JP6480124B2 - Biological detection device, biological detection method, and program - Google Patents

Description

  The present invention relates to a living body detection apparatus, a living body detection method, and a program, and more particularly to a living body detection technique using a voice.

  In recent years, biometric authentication technology based on human physical characteristics and behavior has been recognized as one of personal authentication technologies. In biometric authentication, physical features such as an individual's voice, fingerprint, retina, vein, etc. are used to authenticate the individual. In particular, speaker verification, which is a personal authentication technology based on voice characteristics, can be used to configure an authentication system using a general-purpose device such as a microphone, requires no practice for the speaker, and has a secretarial function application. Due to factors such as the lack of resistance to talking to machines with the spread, further spread is expected in the future.

  However, various techniques have also been discovered that can generate a voice that closely resembles the voice of the speaker, thereby spoofing the speaker and breaking the speaker verification system. For example, Non-Patent Document 1 shows that the speaker verification system can sometimes be deceived by speech synthesized based on text data. Non-Patent Document 2 shows that there is a case where a speaker verification system can be deceived by a voice quality conversion technique for converting a certain voice to resemble a specific speaker's voice.

  On the other hand, various techniques for detecting these spoofed voices have been proposed. Non-Patent Documents 3 and 4 disclose techniques for detecting spoofed speech based on differences in the manner of change in speech parameters between synthesized or voice-converted speech and natural speech.

Lindberg J.H. Et al., “Vulnerability in superverification-a study of technical impulse techniques”, Proc. European Conference on Speech Communication and Technology (Eurospeech), 1999 kinnunen, T .; , Wu, Z .; Z. Lee, K .; A. , Sedlak, F .; , Chng, E .; S. Li, H .; 2012. Vulnerability of spike verification systems against voice conversation attacks: The case of telephone specs, in: Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP). Satoh T. Et al., “A robust speaker verification system against the use of HMM-based speech synthesis system”, Proc. European Conference on Speech Communication and Technology (Eurospeech), 2001 Wu, Z. , Chng, E .; S. Li, H .; , Detecting converted speed and natural speed for anti-spoofing attack in speaker recognition, in: Proc. Interspeech 2012.

  The spoofing detection techniques disclosed in Non-Patent Documents 3 and 4 are all in the stage of extracting a feature value from a transmission point, that is, a speech waveform (FIG. 6). In these detection technologies, impersonation is detected by discriminating between spoofed speech and natural speech based on the characteristics of the speech waveform. However, the accuracy of speech synthesis technology and voice quality conversion technology has improved over time, and the difference from natural speech has been reduced. For this reason, the spoofing detection technique in the transmission point has a problem that improvement according to the technical improvement is constantly required.

  Furthermore, the impersonation detection technique in the transmission point is not a technique for detecting whether a speaker who is supposed to be there is really a living person (biological detection technique). Therefore, there is a problem that it is not a drastic solution to spoofing.

  The present invention has been made to solve such problems, and an object of the present invention is to provide a living body detection apparatus, a living body detection method, and a program capable of detecting a living body by voice.

  Other problems and novel features will become apparent from the description of the specification and the accompanying drawings.

  The living body detection apparatus according to the present invention includes a voice acquisition unit that acquires a voice of a speaker, a pop noise detection unit that detects pop noise from the voice, and the speaker that is a living body based on the detection result of the pop noise. And a determination unit for determining whether or not.

  Further, the living body detection method according to the present invention includes a voice acquisition step of acquiring a speaker's voice, a pop noise detection step of detecting pop noise from the voice, and the speaker based on the detection result of the pop noise. And a determination step of determining whether or not is a living body.

  The program according to the present invention is a program for causing a computer to execute the above method.

  According to the present invention, it is possible to provide a living body detection apparatus, a living body detection method, and a program capable of detecting a living body by voice.

It is a figure which shows the structure of the biological detection apparatus 100 concerning embodiment of this invention. It is a figure which shows operation | movement of the biological detection apparatus 100 concerning embodiment of this invention. It is a figure which shows operation | movement of the biological detection apparatus 100 concerning embodiment of this invention. It is a figure which shows the experimental result concerning Embodiment 3 of this invention. It is a figure which shows operation | movement of the biological detection apparatus 100 concerning embodiment of this invention. It is a figure which shows the structure of a general speaker recognition system. It is a figure which shows the example of the audio | voice acquisition part 110 concerning embodiment of this invention. It is a figure which shows the example of the audio | voice acquisition part 110 concerning embodiment of this invention. It is a figure which shows the example of the audio | voice acquisition part 110 concerning embodiment of this invention. It is a figure which shows the process of the pop noise detection part 130 in embodiment of this invention. It is a figure which shows the process of the pop noise detection part 130 in embodiment of this invention. It is a figure which shows the process of the pop noise detection part 130 in embodiment of this invention. It is a figure which shows the process of the pop noise detection part 130 in embodiment of this invention. It is a figure which shows the process of the pop noise detection part 130 in embodiment of this invention. It is a figure which shows the process of the determination part 150 in embodiment of this invention. It is a figure which shows the example of the microphone used by embodiment of this invention. It is a figure which shows the experimental result concerning Embodiment 5 of this invention.

