit 4 and the speaker is informed of this. The speaker speaks the sentence or word, and inputs the voice to the characteristic parameter extraction unit 5. The characteristic parameter extraction unit 5 converts the input voice into the same expression format as the characteristic parameter extraction unit 11.
At the same time, the sentence / word speech model generation unit 6 accumulates in the phoneme group model accumulation unit 3 according to the sentence or word generated by the recognition sentence / word generation unit 4 or a predetermined sentence or word. Connected phoneme group models to generate a speech model of the sentence or word. Here, in the case of speaker verification, that is, when it is determined whether or not the voice is the voice of the person, the speaker is asked to input the ID of the speaker himself or herself by a key or other input means, and the registered speaker corresponding to the ID is registered. The phoneme group model is input to the sentence / word voice model generation unit 6. In the case of speaker identification, that is, in determining who the voice is registered in advance, models of phonemes or syllables corresponding to all registered speakers are sequentially input to the sentence / word voice model generation unit 6.

After that, the time series of the characteristic parameters obtained by the characteristic parameter extracting section 5 and the voice model of the sentence or the word generated by the sentence / word voice model generating section 6 are input to the similarity calculating section 7, The similarity is calculated, and the value is input to the similarity normalization unit 30 and the speaker recognition determination unit 9. At the same time, the time series of the characteristic parameters obtained by the characteristic parameter extraction unit 5 is input to the normalization similarity calculation unit 29 together with the speech model independent of the utterance content stored in the utterance content independent model storage unit 27, The similarity for normalization is calculated, and the value is input to the similarity normalization unit 30. The similarity normalization unit 30 normalizes the similarity by subtracting the similarity value sent from the normalization similarity calculation unit 29 from the similarity value sent from the similarity calculation unit 7. .. Further, the normalized similarity value is input to the utterance content determination unit 31. The utterance content determination unit 31 compares the normalized similarity value with the utterance content determination threshold value sent from the utterance content determination threshold value accumulating unit 28, and the normalized similarity value. Is greater than the threshold for utterance content determination, it is determined to be the correct utterance content, and if smaller than Threshold is determined to be the incorrect utterance content. The determination result is sent to the speaker recognition determination unit 9. In the speaker recognition determination unit 9,
Based on the value of the similarity sent from the similarity calculator 7,
In the case of speaker verification, the threshold storage unit 1 is used by using the speaker ID.
From 0, a threshold value indicating the range of variation in the degree of similarity that can be regarded as the voice of the person is read out and calculated as described above.
The speech model of a sentence or word connected to the phoneme or syllable group model of the person is compared with the similarity of the input speech of the recognition sentence or word, and the similarity value is larger than the read threshold value. If it is smaller than the threshold value, it is determined to be the voice of another person. In the case of speaker identification, the similarity between the input speech and the speech model of the sentence or word in which the phoneme or syllable group models adapted to all the registered speakers are connected is compared to determine the highest similarity. A large speaker is selected and it is determined that the speaker has uttered. Further, the speaker recognition determination is comprehensively performed by combining the determination result and the determination result sent from the utterance content determination unit 31.

Next, an embodiment in which the invention of claim 8 is applied to the inventions of claims 6 and 7 will be described with reference to FIG. At the stage of registering the speakers, for each speaker to be registered, each learning voice data is input to the characteristic parameter extraction unit 11 and converted into a time series of characteristic parameters (for example, cepstrum, pitch), It is input to the speaker feature calculation unit 15 and the features of the speaker's voice independent of the utterance content (not dependent on the utterance content) are extracted. As a method for extracting a feature of a voice independent of the utterance content, for example, a document “Tomoko Matsui, Sadahiro Furui:“ VQ, comparative study of text-independent speaker recognition methods using discrete / continuous HMM ”, IEICE. The method described in "Speech Study Group Material, SP91-89, 1991", that is, the distribution of feature parameters is represented by a vector quantization codebook, or a method of expressing a plurality of Gaussian distributions can be used. The extracted voice feature is stored in the speaker feature storage unit 16 together with the speaker ID.

