JP2005148640A - Device, method and program of voice recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition device which can recognize the utterer of the voice input by combining and using a corresponding relation of a plurality of voice spectral characteristics and each part in a voice path which becomes its forming factor. <P>SOLUTION: A characteristics extracting part 202 determines the shape parameter of a voice path model based on the voice input from a recognition object. The voice path model consists of a first and a second sound tube corresponding to the oral cavity and the pharyngeal cavity respectively, a connected small sound tube corresponding to the larynx cavity and at least one conic tube corresponding to the piriform recess. The characteristics extracting part 202 determines the shape parameter of the voice path model as a recognition shape parameter based on the voice input from the utterer at the time of recognition. A threshold value comparing part 222 recongizes whether the utterer is a registered recognition object or not based on the comparison result of the recognition shape parameter and the registration shape parameter at a similarity degree calculating part 220. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a voice authentication device, a voice authentication method, and a voice authentication program for performing personal authentication using speaker recognition that automatically determines a speaker based on individual differences in voice.

  Access control to prevent outsiders from entering important facilities and rooms, access management to prevent system destruction from outside the system and unauthorized access from inside the system, and so-called e-commerce The situation where “personal authentication technology” is required to prevent fraudulent acts such as “spoofing” is increasing.

  Conventionally, a method using a combination of a “user ID” and a “password”, a “secret key” in a public key cryptosystem, or the like is employed for such personal authentication.

  Furthermore, in order to further improve the reliability of personal authentication, so-called “biometrics” authentication techniques using physical characteristics and behavioral characteristics of the person such as fingerprints and irises are often used.

  On the other hand, there is an increasing expectation for personal authentication technology using voice, which is a kind of “biometrics”. This is because with recent developments in speech processing technology, it is possible to easily realize a communication-related system if the characteristics of a person to be authenticated, which can use a conventional communication system as it is, can be used for personal authentication. This is because it is expected (see, for example, Non-Patent Document 1).

  However, although voice authentication (hereinafter referred to as “voice authentication”) has the above-mentioned advantages, compared with fingerprints and irises, the relevance to the physical characteristics of individuals is low in the conventional method. In order to use it as a technique for personal authentication, further improvement in accuracy is necessary.

Here, there is an example in which modeling of the vocal tract lower structure is attempted from the correspondence between the lower structure of the vocal tract and the three-dimensional MRI moving image data (see, for example, Non-Patent Document 2). Conventionally, it has not always been clear how to do this.
Furui, “Authentication by Voice, Part 1 Mechanism and Technology Trend of Voice Authentication, Information Processing, Vol. 40, No. 11, 1999, November Takemoto, Honda, Masaki, Shimada, Fujimoto, “Modeling of the vocal tract structure based on 3D MRI video data”, Proceedings of the Acoustical Society of Japan, pp. 281-282, September 2003

  The present invention has been made in order to solve the above-described problems, and its purpose is to provide a correspondence relationship between a plurality of features on the speech spectrum and each part in the vocal tract that is the generation factor. Is used to provide a voice authentication device, a voice authentication method, and a voice authentication program capable of specifying a speaker of an input voice.

  In order to achieve such an object, according to one aspect of the present invention, there is provided a voice authentication device for extracting a feature parameter for determining a shape parameter of a vocal tract model based on a voice input from a person to be authenticated. The vocal tract model includes a first acoustic tube portion corresponding to the oral cavity, a second acoustic tube portion coupled to the first acoustic tube portion and corresponding to the pharyngeal cavity, and a second acoustic tube portion. And at least one conical tube connected to the bottom surface of the second acoustic tube portion and corresponding to the piriform fossa for feature extraction during learning Storage means for storing the shape parameter determined by the means in association with the person to be authenticated as a registered shape parameter, and the feature extraction means is based on the voice input from the speaker at the time of authentication. Shape parameter Determined as testimony shape parameter, to identify whether a person to be authenticated the speaker has been registered, further comprising a similarity comparison means for comparing the authentication shape parameter and registration shape parameter.

  Preferably, the feature extraction means minimizes the difference between the initial value determination means for determining the initial value of the shape parameter based on the voice input, and the transfer function of the vocal tract model based on the initial value and the input spectrum of the voice input. Correcting means for correcting the shape parameter to

  Preferably, the first acoustic tube portion includes a plurality of first acoustic tubes connected to each other, and the second acoustic tube portion includes a plurality of second acoustic tubes connected to each other.

  According to another aspect of the present invention, there is provided a voice authentication method including a step of determining a shape parameter of a vocal tract model based on a voice input from a person to be authenticated at the time of learning. A first acoustic tube portion corresponding to the first acoustic tube portion, a second acoustic tube portion corresponding to the pharyngeal cavity, and a bottom surface of the second acoustic tube portion corresponding to the laryngeal cavity A small acoustic tube connected to the bottom surface of the second acoustic tube portion, and at least one conical tube corresponding to the piriform fossa, and the shape parameter determined at the time of learning as a registered shape parameter A step of determining the shape parameter of the vocal tract model as an authentication shape parameter based on the voice input from the speaker at the time of authentication, and the registration of the authentication shape parameter. Based on the comparison result between Jo parameter, further comprising the step of identifying whether a person to be authenticated the speaker has been registered.