  First, in order to facilitate understanding of the present invention, characteristics of the living body detection method according to the present invention will be described in comparison with a conventional spoofing detection method.

  FIG. 6 is a diagram showing an outline of a general speaker recognition system. Unlike the conventional detection of spoofing in a transmission point, the technique according to the present invention performs a living body detection in a stage of acquiring sound by a microphone point, that is, a microphone.

  In order to perform living body detection by voice, it is necessary to pay attention to features that are possible for a living person and that cannot be reproduced by reproduced voice from an apparatus or the like. Therefore, the inventor has focused on how to breathe into the microphone when a person speaks. When a person speaks, he exhales at the same time as the sound. And the microphone can pick up not only the voice but also the breath. It is known that when a microphone picks up a lot of breath, a unique noise called pop noise is generated. Pop noise is noise for speech, but from a different perspective, it is also information that proves that the person uttering the speech is a living person. This is because pop noise cannot be reproduced by a speaker in principle.

  In view of such knowledge, the present invention is characterized in that pop noise is detected from speech acquired at a Microphone point, and the living body of the speaker is detected using the detection result. Hereinafter, specific embodiments to which the present invention is applied will be described in detail with reference to the drawings.

  Next, a basic configuration of the living body detection apparatus 100 according to the embodiment of the present invention will be described with reference to FIG.

  The living body detection apparatus 100 includes a sound acquisition unit 110, a pop noise detection unit 130, and a determination unit 150. The living body detection device 100 typically includes a storage device that stores a control program and the like, a control device that executes various processes based on the control program, an input / output device that inputs and outputs information with an external device, and the like. An information processing apparatus. The voice acquisition unit 110, the pop noise detection unit 130, and the determination unit 150 are realized as logical processing means using the hardware and the control program described above.

  The voice acquisition unit 110 has a function of acquiring a voice uttered by a speaker. Typically, a microphone that converts sound into an analog signal and a conversion unit that converts the analog signal into a digital signal and outputs the digital signal to the pop noise detection unit 130 are included.

  The sound acquisition unit 110 may be configured with a single microphone or may include a plurality of microphones. If there is a single microphone, the microphone will record both the voice and breath of the speaker. When there are a plurality of microphones, one microphone (referred to as a first microphone) records both the voice and breath of the speaker. The other microphone (referred to as the second microphone) is configured to record only the voice of the speaker as much as possible and to suppress recording by attenuating the breath.

  As a configuration example when using a plurality of microphones, for example, a method of providing a microphone cover or a pop filter only on the second microphone, and a method of installing the first microphone and the second microphone in different spaces (microphones) Array).

  The pop noise detection unit 130 performs a process of detecting a pop noise component generated when the microphone picks up the breath emitted by the speaker from the audio signal (digital signal) output from the audio acquisition unit 110. The pop noise detection unit 130 may include a separation module 131, a feature amount module 132, a phoneme alignment module 133, and an identification module 134.

  The separation module 131 performs a process of separating an audio signal suspected of being a pop noise component from the audio signal output from the audio acquisition unit 110. That is, the separation module 131 is a module that performs audio signal processing. When the sound acquisition unit 110 is configured by a single microphone, the separation module 131 separates a sound signal suspected of being a pop noise component from one sound signal in which a sound component and a pop noise component are mixed. Specific separation methods include, for example, a low-pass filter (smoothing) and a method of re-recording the sound signal that is output from the speaker and propagated in the air. In carrying out the present invention, any of these methods or other equivalent methods can be adopted as necessary.

  On the other hand, when the sound acquisition unit 110 is configured by a plurality of microphones, the separation module 131 includes an output signal (referred to as first sound) of the first microphone including both the sound component and the pop noise component, and the sound component. Is used to separate the audio signal suspected of being a pop noise component using an output signal (referred to as second audio) of a second microphone in which the pop noise component is suppressed. Specific examples of the separation method include a linear filter and spectral subtraction. In carrying out the present invention, any of these methods or other equivalent methods can be adopted as necessary.

  The feature conversion module 132 performs a process of extracting a feature amount from the audio signal suspected of being a pop noise component separated by the separation module 131. The audio signal is time-series information, but it is difficult to handle the time-series information in the subsequent identification module 134. For this reason, the feature quantification module 132 converts the time series information into information based on the frequency. Examples of specific feature amount techniques include MFCC (Mel-Frequency Cepstrum Coefficients), extension of MFCC to a low frequency range, and extension of MFCC to a low frequency range, as well as a technique of dimensional compression. In carrying out the present invention, any of these methods or other equivalent methods can be adopted as necessary. It should be noted that the feature amount module 132 is arbitrarily adopted, and even if it does not exist, the minimum object of the present invention can be achieved.