On the other hand, after the voice data of a large number of speakers is input to the characteristic parameter extraction unit 17 and converted into a time series of characteristic parameters, it is input to the phoneme / syllable model generation unit 18 and common to a large number of speakers. That is, a speaker-independent phoneme / syllable model is created. This method is described in, for example, the document “Yasuhiro Minami, Tatsuo Matsuoka, Kiyohiro Kano:“ Evaluation of Connected Learning HMM by Unspecified Speaker Continuous Speech Database ”, Institute of Electronics, Information and Communication Engineers, Speech Study Group, SP91-113, 1992”. The method described can be used. In this case, the utterance content of the learning voice data is determined in advance, and the utterance content represented by a kana or a phonetic symbol sequence is input to the phoneme / syllable model creating unit 18 together with the voice. As models of each phoneme or syllable, there are a model in which a voice is represented by a time series of characteristic parameters, or a model in which it is represented by a hidden Markov model. The created phoneme / syllable model is stored in the phoneme / syllable model storage unit 2.

At the same time, the characteristic parameter extraction unit 17
The time series of the feature parameters extracted in (4) is input to the utterance content independent model creation unit 26, and a voice model independent of the utterance content is generated. Then, those models are stored in the utterance content independent model storage unit 27. Next, from the phoneme / syllable model sent from the phoneme / syllable model storage unit 2 and the utterance content independent model sent from the utterance content independent model storage unit 27, the allowable variation range of the similarity due to the difference in utterance content is determined. The estimated value is stored in the utterance content determination threshold storage unit 28 as the utterance content determination threshold value.

At the stage of recognizing the speaker, a sentence or word to be uttered is determined in advance, or a new sentence or word is generated by the recognition sentence / word generation unit 4 and the new sentence or word is shown to the speaker. The person utters the sentence or word and inputs the voice into the characteristic parameter extraction unit 5.
The characteristic parameter extraction unit 5 converts the input voice into an expression format using the same characteristic parameter as the characteristic parameter extraction unit 11. On the other hand, in the sentence / word speech model generation unit 6, according to the sentence or word generated by the recognition sentence / word generation unit 4, or in accordance with a predetermined sentence or word, it is registered in the phoneme / syllable model storage unit 2. Connect the phoneme or syllable models that exist to generate a phonetic model of the sentence or word. The time series of the characteristic parameters obtained by the characteristic parameter extraction unit 5 and the voice model of the sentence or the word generated by the sentence / word voice model generation unit 6 are input to the utterance content similarity calculation unit 19, and both of them are input. The degree of similarity is calculated. As a concrete method, for example, the method described in the document “Seiichi Nakagawa:“ Speech recognition by probabilistic model ”, Section 3.1.3, pp. 40-50, The Institute of Electronics, Information and Communication Engineers, 1988” is used. be able to. The calculated similarity value is sent to the similarity normalization unit 30. At the same time, the time series of the characteristic parameters obtained by the characteristic parameter extraction unit 5 is input to the normalization similarity calculation unit 29 together with the speech model independent of the speech content stored in the speech content independent model storage unit 27. , The normalization similarity is calculated, and the value is input to the similarity normalization unit 30. In the similarity normalization unit 30, the utterance content similarity calculation unit 1
The similarity degree is normalized by subtracting the value of the similarity degree for normalization sent from the similarity degree calculation section for normalization 29 from the similarity degree value sent from 9. The normalized similarity value is input to the speech content determination unit 31. Speech content determination unit 3
In 1, the normalized similarity value is compared with the utterance content determination threshold value sent from the utterance content determination threshold value accumulating unit 28, and the normalized similarity value is determined as the utterance content determination value. If it is larger than the usage threshold value, it is determined that it is the correct utterance content, and if it is smaller than it, it is determined that it is the erroneous utterance content. The determination result is sent to the speaker recognition determination unit 25.

On the other hand, the time series of the characteristic parameters obtained by the characteristic parameter extracting section 5 and the characteristics of each speaker's voice stored in the speaker characteristic accumulating section 16 are calculated by the speaker similarity calculating section 2
Sent to 2. In the speaker similarity calculation unit 22, in the case of speaker verification, the speaker ID is used to calculate the similarity with the feature of the speaker's voice. In the case of speaker identification, the speaker feature storage unit 1
The similarity between all the speaker's voice features stored in 6 is calculated. The calculated similarity value is temporarily stored in the speaker similarity storage unit 23, and then the speaker recognition determination unit 2
Input to 5. The speaker recognition determination unit 25 determines the speaker. When performing speaker verification, the threshold storage unit 1
According to the speaker ID, a threshold value indicating the range of variation of the similarity degree that can be regarded as the voice of the person is read from 0, compared with the similarity degree sent from the speaker similarity degree accumulation unit 23, and the similarity degree If the value is larger than the read threshold value, it is determined to be the voice of the person, and if smaller than the threshold value, it is determined to be the voice of another person. In the case of speaker identification, all the similarities between the input voice and all the registered (stored) speakers are compared, the speaker with the highest similarity is selected, and it is determined that the speaker uttered. judge. Further, the speaker recognition determination is comprehensively performed by combining the determination result and the determination result sent from the utterance content determination unit 31.

