JP4255897B2 - Speaker recognition device - Google Patents

Description

  The present invention relates to a speaker recognition device.

  2. Description of the Related Art Conventionally, as a speaker recognition device, a text-dependent speaker recognition device or the like that holds, as a standard pattern, a voice pattern in which a specific speaker (user) utters a password or the like has been proposed. This speaker recognition device is used in ATM (Automatic Tellers Machine) and the like.

  In general, the speaker recognition apparatus calculates the distance after time axis matching between the input voice and the user's standard pattern at the time of recognition, and compares the value with a certain threshold value to determine whether or not the user is the user. At this time, in the speaker recognition device, even if the spoofer utters the same password as the password of the specific speaker registered in advance, the spoofer may be collated as the person himself, so the password is transferred to another person. It is important not to know.

  However, in the conventional speaker recognition device, the user utters the password toward the voice input microphone. For this reason, there is a problem that others around the user can easily hear the password, and information such as the password is easily known to others.

  Therefore, as a method for solving the problem, Patent Document 1 proposes a method for preventing an eavesdropping at the time of voice registration by generating a disturbing sound from a speaker to prevent a user's voice from being heard in the surroundings. At this time, the input sound input to the microphone includes interference sound as well as sound. If this input sound is used as it is for speaker recognition, the accuracy of speaker recognition will be reduced. Therefore, Patent Document 1 improves the speaker recognition accuracy by removing the interfering sound from the input sound using an adaptive filter. I am trying.

JP-A-9-127974

  However, in the method of Patent Document 1, it is difficult to completely remove only the disturbing sound from the waveform in which the disturbing sound is superimposed on the user's voice, and the speaker recognition accuracy is not sufficient. Furthermore, the method of Patent Document 1 has a large operational problem because of a large calculation load for removing the interference sound.

  An object of the present invention is to prevent eavesdropping at the time of voice registration by using a disturbing sound that prevents the user's voice from being heard in the surroundings, and to reduce the amount of calculation load and cost, thereby realizing highly accurate speaker recognition. It is.

The speaker recognition apparatus according to the present invention includes a disturbing sound generating means for generating a disturbing sound, a disturbing sound output unit for outputting the disturbing sound generated by the disturbing sound generating means to an external space, and a user's voice. And the interference sound output unit according to a ratio between a volume of the user's voice included in the input sound input from the sound input unit and a volume of the interference sound output to the external space. Interfering sound changing means for changing the volume of the interfering sound to be output, an operation unit that accepts an operation by the user, a storage unit that stores voice information of the user, and an operation of the user with respect to the operation unit And a state switching means for switching between a registration state for registering the user's voice and a collation state for collating the user's voice, and a voice feature amount based on the input sound. Feature quantity calculator to calculate Voice registration for registering the user's voice as the voice information in the storage unit using the voice feature quantity calculated by the feature quantity calculation means in the registration state switched by the state switching means The user's voice using the voice information stored in the storage unit and the voice feature quantity calculated by the feature quantity calculation means in the collation state switched by the state switching means. Voice collating means for performing collation.

  As a result, the voice feature quantity is calculated by the feature quantity calculation means based on the input sound input from the voice input unit under the interference sound, and when the user's voice is registered, the voice feature quantity is the voice information. Is registered in the storage unit by the voice registration unit, and when the user's voice is collated, the voice collation is performed by comparing the voice feature amount with the voice information registered in the storage unit. An adaptive filter or the like for removing sound is not required, and speaker recognition can be performed with high accuracy under interference sound.

  According to the present invention, it is possible to prevent eavesdropping at the time of voice registration by using a disturbing sound that prevents the user's voice from being heard in the surroundings, and to reduce the amount of calculation load and cost, thereby realizing highly accurate speaker recognition. Can do.

  A first embodiment of the present invention will be described with reference to FIGS.

  FIG. 1 is a block diagram showing a schematic configuration of a speaker recognition device 100 according to the present embodiment. The speaker recognition device 100 according to the present embodiment is an example that performs speaker recognition by a user uttering a specific password.