  The phoneme alignment module 133 performs a process of dividing the featured speech signal into phoneme units (segments). By performing this process, it is possible to verify the correspondence between phonemes and locations where pop noise occurs, so it is possible to eliminate the effects of wind noise and other similar components of pop noise and to perform more accurate pop noise recognition There is. As a specific method, for example, a method using a phoneme alignment function provided by an external automatic speech recognition (ASR) or the like, an acoustic of a reference speech whose segment boundary is known in advance and an input speech are used. There is a method of specifying a segment of input speech by comparison (DTW: Dynamic Time Warping). In carrying out the present invention, any of these methods or other equivalent methods can be adopted as necessary. Note that in the identification module 134 in the subsequent stage, when using a model that itself can recognize the phoneme boundary, such as an HMM (Hidden Markov Model), a configuration corresponding to the phoneme alignment module 133 need not be provided. Note that the use of the phoneme alignment module 133 is arbitrary, and even if it does not exist, the minimum object of the present invention can be achieved.

  The pop noise identification module 134 inputs the feature amount extracted by the feature amount module 132 or the segment amount feature amount segmented by the phoneme alignment module 133 and performs a process of identifying whether or not the feature amount is pop noise. As specific identification methods, for example, use of a statistical processing model such as GMM (Gaussian mixture model), a machine learning processing model such as SVM (Support vector machine), an HMM, and a linear identification model can be considered. When performing phoneme alignment, the identification module 134 identifies the presence or absence of pop noise on a segment basis. On the other hand, when phoneme alignment is not performed, the identification module 134 identifies the presence or absence of a pop noise component in units of sentences that are not divided into segments. In any case, the identification module 134 can detect when a pop noise component, which is an abnormal value, is input, by learning in advance the feature amount of an audio component that does not include the pop noise component. . In carrying out the present invention, any of these methods or other equivalent methods can be adopted as necessary.

  The above-described separation module 131, feature amount module 132, phoneme alignment module 133, and identification module 134 can be used in any combination.

  Based on the pop noise detection result by the pop noise detection unit 130, the determination unit 150 performs a process of determining whether or not the voice speaker acquired by the voice acquisition unit 110 is a living person.

<Embodiment 1>
The first embodiment is a configuration example of the present invention when the sound acquisition unit 110 includes a plurality of microphones, and one microphone includes a microphone cover or a pop filter. In the present embodiment, the phoneme alignment module 133 is not employed.

  The sound acquisition unit 110 includes a plurality of microphones. One microphone (first microphone) aims to pick up the voice and breath of the speaker as much as possible. Typically, the microphone does not include a microphone cover or a pop filter. Alternatively, the first microphone may be provided with a perforated windscreen so that pop noise accompanying speech can be acquired while reducing the influence of surroundings such as wind (FIG. 16).

  The other microphone (second microphone) suppresses the pickup of the speaker's breath and aims to pick up the speaker's voice as much as possible. Typically, the microphone includes a microphone cover and a pop filter. In general, a microphone cover and a pop filter are made of a sponge, a net, or the like, and by covering the vibration film with these, it is possible to reduce pop noise that is generated when breath directly vibrates the vibration film.

  7 to 9 show configuration examples of the voice acquisition unit 110. FIG. 7 is a configuration example using a stereo microphone, FIG. 8 is a configuration example using a plurality of condenser microphones, and FIG. 9 is a configuration example using a plurality of headset microphones. Each of these includes at least one first microphone without a microphone cover and one or more second microphones with a microphone cover.

  The voice acquisition unit 110 may include three or more microphones such as an array microphone. In this case, some microphones in the array microphone are handled as the first microphone without providing the microphone cover, and the remaining microphones are handled as the second microphone with the microphone cover provided.

  Next, the operation of the living body detection apparatus 100 according to the first embodiment will be described using the flowcharts of FIGS. 2 and 3.

S101: Voice acquisition The voice acquisition unit 110 acquires the voice of a speaker. In the present embodiment, the voice acquisition unit 110 includes a first microphone and a second microphone, and each microphone simultaneously acquires the same voice uttered by a speaker.

  The conversion unit of the sound acquisition unit 110 converts the sound acquired by the first microphone and the second microphone into different digital signals. Here, the sound data derived from the first microphone is referred to as first sound, and the sound data derived from the second microphone is referred to as second sound. The conversion unit outputs the first sound and the second sound to the pop noise detection unit 130.

S104: Pop Noise Detection First, the separation module 131 of the pop noise detection unit 130 compares the first sound and the second sound, separates the frequency included only in the first sound, and outputs it as a difference signal. . As a result, an audio signal that may be pop noise can be extracted. For extraction of the difference signal, for example, a known method such as a linear filter, spectral subtraction, Independent component analysis, Independent Vector Analysis, and Blind source separation can be used. Further, a technique for detecting a low frequency fluctuation by a low-pass filter, a blind signal source separation technique, or the like may be used.