  Preferably, the step of determining the shape parameter of the vocal tract model includes the step of determining an initial value of the shape parameter based on the speech input, a transfer function of the vocal tract model based on the initial value, and an input spectrum of the speech input. Modifying the shape parameters to minimize the difference.

  Preferably, the first acoustic tube portion includes a plurality of first acoustic tubes connected to each other, and the second acoustic tube portion includes a plurality of second acoustic tubes connected to each other.

  According to still another aspect of the present invention, there is provided a voice authentication program for causing a computer to execute a voice authentication process, wherein the voice authentication process is performed based on a voice input from a person to be authenticated during learning. Determining a shape parameter of the model, the vocal tract model comprising: a first acoustic tube portion corresponding to the oral cavity; a second acoustic tube portion coupled to the first acoustic tube portion and corresponding to the pharyngeal cavity; A small acoustic tube coupled to the bottom surface of the second acoustic tube portion and corresponding to the laryngeal cavity, and at least one conical tube coupled to the bottom surface of the second acoustic tube portion and corresponding to the piriform fossa Storing the shape parameter determined at the time of learning in association with the person to be authenticated as a registered shape parameter, and recognizing the shape parameter of the vocal tract model based on the voice input from the speaker at the time of authentication. Further comprising determining a shape parameter, and a step of, based on a result of comparison between authentication shape parameter and registration shape parameter, to identify whether a person to be authenticated the speaker has been registered.

  Preferably, the step of determining the shape parameter of the vocal tract model includes the step of determining an initial value of the shape parameter based on the speech input, a transfer function of the vocal tract model based on the initial value, and an input spectrum of the speech input. Modifying the shape parameters to minimize the difference.

  Preferably, the first acoustic tube portion includes a plurality of first acoustic tubes connected to each other, and the second acoustic tube portion includes a plurality of second acoustic tubes connected to each other.

  The voice authentication device, the voice authentication method, and the voice authentication program according to the present invention can improve the relevance with the physical characteristics of an individual in voice authentication and improve the accuracy of voice authentication. Is possible.

Embodiments of the present invention will be described below with reference to the drawings.
[Hardware configuration]
FIG. 1 is a conceptual diagram showing an example of a voice authentication system 1000 using a voice authentication apparatus to which a voice authentication method and a voice authentication program of the present invention are applied.

  Referring to FIG. 1, the voice authentication system 1000 includes a computer 100 that determines whether to permit access to the authentication target person 2 based on the utterance of the authentication target person 2.

  That is, hereinafter, a case where the voice authentication method of the present invention is applied to access right management will be described as an example.

  Referring to FIG. 1, this computer 100 reads / writes information to / from a CD-ROM drive 108 and a flexible disk (hereinafter referred to as FD) 116 for reading information on a CD-ROM (Compact Disc Read-Only Memory). A computer main body 102 including an FD drive 106, a display 104 as a display device connected to the computer main body 102, a keyboard 110 and a mouse 112 as input devices also connected to the computer main body 102, and a voice input device As a microphone 132 and a speaker 134 as an audio output device.

  When the voice authentication method of the present invention is applied to entrance management or the like, the computer 100 operates as a part of the entrance management system. When the user is authenticated, the computer 100 performs an unlocking process of the gate. Become. When the voice authentication method of the present invention is applied to electronic commerce or the like, the voice input from the microphone 132 is converted into a format suitable for communication, and then the partner computer system via the network 310. 300. In the other party's computer system 300, voice authentication processing as described below is performed to authenticate the person 2 to be authenticated.

  FIG. 2 is a block diagram showing the hardware configuration of the computer 100. As shown in FIG.

  As shown in FIG. 2, in addition to the CD-ROM drive 108 and the FD drive 106, the computer main body 102 constituting the computer 100 includes a CPU (Central Processing Unit) 120 connected to the bus BS, and a ROM ( A memory 122 including a read only memory (RAM) and a random access memory (RAM), a direct access memory device, for example, a hard disk 124, and an interface 128 for exchanging data with a microphone 132 or a speaker 134 are included. A CD-ROM 118 is attached to the CD-ROM drive 108. An FD 116 is attached to the FD drive 106.

  The interface 128 can also be used, for example, to communicate with the counterpart computer system 300.

  As will be described later, when the voice authentication program of the present invention operates, a database that stores information that is the basis of the operation will be described as being stored in the hard disk 124.

  The CD-ROM 118 may be another medium, such as a DVD-ROM (Digital Versatile Disc) or a memory card, as long as it can record information such as a program installed in the computer main body. In this case, the computer main body 102 is provided with a drive device that can read these media.

  The main part of the voice authentication apparatus of the present invention is constituted by computer hardware and software executed by the CPU 120. Generally, such software is stored and distributed in a storage medium such as a CD-ROM 118 or FD 116, read from the storage medium by the CD-ROM drive 108 or FD drive 106, and temporarily stored in the hard disk 124. Alternatively, when the device is connected to the network 310, it is temporarily copied from the server on the network to the hard disk 124. Then, the data is further read from the hard disk 124 to the RAM in the memory 122 and executed by the CPU 120. In the case of network connection, the program may be directly loaded into the RAM and executed without being stored in the hard disk 124.