[0037]

As described above, according to the first to fifth aspects of the invention, models of phonemes or syllables adapted to the voice of each speaker to be recognized are created and registered, and those models are connected. Since the similarity between the model of the voice of the sentence or word generated by the input voice and the input voice converted into the expression format using the characteristic parameters is obtained, the speaker who uttered the input voice is recognized. The content to be uttered can be changed, and a high degree of similarity cannot be obtained unless the content is uttered. Therefore, by playing the recorded voice, it is possible to prevent impersonating another person.

According to the sixth aspect of the invention, the input voice for recognizing the speaker is converted into an expression form using the characteristic parameter, and the input voice in the expression form and the utterance content, that is, the words in the expression form are used. It recognizes the speaker who uttered the input voice by combining the similarity with the voice characteristics of each speaker and the similarity with the words to be uttered. You can change it, and you cannot get a high degree of similarity unless you utter that content, so by playing back the recorded voice,
You can prevent impersonating others.

The invention of claim 8 relates to claim 1, 2, 3,
In the speaker recognition methods described in 4, 5, 6 and 7, a voice model independent of the utterance content is created, and the similarity between the voice model and the input voice converted into the expression format using the feature parameter is calculated. By using the phoneme or syllable (group) model, the value of the similarity between the sentence or the word voice model connected to the input voice converted into the expression format using the characteristic parameter is normalized, Is the content of the word, recording time,
It is possible to perform speaker recognition that is not easily affected by variations in similarity due to differences in transmission conditions, microphones, and other test conditions.

Next, an experimental example will be described. The experiment is male 23
, 13 women for 5 months (Time A,
Sentence data for B and C) (1 sentence length is 4 seconds on average)
Target. These voices are converted into a time series at a feature amount that has been used in the past, that is, every fine time of the cepstrum. The cepstrum has a sampling frequency of 12 kHz, a frame length of 32 ms, a frame period of 8 ms, and LPC analysis (Line
ar Predictive Coding, linear predictive analysis) Extracted with degree 16. For the learning, 10 sentences uttered in the period A were used, and in the test, 5 sentences uttered in the periods B and C were used one by one.

The phoneme group model for each speaker is based on the invention of claim 3 and is based on the combination of 64 Gaussian distributions ("Matsui Tomoko,
Satoshi Furui: "Comparison study of text-independent speaker recognition methods using VQ, discrete / continuous HMM," Institute of Electronics, Information and Communication Engineers, Speech Study Group, SP91-89, 1991 "). The number of phoneme groups is 25. The results were evaluated by the rejection rate when the person uttered a sentence different from the designated sentence. The higher the rejection rate, the higher the rate of rejection of the person's voice when the content of the voice varies depending on the reproduction of the recorded voice. Here, the input voice having a degree of similarity higher than the threshold is accepted as the voice uttered by the person himself / herself and rejected otherwise. This threshold was set a posteriori for each speaker so that the voice uttered by the person himself / herself was not rejected. The result is No. 7
Shown in the figure. This result shows that the speaker recognition method of the present invention has a rejection error rate that is almost half that in the case where the similarity degree is not normalized. From the above, the effectiveness of the invention of claim 8 was proved.

[Brief description of drawings]

FIG. 1 is a block diagram showing an example of a recognition device to which an embodiment of the invention of claims 1 and 2 is applied.

FIG. 2 is a block diagram showing an example of a recognition device to which an embodiment of the invention of claim 3 is applied.

FIG. 3 is a block diagram showing an example of a recognition device to which an embodiment of the invention of claim 4 is applied.

FIG. 4 is a block diagram showing an example of a recognition device to which the embodiments of the inventions of claims 6 and 7 are applied.

FIG. 5 is a block diagram showing an example of a recognition device to which an embodiment of the invention of claim 8 relating to claims 1 and 2 is applied.

FIG. 6 is a block diagram showing an example of a recognition device to which an embodiment of the invention of claim 8 relating to claims 3, 4 and 5 is applied.

FIG. 7 is a block diagram showing an example of a recognition device to which an embodiment of the invention of claim 8 relating to claims 6 and 7 is applied.