  FIG. 10 shows a state in which the first sound and the second sound are superimposed. FIG. 11 shows a difference signal between the first sound and the second sound. Here, the amplitude component (unit: dB) protruding around 1.0 (unit: second) on the horizontal axis (time axis) is considered to be an abnormal value for the speech signal corresponding to the voice of the speaker. There may be noise.

  Next, the feature quantity module 132 of the pop noise detection unit 130 extracts a feature quantity from the difference signal. In other words, a feature amount that is information based on a frequency is obtained from a difference signal that is time-series information. The extraction of the feature amount can be performed by a known method such as MFCC (Mel-Frequency Cepstrum Coefficients). Alternatively, since the MFCC is designed to be able to discriminate speech of 70 Hz or higher, a method of extending the discriminable region to a frequency region of less than 70 Hz can also be adopted. This is because pop noise generally has a lower frequency than normal speech. Furthermore, when the frequency domain is expanded in this way, the dimension of the number of features increases and the discrimination performance generally decreases. Therefore, a technique that uses dimensional compression of feature quantities may be employed. Thereby, in MFCC, identification performance can be maintained, corresponding to a low frequency region.

  The identification module 134 of the pop noise detection unit 130 inputs the extracted feature amount of the difference signal to the recognizer. Here, the identification module 134 may divide the feature amount of the difference signal into arbitrary segments and input the segment to the identifier. Note that the segment here does not need to be a segment of phonemes as handled by the phoneme alignment module 133.

  For example, the pop noise detection unit 130 may include two discriminators learned in advance using a sentence including pop noise and a sentence not including pop noise. The discrimination module 134 obtains the likelihood as an output by inputting the feature quantity of the difference signal to these discriminators. Here, if there is a significant difference between the likelihood that the discriminator learned with a sentence including pop noise and the likelihood that the discriminator learned with a sentence not including pop noise outputs, the difference signal is It can be determined that pop noise is included. An appropriate threshold value can be determined in advance as to how much the difference between the two likelihoods is to determine that pop noise exists.

  As the discriminator, for example, a well-known configuration such as GMM, HMM, etc., as well as SVM, which is a two-class pattern discriminator, can be adopted as appropriate. Note that, when the determination by the discriminator is performed, a step is required in which the discriminator learns in advance a model for each sentence including pop noise and each sentence not including pop noise. This process will be described later.

  Further, since the detection of the pop noise component is similar in framework to the voice interval detection (VAD), any known method used in the field of VAD as well as the field of speaker verification may be used.

  When the identification module 134 of the pop noise detection unit 130 identifies the differential signal as pop noise, the identification module 134 outputs to the determination unit 150 that pop noise has been detected. The determination unit 150 determines whether the speaker is a living person based on the detection result of the pop noise. Typically, when pop noise is detected, it is determined that the speaker is a living body, and the process proceeds to S105. On the other hand, if pop noise is not detected, it is determined that the speaker is not a living body, and the process proceeds to S106.

S105: Speaker verification When it is determined that the speaker is a living body, the living body detection apparatus 100 can shift to a speaker verification phase using an arbitrary method. Since various methods are known for speaker verification, detailed description thereof is omitted here. Preferably, the second voice having a relatively small pop noise component can be used for speaker verification.

S106: Rejected as a spoofed voice When it is determined that the speaker is not a living body, the voice acquired by the voice acquisition unit 110 is not generated by a person, but has a high probability of being a voice generated by synthesis or voice quality conversion, for example. Therefore, the living body detection apparatus 100 determines that this is a spoofed voice and rejects it without performing speaker verification. That is, error processing, end processing, and the like are performed.

  Here, the pre-learning process required when performing the pop noise determination by the discriminator will be described with reference to FIG. Here, the case where SVM is used as a learning device and a discriminator will be described as an example.

S201: Voice acquisition As in S101, the voice acquisition unit 110 acquires the voice of the speaker.

S202: Learning Pop Noise / Non-Pop Noise Model First, as in S104, the separation module 131 of the pop noise detection unit 130 compares the first sound and the second sound and includes only the first sound. Are separated and output as a differential signal. Next, the feature amount conversion module 132 converts the difference signal into a feature amount. Then, the identification module 134 inputs both the feature quantity of the difference signal and the teacher signal indicating that it is pop noise to one learning device. In addition, the feature amount conversion module 132 converts the second sound that does not include the pop noise component into a feature amount. Then, the identification module 134 inputs both the feature amount of the second sound and the teacher signal indicating that it is non-pop noise to the other learning device. In other words, in the present embodiment, a total of two learners are generated, one each for determining the likelihood of each of pop noise and non-pop noise.

  When SVM or GMM is used as a learning device, it is preferable that a segment corresponding to the pop noise component is cut out in advance from the feature quantity of the difference signal, and the extracted segment is input to the learning device. On the other hand, when the HMM is used as a learning device, the model itself has a function of automatically recognizing a phoneme boundary, and thus the above-described clipping process is not particularly necessary.