  The computer hardware itself and its operating principle shown in FIGS. 1 and 2 are general. Therefore, the most essential part of the present invention is software stored in a storage medium such as the FD 116, the CD-ROM 118, and the hard disk 124.

  As a general tendency, various program modules are prepared as a part of a computer operating system, and an application program generally calls a module in a predetermined arrangement and advances the processing when necessary. In such a case, the software itself for realizing the voice authentication device does not include such a module, and the voice authentication device is realized only when the computer cooperates with the operating system. However, as long as a general platform is used, it is not necessary to distribute software including such modules, and the software itself not including these modules and the recording medium storing the software (and the software distributes on the network). Data signal) can be considered to constitute the embodiment.

[Voice authentication based on personality generation factors]
FIG. 3 is a diagram illustrating a correspondence relationship between features on the voice spectrum and parts in the vocal tract.

  As will be described below, the voice authentication device and the voice authentication method of the present invention use a combination of a plurality of features on the voice spectrum and corresponding parts of each part in the vocal tract as a generation factor thereof. It is possible to specify the speaker of the input voice.

  In particular, in the present invention, voice authentication is performed by extracting a speaker's personality factor from input voice (vowel). At this time, in the voice authentication, the “text-dependent” authentication in which the utterance content (keyword) of the voice used for the authentication is determined in advance, the “text-independent” authentication in which any words can be generated, the device The voice authentication according to the present invention can be applied to any of “text designation type” authentication in which a new keyword is designated from the apparatus side to the authentication subject every time it is used.

  On the other hand, FIG. 4 is a conceptual diagram showing a mid-sectional view of the speech generation system.

  Referring to FIG. 3 and FIG. 4, the distribution pattern of the maximum (maximum point) of the voice spectrum corresponds to “the length of the vocal tract”. The “vocal tract” refers to the space from the glottis through the pharyngeal cavity, mouth and lips.

  The low-order formant corresponds to the “relationship between pharyngeal cavity and oral cavity length, cross-sectional area and volume”. The fourth formant, which is the fourth maximum point from the low frequency side, corresponds to the “shape of the laryngeal cavity”.

  In addition, the number of zeros (minimum points) existing in the high frequency band, the frequency, the bandwidth, and the relative energy difference with the surrounding poles, as described in detail later, are as follows. Corresponds to “shape”.

  The vocal tract shape is the main factor that determines the voice quality of each person, that is, the individuality of the voice. In other words, it can be said that the main generation factor of the voice personality is the individual difference of the vocal tract shape.

  Hereinafter, the part in the vocal tract in FIG. 3 will be described in more detail.

(Voice tract chief)
“Vocal tract length” refers to the length from the glottis to the lips. Vocal tract length is highly correlated with age, gender, and individual physique. The longer the acoustic tube is, the lower its resonance frequency is, so there is a correspondence between the vocal tract length and the distribution pattern of the speech spectrum poles. Therefore, the vocal tract length of the speaker can be obtained from the voice.

(Relationship between pharyngeal cavity and oral cavity length, cross-sectional area, volume)
The relationship between the length of the pharyngeal cavity and oral cavity, cross-sectional area, and volume determines the lower-order formant.

  FIG. 5 is a diagram showing a theoretical relationship between a change in the cross-sectional area of the pharyngeal cavity and the oral cavity and a low-order formant in a two-section model of the vocal tract.

  FIG. 5A shows a case where the cross-sectional area of the pharyngeal cavity is larger than that of the oral cavity and the cross-sectional area of the pharyngeal cavity is larger than that of the oral cavity. It shows a small case. On the other hand, FIG. 5B is a diagram showing changes in formant frequency corresponding to two cases in which the cross-sectional area relationship between the pharyngeal cavity and the oral cavity shown in FIG.

  First, when the cross-sectional area of the pharyngeal cavity increases as shown in the upper side of FIG. 5A, as shown in the upper side of FIG. 5B, the first formant (F1) decreases and the second formant (F2) decreases. A rise occurs.

  On the other hand, as shown in the lower side of FIG. 5A, when the cross-sectional area of the oral cavity increases, the first formant (F1) rises and the second formant (F2) falls.

  The relationship shown in FIG. 5 is also observed between the measured value based on nuclear magnetic resonance imaging (MRI) and the recorded sound, and the first formant frequency is the pharyngeal cavity cross-sectional area, The formant frequency is correlated with the oral cross section.

  FIG. 6 is a diagram showing the correlation between the average area of the pharyngeal cavity and the first formant frequency as a measured value by MRI. FIG. 6A shows the relationship between the average area of the pharyngeal cavity measured by MRI and the measured value of the frequency of the first formant of the vowel “A” for a plurality of subjects. FIG. 6B shows the relationship between the average area of the pharyngeal cavity measured by MRI and the measured value of the first formant frequency of the vowel “E” for a plurality of subjects.