FIG. 8 is a diagram showing an experiment related to claim 8;

[Explanation of symbols]

 1 speaker adaptation unit 2 phoneme / syllable model storage unit for unspecified speaker 3 adapted phoneme / syllable model storage unit 4 recognition sentence / word generation unit 5 feature parameter extraction unit 6 sentence / word speech model creation unit 7 similarity Calculation unit 9 Speaker recognition determination unit 10 Threshold storage unit 11 Feature parameter extraction unit 30 Similarity normalization unit

Claims (8)

[Claims] 1. Means for creating and registering a phoneme or syllable model adapted to the voice of each speaker to be recognized, and means for connecting these phoneme or syllable models to generate a sentence or word speech model. A sentence or word speech model having means for converting an input speech into an expression format using a characteristic parameter, and connecting the input speech converted into the expression format with a phoneme or syllable model registered in advance for a speaker. A speaker recognition method characterized by recognizing a speaker who uttered the input voice by obtaining a degree of similarity with the speaker. 2. Each time recognition is performed, a speaker designates a sentence or word to be uttered, and the input speech in which the sentence or word is uttered is converted into an expression format using a characteristic parameter, and a phoneme or syllable. 2. The speaker recognition method according to claim 1, further comprising: recognizing the speaker who uttered the input voice by obtaining the similarity between the model of FIG. 1 and the voice model of the sentence or word generated. .. 3. A means for converting an input voice into an expression form using a characteristic parameter, a means for calculating a plurality of Gaussian distributions expressing the characteristics of each speaker's voice independent of utterance content, and those Gaussian distributions. A means for creating a phoneme or syllable group model by giving an appropriate weight to each phoneme or syllable group, and a means for connecting the phoneme or syllable group models in time to create a speech model of a sentence or word. Using the training speech data of each speaker, the distribution of the characteristic parameters in the speech data is approximated by a plurality of Gaussian distributions, and the weighted addition of these Gaussian distributions is performed for each phoneme or syllable group. Create a phoneme or syllable group model that approximates the distribution of characteristic parameters, and use the phoneme or syllable group model according to the utterance content (sentence or word) of the input speech that should recognize the speaker. A speaker recognition method characterized by recognizing a speaker by comparing a model of a speech of a sentence or a word connected in terms of time with a time-series representation of the above-mentioned input speech. 4. A means for converting an input voice into an expression format using a feature parameter, a means for calculating a vector quantization codebook expressing the features of each speaker's voice independent of utterance content, A method for creating a phoneme or syllable group model by giving appropriate weights to the codebook corresponding to each phoneme or syllable group, and creating a phonetic model of a sentence or word by connecting the phoneme or syllable group model in time. Means for approximating the distribution of the characteristic parameters in the speech data with a codebook by using the learning speech data of each speaker, and performing weighted addition of these codebook elements for each phoneme or syllable group. A phoneme syllable group model that approximates the distribution of the characteristic parameters of is created, and the phoneme syllable group model is temporally connected according to the utterance content (sentence or word) of the input speech that should recognize the speaker. Is a speaker recognition method characterized by recognizing a speaker by comparing a model of a voice of a word with a time-series representation of the input voice of the feature parameter. 5. Each time recognition is performed, a speaker designates a sentence or word to be uttered, and the input speech in which the sentence or word is uttered is converted into an expression format using a characteristic parameter, and the phoneme or 4. The speaker that uttered the input speech is recognized by comparing with a speech model of the sentence or word connected with a syllable group model.
Or the speaker recognition method described in 4. 6. A means for recognizing a speaker by extracting the characteristics of a speaker's voice independent of the utterance content, a means for recognizing the utterance content (word) of the voice of an unspecified speaker, and an input voice. Means for converting into an expression format using characteristic parameters, and obtaining the respective similarity between the input speech converted into the above expression format, the characteristics of the voice of each speaker and the words to be uttered, Combining degrees logically and numerically,
A speaker recognition method comprising recognizing a speaker who has generated the input voice. 7. The above-mentioned two types of the above-mentioned two types regarding a sentence or word to be uttered by a speaker for each recognition, and an input voice in which the sentence or word is uttered is converted into an expression format using a characteristic parameter. Combining similarities,
7. The speaker recognition method according to claim 6, wherein the speaker uttering the input voice is recognized. 8. Claims 1, 2, 3, 4, 5, 6 and 7
In the speaker recognition method described in any of the above, the similarity between the input speech converted into the expression format using the characteristic parameter and the sentence or word speech model in which the phoneme or syllable (group) model is connected. A method for recognizing a speaker, characterized by normalizing variations in the degree of text or words, recording time, transmission system, microphone and other test conditions.