  By repeating the processes according to S201 to S202 a plurality of times, models of pop noise and non-pop noise speech are formed in the learning device. As a result, a discriminator that outputs a corresponding model in response to the input of the feature amount of the difference signal is formed.

  According to the present embodiment, the voice acquisition unit 110 acquires information indispensable for living body detection in the Microphone point. As a result, it is possible to realize living body detection that was not possible with impersonation detection at the transmission point.

  Further, according to the present embodiment, the determination unit 150 performs living body detection based on the pop noise detection result by the pop noise detection unit 130. Since pop noise is a phenomenon that cannot be reproduced in principle by a speaker, it is possible to realize biometric authentication that is robust against speaker misrepresentation.

  Further, according to the present embodiment, the voice acquisition unit 110 outputs a voice signal including pop noise and a voice signal not including the pop noise by using a plurality of microphones. As a result, the pop noise detection unit 130 can efficiently separate the pop noise component by applying a known separation technique.

<Embodiment 2>
The second embodiment is a configuration example in which the sound acquisition unit 110 includes a plurality of microphones, and these microphones are arranged in different spaces. The rest of the configuration is the same as in the first embodiment.

  In the present embodiment, one of the plurality of microphones included in the voice acquisition unit 110 (the first microphone) is intended to pick up pop noise along with the voice of the speaker. It is placed in a position where it can be directly applied. For example, the first microphone is arranged at a position facing the speaker. The other microphone (second microphone) is placed at a position where it is difficult for the speaker to breathe directly in order to suppress pop noise and to pick up only the speaker's voice as much as possible. . For example, the second microphone is arranged on the side of the speaker or at a position away from the first microphone. In the present embodiment, the second microphone is not necessarily provided with a microphone cover or a pop filter.

  For example, when an array microphone is used as the voice acquisition unit 110, a microphone located near the speaker is treated as a first microphone, and a microphone located farther from the speaker than the first microphone is treated as a second microphone. Can do.

  According to the present embodiment, the audio acquisition unit 110 can realize the same pop noise detection process as that of the first embodiment without using a microphone cover or a pop filter.

<Embodiment 3>
The third embodiment is a configuration example in which the sound acquisition unit 110 includes a single microphone and a low-pass filter is employed as the separation module 131. The rest of the configuration is the same as in the first embodiment.

  The voice acquisition unit 110 in the third embodiment is configured by a single microphone. The purpose of this microphone is to pick up as much of the speaker's voice and pop noise as possible. Therefore, it is preferable that the microphone is not provided with a microphone cover or a pop filter, which is disposed at a position where the speaker can easily breathe. Alternatively, the first microphone may be provided with a perforated windscreen so that only the pop noise accompanying the speech can be acquired while reducing the influence of the surroundings such as wind.

  In addition, the separation module 131 in the third embodiment includes a low cut filter and a low pass filter.

  Next, a characteristic operation of the third embodiment will be described.

S101: Voice acquisition The voice acquisition unit 110 acquires the voice of a speaker. In the present embodiment, voice acquisition unit 110 is a single microphone. The conversion unit of the sound acquisition unit 110 converts the sound acquired by the microphone into a digital signal and outputs the digital signal to the pop noise detection unit 130.

S104: Pop Noise Detection The separation module 131 of the pop noise detection unit 130 uses the low cut filter to extract only the audio component from the audio signal output from the audio acquisition unit 110. Pop noise detection unit 130 uses the extracted audio component as equivalent to the second audio in the first embodiment. Further, the pop noise detection unit 130 extracts only a noise component from the audio signal output from the audio acquisition unit 110 using a low-pass filter. The pop noise detection unit 130 can use the extracted noise component as an equivalent of the difference signal in the first embodiment.

  The low-pass filter smoothes the speech waveform, but has a property of leaving an abnormal value included in the speech signal. Therefore, the pop noise component can be revealed from the audio signal including the pop noise by the low-pass filter.

  And the pop noise detection part 130 determines whether the extracted noise component is a pop noise using a discriminator similarly to Embodiment 1. FIG. That is, the feature quantity module 132 converts the noise component extracted by the low-pass filter instead of the difference signal in the first embodiment into a feature quantity, and preferably divides it into several segments. Then, the identification module 134 inputs the feature amount of the noise component to the two discriminators, and determines the presence / absence of pop noise based on the output likelihood difference.

S105 to S106:
The operation is the same as in the first embodiment.

The operation in the pre-learning process of the present embodiment is as follows.
S201: Voice acquisition As in S101, the voice acquisition unit 110 acquires the voice of the speaker.

S202: Learning of Pop Noise / Non-Pop Noise Model First, as in S104, the separation module 131 extracts a noise component from an audio signal using a low-pass filter. Next, the feature quantity conversion module 132 converts the noise signal into a feature quantity. Then, the identification module 134 inputs both the feature quantity of the noise signal and the teacher signal indicating that it is pop noise to one learning device. Further, the separation module 131 extracts a sound component from the sound signal using a low cut filter. The feature amount module 132 converts the sound component into a feature amount. Then, the identification module 134 inputs both the feature amount of the speech component and the teacher signal indicating that it is non-pop noise to the other learning device. The other operations are the same as those in the first embodiment.