  There is a negative correlation between the measured value of the average area of the pharyngeal cavity and the measured value of the frequency of the first formant of the vowel.

  FIG. 7 is a diagram showing the correlation between the average oral cavity area and the first formant frequency as a measured value by MRI. FIG. 7A shows the relationship between the average area of the oral cavity measured by MRI and the measured value of the frequency of the first formant of the vowel “A” for a plurality of subjects. FIG. 7B shows the relationship between the average area of the oral cavity measured by MRI and the measured value of the frequency of the first formant of the vowel “E” for a plurality of subjects.

  There is a positive correlation between the measured value of the average area of the oral cavity and the measured value of the frequency of the first formant of the vowel.

  Using the above relationship, the rough shape of the pharyngeal cavity and the oral cavity can be estimated from the low-order formant.

(Shape of laryngeal cavity)
The “laryngeal cavity” is a thin tube that forms part of the hypopharyngeal cavity.

  FIG. 8 is a drawing for explaining the shape of the laryngeal cavity. 8A shows the laryngeal cavity surrounded by a white line in the MRI image, and FIG. 8B shows the three-dimensional shape of the hypopharyngeal cavity obtained from the MRI image in a wire frame. Shows the wire frame with bold lines and the dark gray scale. As shown in FIG. 8 (b), the hypopharyngeal cavity includes a laryngeal cavity and, in principle, piriform fossa, which will be described later, on both sides of the laryngeal cavity.

  FIG. 9 is a diagram showing the three-dimensional shape of the hypopharyngeal cavity for three persons. In FIG. 9, FIGS. 9 (a1) to (a3) are views of the three-dimensional shape of the hypopharyngeal cavity obtained from the MRI images for three subjects, respectively, as viewed from the front, FIGS. 9B1 to 9B3 are views of these wire frames as viewed from the left side.

  As shown in this figure, there are individual differences in the shape and size of the laryngeal cavity.

  FIG. 10 shows the crossing of each part of the hypopharyngeal cavity when vowels (/ a /, / i /, / u /, / e /, / o /) are uttered for the three persons shown in FIG. It is a figure which shows a surface shape as a parameter from the distance from a glottis. Each of FIGS. 10 (c1) to (c3) corresponds to each person shown in FIGS. 9 (a1) to (a3).

  As shown in FIG. 10, even if the vowel to be uttered changes, the shape change is extremely small for each person.

  FIG. 11 is a diagram illustrating a speech spectrum of the vowel “e”. In FIG. 11, the fourth formant is indicated by an arrow.

  The laryngeal cavity is acoustically independent in the vocal tract and acts as a Helmholtz resonator. The shape and size of the laryngeal cavity determines the frequency, bandwidth, and energy of the fourth formant of the voice spectrum. That is, the individual difference in the shape of the laryngeal cavity appears in the fourth formant.

  FIG. 12 is a diagram illustrating the frequency of the fourth formant of speakers A to K.

  As shown in FIG. 12, the fourth formant frequency varies among individuals. Therefore, it can be said that the laryngeal cavity is a factor in generating personality of speech.

  By determining the shape of the Helmholtz resonator corresponding to the fourth formant, the shape of the speaker's laryngeal cavity can be determined.

  FIG. 12 also shows the zero point frequency due to the piriform fossa for each speaker, which will be described later.

(Piriform shape)
FIG. 13 is a diagram showing the position of the piriform fossa in the hypopharyngeal cavity. FIG. 13 shows the three-dimensional shape of the hypopharyngeal cavity obtained from the MRI image as a wire frame when viewed from the front, and the piriform fossa portion is indicated by a thick line and a dark gray scale. is there.

  The piriform fossa is a branch duct that exists in the hypopharyngeal cavity in principle, one on each side. When viewed from the front, the piriform fossa has a shape as shown in FIG. 13, and this shape can be approximated by a cone.

  As shown in FIG. 10, like the laryngeal cavity, the shape, length, and size of the piriform fossa vary among individuals, and even if the vowel to be uttered changes, the shape change is extremely small.

  Since the piriform fossa is a branch pipe in the vocal tract, a zero point (minimum point) is generated on the voice spectrum. The shape, length, and size of the piriform fossa determine the number of zeros that appear in the high frequency band of the speech spectrum, the frequency, the bandwidth, and the relative energy difference from the poles around the zeros.

  FIG. 14 is a diagram illustrating the position of the zero point due to the piriform fossa on the speech spectrum of the vowel “e”.

  In FIG. 14, an arrow is added to the zero point due to the piriform fossa.

  In addition, FIG. 12 shows the frequency of the zero point by ten piriform fossae of speakers A to K as described above.

  From FIG. 12, it can be seen that there is an individual difference in the frequency of the zero point due to the piriform fossa. This frequency corresponds to individual differences in the shape of the piriform fossa. Therefore, it can be said that the piriform fossa is also a factor in generating personality of speech.

  When the shape, length and size of the two piriform fossa are different, two zeros appear, and when the shape, length and size are the same or close, only one zero appears. Generally, there are two right and left piriform fossa, but there are some people who have a piriform fossa only on one side as shown in FIG. 9 (a3). In this case, only one zero appears.