  As shown in FIG. 1, the speaker recognition device 100 includes a disturbing sound generating unit 1 that generates a disturbing sound that prevents a user's voice from being heard in the surroundings, and a disturbing sound output unit 2 that outputs the generated disturbing sound. A voice input unit 3 for inputting a user's voice, a feature amount calculation unit 4 for calculating a voice feature amount based on an input sound input to the voice input unit 3, and an operation unit 5 for accepting an operation by the user A state switching unit 6 that switches between a registration state for registering a user's voice and a collation state for collating the user's voice in accordance with a user's operation on the operation unit 5; 4, a voice registration unit 7 that registers a user's voice in a registered state using the voice feature amount calculated by 4, a standard pattern DB (database) 8 that stores voice information from the voice registration unit 7 as a standard pattern, and a standard Store in pattern DB8 The voice collation unit 9 that collates the user's voice in the collation state using the standard pattern and the voice feature amount calculated by the feature amount calculation unit 4, and the input sound input to the voice input unit 3 Based on the disturbing sound changing unit 10 or the like for changing the disturbing sound based on this.

  The disturbing sound is output from the disturbing sound output unit 2 and looped back and input to the sound input unit 3. Therefore, the input sound is generated by overlapping (mixing) the disturbing sound output from the disturbing sound output unit 2 and the user's voice and is input to the sound input unit 3.

  The interfering sound generation unit 1 generates interfering sounds such as music, beep sounds, synthesized sounds, and radio sounds as digital signals and sends them to the interfering sound output unit 2. This disturbing sound drowns out the user's voice, thereby preventing the user's voice from being heard in the surroundings. Such an interference sound generator 1 functions as an interference sound generator.

  The interfering sound output unit 2 includes a D / A converter that converts the generated digital signal into an analog signal, an amplifier that amplifies the converted analog signal, and a speaker that outputs the amplified analog signal as output sound (both shown in FIG. Not shown). Such an interference sound output unit 2 converts the digital signal of the interference sound generated by the interference sound generation unit 1 into an analog signal, amplifies it, and outputs it as an output sound to the outside.

  The voice input unit 3 includes a microphone for inputting a sound such as a user's voice as an analog signal, an amplifier for amplifying the input analog signal, and an A / D converter for converting the amplified analog signal into a digital signal ( Neither is shown). The user's voice is mainly input to such a voice input unit 3, but the disturbing sound output from the disturbing sound output unit 2 other than the user's voice also overlaps (mixes) with the user's voice. ) Is input. Therefore, the voice input unit 3 amplifies the input sound (synthetic sound) generated from the interference sound and the user's voice and converts it from an analog signal to a digital signal. The feature amount calculation unit 4 and the interference sound change unit 10 Send to.

  The feature quantity calculation unit 4 performs linear prediction analysis based on the input sound sent from the voice input unit 3 to obtain the voice feature quantity. Linear prediction analysis is a technique for obtaining a spectral envelope from input sound, and is a commonly known speech feature extraction method that reflects the vocal tract characteristics of the utterance mechanism (Kiyohiro Shikano (4 others) “Speech recognition system” Ohm Publication 1st edition (May 2001) See P1-P13). Such a feature quantity calculation unit 4 functions as a feature quantity calculation means.

  The operation unit 5 is an operation panel operated by a user, and includes a numeric keypad, a plurality of selection buttons (all not shown), and the like. For example, the user switches between a registration state and a collation state by pressing a selection button or the like.

  The state switching unit 6 switches the state of the speaker recognition device 100 to a registration state or a collation state in accordance with a user operation on the operation unit 5. In the registration state, the output of the feature amount calculation unit 4 passes to the speech registration unit 7, and in the recognition state, the output of the feature amount calculation unit 4 passes to the speech collation unit 9. That is, the state switching unit 6 transmits the voice feature amount calculated by the feature amount calculation unit 4 to the voice registration unit 7 in the registration state, and the voice feature amount calculated by the feature amount calculation unit 4 in the recognition state. It transmits to the voice collation part 9. Such a state switching unit 6 functions as a state switching unit.

  In the registration state, the voice registration unit 7 registers the voice feature amount (voice pattern) sent from the state switching unit 6 in the standard pattern DB 8 as a standard pattern that is voice information. At this time, the voice feature amount is a feature amount obtained from the input sound (input sound including the voice and the disturbing sound) when the user (specific speaker) utters the password under the disturbing sound. Such a voice registration unit 7 functions as a voice registration unit.

  The standard pattern DB 8 is a storage unit that stores a standard pattern that is audio information. As the standard pattern DB, for example, an HDD (hard disk) or a memory is used.

  The voice collation unit 9 calculates the distance after time axis matching between the voice feature amount (speech pattern) sent from the state switching unit 6 and the standard pattern stored in the standard pattern DB 8 in the collation state. Is compared with a certain threshold value. At this time, the voice feature amount sent from the state switching unit 6 is obtained from the input sound (the input sound including the voice and the disturbing sound) when the user (specific speaker) utters the password under the disturbing sound. Feature amount. Such a voice collating unit 9 functions as a voice collating unit.