(Experimental result)
FIG. 4 shows the results of a demonstration experiment using the configuration of the third embodiment. The inventor tried to detect pop noise for 10 speakers from F001 to F010. As a result, for all speakers, the likelihood that the classifier that has learned the sentence containing pop noise (“voice with pop noise”) outputs the likelihood that the classifier that has learned the sentence that does not contain pop noise outputs. (“Speech without pop noise”). That is, it has been found that it is possible to detect with high accuracy that the input voice of the speaker includes pop noise.

  According to the present embodiment, the voice acquisition unit 110 is configured with a single microphone, and the pop noise detection unit 130 extracts a voice component and a noise component using a low cut filter and a low pass filter. Thereby, a living body detection is realizable with a simple structure compared with the case where a plurality of microphones are used.

<Embodiment 4>
The fourth embodiment is a configuration example in which the voice acquisition unit 110 includes a single microphone for directly acquiring a speaker's voice, and a speaker and a microphone are employed as the separation module 131. The rest of the configuration is the same as in the first embodiment.

  The voice acquisition unit 110 according to the fourth embodiment includes a single microphone (first microphone) for directly acquiring a speaker's voice. The first microphone aims to pick up the voice of the speaker and pop noise as much as possible. Therefore, it is preferable that the microphone is not provided with a microphone cover or a pop filter, which is disposed at a position where the speaker can easily breathe. Alternatively, the first microphone may be provided with a perforated windscreen so that only the pop noise accompanying the speech can be acquired while reducing the influence of the surroundings such as wind.

  In addition, the separation module 131 of the pop noise detection unit 130 includes a speaker that outputs recorded sound from the first microphone, and a second microphone that records sound that is output from the speaker and propagated in the air. Note that since the sound recording by the second microphone may be performed inside the living body detection apparatus 100, the speaker and the second microphone do not need to be exposed to the speaker.

  Next, a characteristic operation of the fourth embodiment will be described.

S101: Voice acquisition The first microphone of the voice acquisition unit 110 acquires the voice of the speaker. In the present embodiment, only the first microphone directly records the voice of the speaker. The voice acquired by the first microphone includes pop noise generated by a speaker.

  Next, the speaker of the separation module 131 reproduces the voice of the speaker recorded by the first microphone. Then, the second microphone records the sound reproduced from the speaker. Since the sound output from the speaker does not generate pop noise in principle, the sound acquired by the second microphone does not include the pop noise component.

  The sound acquisition unit 110 and the separation module 131 convert the sound acquired by the first microphone and the second microphone into different digital signals. Here, the sound data derived from the first microphone is referred to as first sound, and the sound data derived from the second microphone is referred to as second sound. The conversion unit outputs the first sound and the second sound to the pop noise detection unit 130.

  Since the operations in S104 to S106 and S201 to S202 are the same as those in the first embodiment, detailed description thereof is omitted.

  According to the present embodiment, the voice acquisition unit 110 includes one microphone exposed to the speaker, a speaker provided in the apparatus, and a second microphone. Thereby, compared with the case where a plurality of microphones exposed to the speaker are used, the living body detection device 100 can be configured with a simple appearance.

<Embodiment 5>
In the first to fourth embodiments, the example in which the living body detection apparatus 100 performs living body detection by determining whether or not pop noise is included in the audio signal has been described. However, in the methods of the first to fourth embodiments, when a sound similar to pop noise is input due to factors other than the speaker's breath such as wind, for example, the living body detection device 100 mistakes this as pop noise generated by the living body. May end up. Therefore, the fifth embodiment presents a living body detection method that is particularly robust against the influence of wind or the like as compared with the first to fourth embodiments.

  Usually, it is known that pop noise mainly occurs in some specific consonants such as plosives. Therefore, in this embodiment, robustness is enhanced by inspecting whether or not pop noise is generated at an appropriate place.

  The living body detection apparatus 100 according to the present embodiment is characterized in that a phoneme alignment module 133 is included in the pop noise detection unit 130 in addition to the constituent elements of the living body detection apparatus 100 according to the first to fourth embodiments. The rest of the configuration is the same as in Embodiments 1 to 4 unless otherwise specified.

  The voice acquisition unit 110 outputs a voice signal to the phoneme alignment module 133. The separation module 131 of the pop noise detection unit 130 separates the pop noise component from the audio signal. Then, the feature amount module 132 converts the pop noise component into a feature amount.

  The phoneme alignment module 133 receives a voice signal from the voice acquisition unit 110 and performs a process of identifying a phoneme in the voice signal. Then, a segment obtained by dividing the audio signal by the time length of each phoneme is defined, and the feature amount of the pop noise component is divided into segments.