  From the above, the shape, length, and size of the speaker's piriform fossa can be obtained by using information about the zero point due to the piriform fossa.

  Note that the frequency band in which the effect of the piriform fossa appears on the voice spectrum is higher than the fixed telephone frequency band (4 kHz or less). Therefore, when this method is used for a telephone, it is necessary to target a mobile phone or an IP phone having a wider frequency band.

[Speaker registration and authentication by optimizing the shape parameters of the vocal tract model]
In the present invention, personal registration and authentication are performed by combining the individuality generation factors described above.

  Back-estimating the vocal tract cross-sectional area function from speech is one of the difficult tasks, because the conventional speech generation model does not take into account the resonance phenomenon between the piriform fossa and the laryngeal cavity described above. The complexity of the high frequency spectrum cannot be taken into the inverse estimation.

  FIG. 15 is a diagram for explaining a conceptual diagram of the speech generation model of the present invention.

  That is, in the present invention, sound generation is modeled as a result of the sound from the sound source being affected by the resonance of the main vocal tract and the resonance of the hypopharyngeal cavity as sound.

  By using a model as shown in FIG. 15, it is possible to incorporate the complexity of the high frequency spectrum into the inverse estimation.

  That is, in the conventional voice generation model, the voice is expressed by a linear combination of the sound source and the vocal tract, whereas the voice generation model in the present invention adds hypopharyngeal cavity resonance in addition to the sound source and the main vocal tract resonance. Based on this model, it is possible to accurately estimate the cross-sectional area function of the main vocal tract by removing the hypopharyngeal resonance component contained in the speech spectrum.

  Specifically, any of the following methods can be used to determine the personality parameter.

(First personality parameter determination method)
First, as a method for determining the first personality parameter, the shape parameter of the pharyngeal cavity / oral cavity and the shape parameter of the laryngeal cavity / piriform fossa are obtained from the spectrum of the input speech, and these are used as the individuality parameters as they are. Can be used.

(Second personality parameter determination method)
Alternatively, the above parameters are applied to the vocal tract model and optimized so that the input speech spectrum matches the transfer function calculated by the vocal tract model, and the parameters of the vocal tract model at that time are adopted as personality parameters. It is also possible to use a method to do this.

  Hereinafter, the method for determining these two personality parameters will be described in more detail.

[Details of First Personality Parameter Determination Method]
First, because the main vocal tract resonance and the hypopharyngeal cavity resonance are not in a linear relationship and interact with each other, in order to extract personality factors from speech, the transfer function obtained from the vocal tract model and the input speech spectrum Individuality parameters must be obtained by minimizing the error between them. A general error minimization method can be used for this optimization.

  Hereinafter, a method for optimizing the shape parameter of the vocal tract model will be described.

  FIG. 16 is a diagram showing an outer shape of a vocal tract model composed of each part of the vocal tract described in FIG. In this vocal tract model, the oral cavity and the laryngeal cavity are basically approximated by two acoustic tubes, and these two acoustic tubes are connected. Further, it is assumed that a piriform fossa represented by two cones and a laryngeal cavity approximated by the connection of two small acoustic tubes are connected to the bottom of the acoustic tube in the laryngeal cavity. It is assumed that sound from the sound source is input to the vocal tract model from the bottom of the laryngeal cavity.

  FIG. 17 is a diagram showing parameters for specifying the shape of the three-dimensional vocal tract model shown in FIG.

  As shown in FIG. 17, first, the acoustic tube corresponding to the oral cavity has a length Lor, a cylindrical shape with a radius Ror of the cross section, and the upper surface side (oral cavity side) is open. On the other hand, the acoustic tube corresponding to the pharyngeal cavity has a cylindrical shape with a length Lph and a cross-sectional radius Rph, and its upper surface is connected to the lower opening of the acoustic tube corresponding to the oral cavity. On the other hand, a small connected acoustic tube corresponding to the laryngeal cavity is connected to the lower surface of the acoustic tube corresponding to the pharyngeal cavity, and two conical tubes corresponding to the piriform fossa on both sides of the connected acoustic tube. Are linked. A connecting small acoustic tube corresponding to the laryngeal cavity includes a cylindrical first small acoustic tube having a cross-sectional radius Rla1 and a length Lla1 connected to the lower surface of the acoustic tube corresponding to the pharyngeal cavity, and the first small acoustic tube. A cylindrical second small acoustic tube having a cross-sectional radius Rla2 and a length Lla2 connected to the lower surface is provided, and the lower side of the second small acoustic tube is open. The sound of the sound source is input to the vocal tract model from the lower side of the second small acoustic tube.

(Functional configuration of voice authentication system)
FIG. 18 is a functional block diagram for explaining a functional configuration of the voice authentication system 1000 realized by software operating on the computer 100.

  The basic configuration of the voice authentication system shown in FIG. 18 is the same as that described in Non-Patent Document 1 described above, but as described below, in the present invention, a speaker model is used. The configuration is expressed by a combination of parameters of the vocal tract model shown in FIGS.

  Hereinafter, the functional configuration of the voice authentication system 1000 will be briefly described.