  The interfering sound changing unit 10 can freely change the volume of the interfering sound at the interfering sound output unit 2 based on the input sound sent from the audio input unit 3 according to the ratio of the user's utterance volume and the interfering sound volume. To do. Such a disturbing sound changing unit 10 functions as a disturbing sound changing means.

  Next, voice registration in the voice registration unit 7 will be described with reference to FIGS. FIG. 2 is a schematic diagram showing a waveform of a voice at the time of voice registration in a quiet environment, FIG. 3 is a schematic diagram showing a waveform of a voice under a disturbing sound, and FIG. 4 is a diagram showing the sound under the disturbing sound after adaptive filter processing. It is a schematic diagram which shows the waveform of an audio | voice. FIG. 5 is a schematic diagram showing a waveform of a voice at the time of voice registration under an interfering sound, and FIG. 6 is a schematic diagram showing a waveform of a voice at the time of voice collation under the disturbing sound.

  The voice at the time of voice registration in a quiet environment has a waveform of x (t) as shown in FIG. Further, the sound under the disturbing sound has a waveform of x (t) + y (t) as shown in FIG. 3 where the disturbing sound is y (t). Furthermore, the waveform of the sound under the interference sound after the adaptive filter processing is a waveform of x (t) + y ′ (t) as shown in FIG. 4 where the interference sound after the adaptive filter processing is y ′ (t). become.

  Here, in the conventional technique, the speech x (t) as shown in FIG. 2 is compared with the speech x (t) + y ′ (t) as shown in FIG. Since there is a difference y ′ (t) between them, the speaker recognition accuracy is lowered. This is because the voice in a quiet environment is used as the voice at the time of voice registration.

  Therefore, in the present embodiment, the voice under the disturbing sound is used as the voice at the time of voice registration. The voice at the time of voice registration under the interference sound has a waveform of x (t) + y (t) as shown in FIG. Further, the voice at the time of voice collation under the interference sound has a waveform of x (t) + y (t) as shown in FIG. At this time, the difference between the waveform of x (t) + y (t) as shown in FIG. 5 and the waveform of x (t) + y (t) as shown in FIG. 6 takes into account the steadiness of y (t). Then it is very small and their waveforms are almost the same. Thereby, speaker recognition accuracy is improved.

  In such a configuration, the speaker recognition device 100 generates a disturbing sound by the disturbing sound generation unit 1 and outputs the generated disturbing sound to the outside by the disturbing sound output unit 2. The user utters a password toward the microphone of the voice input unit 3 while the disturbing sound is generated. At this time, the user's voice is input to the voice input unit 3 as an input sound, overlapping (mixed) with the disturbing sound output from the disturbing sound output unit 2.

  The speaker recognition device 100 obtains a speech feature value by the feature value calculation unit 4 based on the input sound (input sound including speech and interference sound) input from the speech input unit 3. Thereafter, when the registration state is selected by the state switching unit 6, the voice registration unit 7 registers the voice feature amount in the standard pattern DB 8 as a standard pattern. On the other hand, when the collation state is selected by the state switching unit 6, the voice collation unit 9 compares the voice feature amount with the standard pattern stored in the standard pattern DB 8, and performs voice collation. The user switches the registration state and the collation state of the speaker recognition device 100 by operating the operation unit 5.

  As described above, in the present embodiment, when the voice feature amount is calculated by the feature amount calculation unit 4 based on the input sound input from the voice input unit 3 under the interference sound, and the user's voice is registered, The voice feature quantity is registered as a standard pattern in the standard pattern DB 8 by the voice registration unit 7. When the user's voice is collated, the voice feature quantity is compared with the standard pattern registered in the standard pattern DB 8. Thus, voice collation is performed, so that an adaptive filter or the like for removing the interfering sound is not required, and speaker recognition can be performed with high accuracy under the interfering sound. Thereby, it is possible to prevent eavesdropping at the time of voice registration by using a disturbing sound that prevents the user's voice from being heard in the surroundings, reduce the calculation load and cost, and realize speaker recognition with high accuracy.

  Further, in the present embodiment, since the interference sound changing unit 10 which is an interference sound changing means for changing the interference sound based on the input sound is provided, the interference sound has, for example, the volume of the sound of the user's voice. Since it is adjusted and changed accordingly, it is not necessary to increase the volume of the disturbing sound more than necessary, and it is possible to prevent the surrounding people from feeling uncomfortable.