  The identification module 134 detects pop noise on a segment basis. That is, the segmented feature amount is input to the discriminator, and the presence / absence of pop noise is identified for each segment.

  The determination unit 150 verifies whether or not the correspondence between the pop noise detection result of each segment by the pop noise detection unit 130 and the phoneme is correct.

  Next, the operation of the living body detection apparatus 100 will be described using the flowchart of FIG.

S301: Voice acquisition As in S101, the voice acquisition unit 110 acquires the voice of the speaker.

S302: Phoneme Identification The phoneme alignment module 133 extracts phonemes from the voice signal input from the voice acquisition unit 110. Preferably, the second sound with relatively little pop noise can be used.

  Here, the phoneme is a minimum unit of speech having linguistic value. For example, individual vowels and consonants correspond to phonemes. Recognition of phonemes from a speech signal can be performed using a known method such as HMM.

  The phoneme alignment module 133 defines a plurality of segments corresponding to the recognized time length of each phoneme for the speech signal (FIG. 12). A segment can typically be defined by the start point (time) and time length of each phoneme. Moreover, the phoneme alignment module 133 assigns a phoneme name as a label to each segment. For example, the phoneme alignment module 133 can realize this by creating a record in which a phoneme name, a phoneme start point, and a phoneme time length are associated with each other, and holding it in a storage area (not shown). (FIG. 13).

S303: Pop Noise Detection The separation module 131 of the pop noise detection unit 130 separates and extracts the pop noise component from the audio signal. Subsequently, the feature amount module 132 converts the pop noise component into a feature amount. Then, the phoneme alignment module 133 divides the pop noise component that has been featured into segments. The identification module 134 performs a pop noise detection process for each segment.

  For example, consider a case where a segment as shown in FIG. 13 is defined for an audio signal. First, since a segment (ID = 1) having a start point of 00:00:00 and a time length of 00: 00: 00: 12 is defined, the phoneme alignment module 133 determines the feature amount of the pop noise component. Among them, a feature amount of a region corresponding to this segment, that is, an area from 00:00:00 to 00: 00: 00: 12 is cut out. Then, the identification module 134 inputs the extracted feature amount to the classifier, and obtains the detection result of the pop noise.

  Subsequently, the phoneme alignment module 133 sets the time width corresponding to the next segment (ID = 2) among the feature quantities of the pop noise component, that is, from 00:00:00 to 12: 00: 00: 20. Extract region features. Similarly, the pop noise is detected using the discriminator. Similarly, the pop noise detector 130 detects pop noise for all segments. The pop noise detection unit 130 stores the detection result for each segment for which detection was attempted (FIG. 14).

S304: Validity verification of relationship between pop noise and phoneme The determination unit 150 verifies the validity of the detection result of the pop noise obtained in S303. As described above, it is known that pop noise mainly occurs in some specific consonants such as plosives (eg, “p”). In the present embodiment, it is assumed that living body detection apparatus 100 holds in advance a pattern table in which such phonemes are associated with the possibility of occurrence of pop noise (FIG. 15).

  The determination unit 150 determines the correspondence between the phoneme and the pop noise detection result (FIG. 14) in the pop noise detection result obtained in S303, and the correspondence between the phoneme given in advance and the possibility of pop noise generation (FIG. 15). ) And verify whether they are consistent. For example, in FIG. 14, the pop noise is “non-detection” for the phoneme “t” with the segment ID = 1. On the other hand, in FIG. 15, for the phoneme “t”, the possibility of occurrence of pop noise is defined as “none”. Therefore, the determination unit 150 determines that the determination result of the segment ID = 1 is appropriate.

  Also, pop noise is “detected” for the phoneme “p” of the segment ID = 5. On the other hand, in FIG. 15, the possibility of pop noise is defined as “present” for the phoneme “p”. Therefore, the determination unit 150 determines that the determination result of the segment ID = 5 is also valid. Similarly, the determination unit 150 verifies the validity of each of the pop noise detection results obtained in S303.

  If the validity of all segments is confirmed, the determination unit 150 determines that the speaker is a living body, and the process proceeds to S305. On the other hand, if the validity cannot be confirmed in all segments, it is determined that the speaker is not a living body, and the process proceeds to S306.

  Here, the determination unit 150 can appropriately adopt any criterion other than the above as a determination criterion for living body detection. For example, the speaker may be determined to be a living body when the proportion of segments whose validity has been verified exceeds a predetermined threshold.

S305: Speaker verification When it is determined that the speaker is a living body, the living body detection apparatus 100 can shift to a speaker verification phase using an arbitrary method.

S306: Rejected as a spoofed voice When it is determined that the speaker is not a living body, the voice acquired by the voice acquisition unit 110 is not generated by a person, but has a high probability of being a voice generated by synthesis or voice quality conversion, for example. Therefore, the living body detection apparatus 100 determines that this is a spoofed voice and rejects it without performing speaker verification. That is, error processing, end processing, and the like are performed.