  Referring to FIG. 18, first, an input voice wave is subjected to spectrum conversion at a fine time of about 20 milliseconds in voice analysis unit 200. A method for expressing such a spectrum is not particularly limited, and for example, a cepstrum parameter can be used. Hereinafter, a parameter for expressing a speech spectrum such as a cepstrum parameter is referred to as a “speech parameter”.

  In the speaker model registration process (learning process), the switching unit 204 is switched to connect the process from the feature extraction unit 202 to the speaker model creation unit 206.

  Therefore, the feature extraction unit 202 is a parameter that expresses the feature of the speaker based on the time series of voice parameters, that is, each parameter that defines the shape of the above-described vocal tract model (hereinafter referred to as “voice tract model shape parameter”). Value).

  The speaker model creation unit 206 registers each speaker and the corresponding vocal tract model shape parameter in a storage device such as the hard disk 124 in association with each other.

  Subsequently, the threshold value setting unit 210 examines the range of fluctuation of each speaker's voice in advance from a plurality of input voices for the same speaker, and determines a threshold of an allowable limit for determining the voice of the speaker. To do.

  On the other hand, in the authentication process, the switching unit 204 is switched to connect the processing from the feature extraction unit 202 to the similarity calculation unit 220.

  Therefore, also in the authentication process, the vocal tract model shape parameter corresponding to the input voice is extracted by the processes of the voice analysis unit 200 and the feature extraction unit 202 in the same manner as in the learning process.

  The similarity calculation unit 220 compares the vocal tract model shape parameter extracted by the feature extraction unit 202 with each registered speaker model, calculates the degree of similarity, for example, the distance between them, If the degree of similarity is greater than a preset threshold value, the threshold value comparison unit 222 outputs an authentication result indicating that the voice is accepted and accepted, and if not, the threshold value is calculated as another person's voice. Judgment is made and an authentication result to be rejected or rejected is output.

  That is, in the present invention, as described above, the shape parameter of the vocal tract model is determined from the voice of the speaker, and authentication is performed using this. Determination of the shape parameter of the vocal tract model for a certain speaker's voice is performed by the following method.

(Determination of vocal tract model shape parameters for a speaker's voice)
FIG. 19 is a flowchart for explaining a procedure of a speaker model registration process for determining and registering a vocal tract model shape parameter for a speaker's voice by the first personality parameter determination method. .

  Referring to FIG. 19, when the speaker model registration process is started, the vocal tract length is first determined (step S100). That is, the voice analysis unit 200 performs spectrum analysis on the voice. Then, the feature extraction unit 202 determines the vocal tract length based on the number of poles that appear in a certain frequency band. At that time, reference is made to a standard vocal tract length obtained in advance by MRI measurement.

  Subsequently, the feature extraction unit 202 determines sexes by threshold processing from the vocal tract length and the fundamental frequency (step S102). Such a threshold value is experimentally determined in advance.

  Next, the feature extraction unit 202 determines the shape parameter of the laryngeal tube (step S104). That is, the parameters Lla1, Lla2, and Rla2 in FIG. 17 are determined with reference to standard values obtained by MRI measurement. Since the laryngeal canal can be regarded as a Helmholtz resonator, Rla1 can be determined from these three parameters and the fourth formant frequency.

  Next, the feature extraction unit 202 determines the shape parameter of the piriform fossa (step S106).

  At this time, as described above, the piriform fossa is approximated by a conical shape. In order to determine the shape parameter of the piriform fossa, for example, the relationship between the radius and height of the bottom surface of the conical shape and the frequency and bandwidth of the zero point created by the conical shape is previously set in a table. Next, the number of zeros appearing in a frequency band of 4 kHz or higher on the speech spectrum is specified, and if there is one zero, one cone is used, and if there are two zeros, two cones are used. Then, the radius (parameters Rpr1, Rpr2 in FIG. 17) and the height (Lpr1, Lpr2 in FIG. 17) and the height of the bottom of the cone are determined by reverse table lookup from the frequency and bandwidth of the zero point on the speech spectrum.

  Subsequently, the feature extraction unit 202 determines the shape parameters of the oral cavity and the pharyngeal cavity (step S108).

  Here, the vocal tract length is divided into two equal parts to create a two-section vocal tract model consisting of the pharyngeal cavity and the oral cavity, and the cross-sectional area of the pharyngeal cavity and the oral cavity is obtained by analysis of low-order formants.

  Further, the feature extraction unit 202 completes the vocal tract model by adding the hypopharyngeal cavity to the two-section vocal tract model (step S110). That is, the hypopharyngeal cavity is added to the vocal tract model obtained in step S108.

  Next, the speaker model creation unit 206 registers the shape parameter of the vocal tract model obtained in step S110 in the storage device as a personality parameter related to the speaker (step S112).

  This completes the speaker model registration process based on the first individuality parameter determination method.

[Details of second personality parameter determination method]
Next, speaker model registration processing based on the above-described second personality parameter determination method and second personality parameter determination method will be described.