  A second embodiment of the present invention will be described with reference to FIG.

  FIG. 7 is a block diagram showing a schematic configuration of the speaker recognition apparatus 101 of the present embodiment. The present embodiment is an example of a speaker recognition device 101 that performs speaker recognition by a user uttering a specific password. In addition, the same part as 1st embodiment is shown with the same code | symbol, and the description is also abbreviate | omitted.

  As shown in FIG. 7, the speaker recognition device 101 includes an interference sound generator 1 that generates an interference sound that prevents the user's voice from being heard in the surroundings, and an interference sound output unit 2 that outputs the generated interference sound. A voice input unit 3 for inputting a user's voice, a feature amount calculation unit 4 for calculating a voice feature amount based on an input sound input to the voice input unit 3, and an operation unit 5 for accepting an operation by the user A state switching unit 6 that switches between a registration state for registering a user's voice and a collation state for collating the user's voice in accordance with a user's operation on the operation unit 5; 4, a voice registration unit 7 that registers a user's voice in a registered state using the voice feature amount calculated by 4, a standard pattern DB (database) 8 that stores voice information from the voice registration unit 7 as a standard pattern, and a standard Store in pattern DB8 The voice collation unit 9 that collates the user's voice in the collation state using the standard pattern and the voice feature amount calculated by the feature amount calculation unit 4, and the input sound input to the voice input unit 3 Based on the disturbing sound changing unit 10 or the like for changing the disturbing sound based on this.

  The disturbing sound is output from the disturbing sound output unit 2 and looped back and input to the sound input unit 3. Therefore, the input sound is generated by overlapping (mixing) the disturbing sound output from the disturbing sound output unit 2 and the user's voice and is input to the sound input unit 3.

  The interfering sound generation unit 1 generates interfering sounds such as music, beep sounds, synthesized sounds, and radio sounds as digital signals and sends them to the interfering sound output unit 2. This disturbing sound drowns out the user's voice, thereby preventing the user's voice from being heard in the surroundings. Such an interference sound generator 1 functions as an interference sound generator.

  The interfering sound output unit 2 includes a D / A converter that converts the generated digital signal into an analog signal, an amplifier that amplifies the converted analog signal, and a speaker that outputs the amplified analog signal as output sound (both shown in FIG. Not shown). Such an interference sound output unit 2 converts the digital signal of the interference sound generated by the interference sound generation unit 1 into an analog signal, amplifies it, and outputs it as an output sound to the outside.

  The voice input unit 3 includes a microphone for inputting a sound such as a user's voice as an analog signal, an amplifier for amplifying the input analog signal, and an A / D converter for converting the amplified analog signal into a digital signal ( Neither is shown). The user's voice is mainly input to such a voice input unit 3, but the disturbing sound output from the disturbing sound output unit 2 other than the user's voice also overlaps (mixes) with the user's voice. ) Is input. Therefore, the voice input unit 3 amplifies the input sound (synthetic sound) generated from the interference sound and the user's voice and converts it from an analog signal to a digital signal. The feature amount calculation unit 4 and the interference sound change unit 10 Send to.

  The feature quantity calculation unit 4 includes a noise removal unit that estimates and removes noise other than the interference sound from the input sound sent from the voice input unit 3, and removes the estimated noise from the input sound. Based on the above, a linear prediction analysis is performed to obtain a speech feature amount. Such a feature quantity calculation unit 4 functions as a feature quantity calculation means. In the present embodiment, the spectrum subtraction method is used as the noise estimation means, but the present invention is not limited to this. The method is a method in which the spectrum of the input sound sent from the voice input unit 3 is time-averaged for each frequency and subtracted sequentially (Boll SF: Suppression of Acoustic Noise in Speech Using Spectral Subtraction, IEEE Trans.ASSP-27). , P. 113-120, 1979). Linear prediction analysis is a technique for obtaining a spectral envelope from input sound, and is a generally known speech feature extraction method that reflects the vocal tract characteristics of the utterance mechanism (Kiyohiro Shikano (4 others) “Speech recognition system” "Ohm Publishing Co., Ltd. 1st edition (May 2001) See P1-P13).

  The operation unit 5 is an operation panel operated by a user, and includes a numeric keypad, a selection button (both not shown), and the like. For example, the user switches between a registration state and a collation state by pressing a selection button or the like. Further, the user inputs the report information such as a password and an ID number by operating the numeric keypad, and self-reports himself / herself.