  According to the present embodiment, the phoneme alignment module 133 classifies the speech signal into phoneme levels, and the pop noise detection unit 130 detects the occurrence of pop noise at the phoneme level. Then, the determination unit 150 verifies whether the pop noise is generated at an appropriate position (phoneme). Thereby, it is possible to realize more robust living body detection that can eliminate the influence of pop noise derived from, for example, wind, which does not depend on human speech.

(Experimental result)
The inventor conducted a living body detection demonstration experiment using the configuration of the fifth embodiment. In the experiment, the voice of the speaker and the voice output from the speaker were input to the voice acquisition unit 110, respectively. As the sound acquisition unit 110, a first microphone without a pop filter and a second microphone having a pop filter are employed. A condenser microphone, a stereo microphone, and a headset microphone were used as the first microphone and the second microphone. Further, the feature amount module 132 is not employed, and the difference signal between the first sound and the second sound is input to the discriminator as it is. Note that the discriminator has already learned the differential signal when pop noise occurs. The identification module 134 determines that pop noise has been detected when the likelihood that the identifier is output is equal to or greater than a predetermined threshold.

  FIG. 17 shows the results of this demonstration experiment. In either case of using a condenser microphone or a stereo microphone, when a speaker's live voice was input, it was detected as a living body with a high probability (100%). In addition, when a sound output from a speaker is input, it is determined that it is not a living body with a high probability (100%). Even when a headset microphone was used, a generally good living body detection rate of 86.9% was observed.

<Other embodiments>
Note that the present invention is not limited to the above-described embodiment, and can be changed as appropriate without departing from the spirit of the present invention.

  For example, in the fifth embodiment, an example (FIG. 15) in which presence / absence of occurrence of pop noise is associated in advance with one phoneme has been shown. However, pop noise is generated for a plurality of continuous phonemes. The presence or absence of possibility may be associated.

  In the above-described embodiment, the present invention has been mainly described as a hardware configuration. However, the present invention is not limited to this, and a CPU (Central Processing Unit) executes a computer program for arbitrary processing. Can also be realized. In this case, the computer program can be stored and provided to the computer using various types of non-transitory computer readable media. Non-transitory computer readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (for example, flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (for example, magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / W, semiconductor memory (for example, mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, RAM (Random Access Memory)). The program may also be supplied to the computer by various types of transitory computer readable media. Examples of transitory computer readable media include electrical signals, optical signals, and electromagnetic waves. The temporary computer-readable medium can supply the program to the computer via a wired communication path such as an electric wire and an optical fiber, or a wireless communication path.

DESCRIPTION OF SYMBOLS 100 Living body detection apparatus 110 Audio | voice acquisition part 130 Pop noise detection part 131 Separation module 132 Feature-quantization module 133 Phoneme alignment module 134 Identification module 150 Determination part

Claims (11)

A voice acquisition unit that acquires the voice of the speaker;
A pop noise detector for detecting pop noise from the voice;
A living body detection apparatus comprising: a determination unit that determines whether or not the speaker is a living body based on a detection result of the pop noise. The sound acquisition unit acquires a first microphone that acquires the sound as a first sound, and a second microphone that acquires the sound in which the pop noise is reduced compared to the first sound as a second sound. A microphone, and
The living body detection device according to claim 1, wherein the pop noise detection unit detects the pop noise using a difference between the first sound and the second sound. The living body detection device according to claim 2, wherein the second microphone includes a cover for reducing the pop noise. The living body detection device according to claim 2, wherein the first microphone and the second microphone are arranged at different positions in a certain space. The sound acquisition unit acquires, as a second sound, a first microphone that acquires the sound as a first sound, a speaker that outputs the first sound, and the first sound output from the speaker. A second microphone that
The living body detection device according to claim 1, wherein the pop noise detection unit detects the pop noise using a difference between the first sound and the second sound. The voice acquisition unit includes a microphone that acquires the voice,
The living body detection device according to claim 1, wherein the pop noise detection unit detects the pop noise from the voice using a low-pass filter. Further comprising a phoneme identifying unit for identifying a plurality of phonemes from the speech;
The living body detection device according to any one of claims 1 to 6, wherein the pop noise detection unit detects the pop noise for each of a plurality of segments obtained by dividing the voice by a time length of the phoneme. The determination unit further determines the validity of the correspondence between the phoneme and the detection result of the pop noise for each of the segments, and whether the speaker is a living body based on the determination result of the validity The living body detection device according to claim 7 which judges whether or not. The living body detection device according to any one of claims 1 to 8, wherein the pop noise detection unit includes a discriminator that has previously learned models of pop noise and non-pop noise. A voice acquisition step for acquiring the voice of the speaker;
A pop noise detecting step for detecting pop noise from the voice;
And a determination step of determining whether or not the speaker is a living body based on the detection result of the pop noise.   The program for making a computer perform the method of Claim 10.