(First Example of Second Personality Parameter Determination Method)
In the vocal tract model obtained in step S108 of FIG. 19, a four-section vocal tract model is created by further dividing the pharyngeal cavity and the oral cavity into two equal parts. As an initial value, each part divided into two equal parts has the same cross-sectional area as before. Then, the shape parameters of the four sections and the hypopharyngeal cavity are optimized so as to minimize the difference between the transfer function of the four section vocal tract model and the input spectrum. If necessary, the number of divisions can be further increased to use an 8-section vocal tract model. By the above optimization, the sectional area of each divided part is determined individually. The shape parameter of the vocal tract model obtained as a result is used as the individuality parameter for the speaker. Such a method for determining the shape parameter is performed both at the time of registration (learning) and at the time of authentication.

  The number of the acoustic tube corresponding to the pharyngeal cavity and the acoustic tube corresponding to the oral cavity is not limited to the above-described two divisions or four divisions, and the difference between the transfer function and the input spectrum obtained corresponding to the division result. As long as the degree of freedom can be minimized by calculation, it is possible to set the number of divisions for each of the acoustic tube corresponding to the pharyngeal cavity and the acoustic tube corresponding to the oral cavity.

(Second Example of Second Personality Parameter Determination Method)
FIG. 20 is a flowchart showing the procedure of the second example of the method for determining the second personality parameter.

  First, the feature extraction unit 202 obtains a vocal tract cross-sectional area function from speech (step S200). This is possible, for example, using so-called PARCOR analysis.

  Next, the feature extraction unit 202 obtains the shape parameter of the laryngeal cavity from the speech spectrum by the same method as step S104 in FIG. 19 (step S202).

  Subsequently, the feature extraction unit 202 obtains a portion corresponding to the oral cavity and the pharyngeal cavity of the vocal tract cross-sectional area function obtained in step S200 from the correlation between the low-order formant and the average oral cavity area, and the low-order formant and the average pharyngeal cavity area. Is corrected (step S204).

  The vocal tract cross-sectional area function obtained in step S200 does not include a branch pipe. Therefore, the feature extraction unit 202 obtains the shape parameter of the piriform fossa approximated by a cone from the speech spectrum by the same method as step S106 in FIG. 19 (step S206).

  Then, the feature extraction unit 202 creates a vocal tract model as shown in FIG. 17 using the shape parameters of the laryngeal cavity, oral cavity, pharyngeal cavity, and piriform fossa obtained as described above as initial values (step S208). Note that the number of divisions of the oral cavity and the pharyngeal cavity, that is, the accuracy, changes according to the analysis order of the PARCOR analysis.

  Subsequently, the feature extraction unit 202 calculates a transfer function of this vocal tract model, and corrects the shape parameter of the vocal tract model until the error between it and the speech spectrum is minimized (step S210).

  The shape parameter of the vocal tract model obtained in step S210 is registered as a personality parameter for the speaker in the storage device (step S212).

  Such a method for determining the shape parameter is performed not only at the time of registration (learning) but also at the time of authentication.

  As described above, when a speaker is registered in the voice authentication system as shown in FIG. 18, the personality of the registered speaker is determined from the voice by using the first or second personality parameter determination method. Determine and register parameters. When collating a speaker, the personality parameter of the speaker is determined from the input voice, and the speaker of the input voice is determined by collating with the personality parameter of the registered speaker.

  With such a configuration, in voice authentication, it is possible to increase the relevance with the physical characteristics of an individual and perform personal authentication, and it is possible to improve the accuracy of voice authentication.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

It is a conceptual diagram which shows an example of the voice authentication apparatus 1000 for enforcing the voice authentication method of this invention. It is a figure which shows the hardware constitutions of the computer 100 in a block diagram format. It is a figure which shows the correspondence of the characteristic on an audio | voice spectrum, and the site | part in a vocal tract. It is a conceptual diagram which shows the median cross-sectional view of an audio | voice production | generation system. It is a figure which shows the theoretical relationship with the cross-sectional area change of a pharyngeal cavity and an oral cavity in a two-section model of a vocal tract, and a low-order formant. It is a figure which shows the correlation of a pharyngeal cavity average area and a 1st formant frequency by the measured value by MRI. It is a figure which shows the correlation of an oral cavity average area and a 1st formant frequency by the measured value by MRI. It is drawing for demonstrating the shape of a laryngeal cavity. It is a figure which shows the three-dimensional shape of the hypopharyngeal cavity for three persons. It is a figure which shows the cross-sectional shape of each part of the hypopharyngeal space when each vowel is uttered about three persons shown in FIG. 9 as a parameter from the glottis. It is a figure which shows the audio | voice spectrum of vowel "e". It is a figure which shows the frequency of the 4th formant of speaker A-ko. It is a figure which shows the position of the piriform fossa in a hypopharyngeal cavity. It is a figure which shows the position of the zero point by the piriform fossa on the audio | voice spectrum of vowel "e". It is a figure for demonstrating the conceptual diagram of the audio | voice production | generation model of this invention. It is a figure which shows the external shape of the vocal tract model comprised from each part of the vocal tract demonstrated in FIG. It is a figure which shows each parameter for pinpointing the shape of the three-dimensional vocal tract model shown in FIG. 2 is a functional block diagram for explaining a functional configuration of a voice authentication system 1000 realized by software operating on a computer 100. FIG. It is a flowchart for demonstrating the procedure of the registration process of the speaker model for determining and registering the shape parameter of the vocal tract model with respect to a speaker's audio | voice. It is a flowchart which shows the procedure of the 2nd example of the determination method of a 2nd individuality parameter.

Explanation of symbols

  100 computer, 102 computer main body, 104 display, 106 FD drive, 108 CD-ROM drive, 110 keyboard, 112 mouse, 116 flexible disk, 118 CD-ROM, 120 CPU, 122 memory, 124 hard disk, 128 communication interface, 132 microphone 134 Speaker, 300 partner computer, 310 network, 1000 voice authentication system.

Claims (9)

A feature extraction unit for determining a shape parameter of a vocal tract model based on a voice input from a person to be authenticated,
The vocal tract model is
A first acoustic tube portion corresponding to the oral cavity;
A second acoustic tube portion coupled to the first acoustic tube portion and corresponding to the pharyngeal cavity;
A small acoustic tube connected to the bottom surface of the second acoustic tube portion and corresponding to the laryngeal cavity;
At least one conical tube coupled to the bottom surface of the second acoustic tube portion and corresponding to the piriform fossa;
At the time of learning, it further comprises storage means for storing the shape parameter determined by the feature extraction means in association with the person to be authenticated as a registered shape parameter,
The feature extraction means determines a shape parameter of the vocal tract model as an authentication shape parameter based on a voice input from a speaker at the time of authentication,
A voice authentication device further comprising similarity comparison means for comparing the authentication shape parameter with the registered shape parameter in order to specify whether or not the speaker is the registered person to be authenticated. The feature extraction means includes
An initial value determining means for determining an initial value of the shape parameter based on the voice input;
The voice authentication device according to claim 1, further comprising a correction unit that corrects the shape parameter so as to minimize a difference between a transfer function of the vocal tract model based on the initial value and an input spectrum of the voice input. The first acoustic tube portion includes a plurality of first acoustic tubes connected to each other,
The voice authentication device according to claim 2, wherein the second acoustic tube portion includes a plurality of second acoustic tubes connected to each other. At the time of learning, comprising the step of determining the shape parameter of the vocal tract model based on the voice input from the person to be authenticated,
The vocal tract model is
A first acoustic tube portion corresponding to the oral cavity;
A second acoustic tube portion coupled to the first acoustic tube portion and corresponding to the pharyngeal cavity;
A small acoustic tube connected to the bottom surface of the second acoustic tube portion and corresponding to the laryngeal cavity;
At least one conical tube coupled to the bottom surface of the second acoustic tube portion and corresponding to the piriform fossa;
Storing the shape parameter determined at the time of learning in a storage device in association with the authentication target person as a registered shape parameter;
Determining the shape parameter of the vocal tract model as an authentication shape parameter based on a voice input from a speaker at the time of authentication;
A voice authentication method further comprising: specifying whether or not the speaker is the registered person to be authenticated based on a comparison result between the authentication shape parameter and the registered shape parameter. Determining the shape parameter of the vocal tract model comprises:
Determining an initial value of the shape parameter based on the speech input;
The speech authentication method according to claim 4, further comprising the step of modifying the shape parameter so as to minimize a difference between a transfer function of the vocal tract model based on the initial value and an input spectrum of the speech input. The first acoustic tube portion includes a plurality of first acoustic tubes connected to each other,
The voice authentication method according to claim 5, wherein the second acoustic tube portion includes a plurality of second acoustic tubes connected to each other. A voice authentication program for causing a computer to execute voice authentication processing,
The voice authentication process includes:
At the time of learning, comprising the step of determining the shape parameter of the vocal tract model based on the voice input from the person to be authenticated,
The vocal tract model is
A first acoustic tube portion corresponding to the oral cavity;
A second acoustic tube portion coupled to the first acoustic tube portion and corresponding to the pharyngeal cavity;
A small acoustic tube connected to the bottom surface of the second acoustic tube portion and corresponding to the laryngeal cavity;
At least one conical tube coupled to the bottom surface of the second acoustic tube portion and corresponding to the piriform fossa;
Storing the shape parameter determined at the time of learning in a storage device in association with the authentication target person as a registered shape parameter;
Determining the shape parameter of the vocal tract model as an authentication shape parameter based on a voice input from a speaker at the time of authentication;
A voice authentication program further comprising: specifying whether or not the speaker is the registered person to be authenticated based on a comparison result between the authentication shape parameter and the registered shape parameter. Determining the shape parameter of the vocal tract model comprises:
Determining an initial value of the shape parameter based on the speech input;
The voice authentication program according to claim 7, further comprising the step of correcting the shape parameter so as to minimize a difference between a transfer function of the vocal tract model based on the initial value and an input spectrum of the voice input. The first acoustic tube portion includes a plurality of first acoustic tubes connected to each other,
The voice authentication program according to claim 8, wherein the second acoustic tube portion includes a plurality of second acoustic tubes connected to each other.