  The state switching unit 6 switches the state of the speaker recognition device 101 to a registration state or a collation state in accordance with a user operation on the operation unit 5. In the registration state, the output of the feature amount calculation unit 4 passes to the speech registration unit 7, and in the recognition state, the output of the feature amount calculation unit 4 passes to the speech collation unit 9. That is, the state switching unit 6 transmits the voice feature amount calculated by the feature amount calculation unit 4 to the voice registration unit 7 in the registration state, and the voice feature amount calculated by the feature amount calculation unit 4 in the recognition state. It transmits to the voice collation part 9. Such a state switching unit 6 functions as a state switching unit.

  In the registration state, the voice registration unit 7 registers the voice feature amount (voice pattern) sent from the state switching unit 6 in the standard pattern DB 8 as a standard pattern that is voice information. At this time, the voice feature amount is a feature amount obtained from the input sound (input sound including the voice and the disturbing sound) when the user (specific speaker) utters the password under the disturbing sound. Such a voice registration unit 7 functions as a voice registration unit.

  The standard pattern DB 8 is a storage unit that stores a standard pattern that is audio information. As the standard pattern DB, for example, an HDD (hard disk) or a memory is used.

  The voice collation unit 9 calculates the distance after time axis matching between the voice feature amount (speech pattern) sent from the state switching unit 6 and the standard pattern stored in the standard pattern DB 8 in the collation state. Is compared with a certain threshold value. At this time, the voice feature amount sent from the state switching unit 6 is obtained from the input sound (the input sound including the voice and the disturbing sound) when the user (specific speaker) utters the password under the disturbing sound. Feature amount. Such a voice collating unit 9 functions as a voice collating unit.

  The interfering sound changing unit 10 can freely change the volume of the interfering sound at the interfering sound output unit 2 based on the input sound sent from the audio input unit 3 according to the ratio of the user's utterance volume and the interfering sound volume. To do. Such a disturbing sound changing unit 10 functions as a disturbing sound changing means.

  Here, the voice collation unit 9 performs speaker recognition based on the report information input by the operation unit 5 and sends the recognition result to the interference sound generation unit 1. The interfering sound generating unit 1 selects interfering sounds such as music, beep sound, synthesized speech, and radio sound according to the recognition result based on the report information by the voice collating unit 9, and generates the interfering sound as a digital signal. For example, the interfering sound generation unit 1 includes a storage unit (not shown) that stores a file or the like in which an interfering sound generated for each user (registrant) is set in advance. According to the recognition result based on, the interference sound generated from the file is selected, and the interference sound is generated as a digital signal.

  In such a configuration, when the user inputs report information such as a password and an ID number by the operation unit 5, the speaker recognition device 101 performs speaker recognition by the voice collation unit 9 based on the report information. Based on the recognition result, the disturbing sound generating unit 1 generates a disturbing sound, and the disturbing sound output unit 2 outputs the generated disturbing sound to the outside. The user utters a password toward the microphone of the voice input unit 3 while the disturbing sound is generated. At this time, the user's voice is input to the voice input unit 3 as an input sound, overlapping (mixed) with the disturbing sound output from the disturbing sound output unit 2.

  The speaker recognition apparatus 101 obtains a voice feature value by the feature value calculation unit 4 based on the input sound (input sound including voice and interference sound) input from the voice input unit 3. Thereafter, when the registration state is selected by the state switching unit 6, the voice registration unit 7 registers the voice feature amount in the standard pattern DB 8 as a standard pattern. On the other hand, when the collation state is selected by the state switching unit 6, the voice collation unit 9 compares the voice feature amount with the standard pattern and performs voice collation. The user switches the registration state and the collation state of the speaker recognition device 101 by operating the operation unit 5.

  As described above, in the present embodiment, when the voice feature amount is calculated by the feature amount calculation unit 4 based on the input sound input from the voice input unit 3 under the interference sound, and the user's voice is registered, The voice feature quantity is registered as a standard pattern in the standard pattern DB 8 by the voice registration unit 7. When the user's voice is collated, the voice feature quantity is compared with the standard pattern registered in the standard pattern DB 8. Thus, voice collation is performed, so that an adaptive filter or the like for removing the interfering sound is not required, and speaker recognition can be performed with high accuracy under the interfering sound. Thereby, it is possible to prevent eavesdropping at the time of voice registration by using a disturbing sound that prevents the user's voice from being heard in the surroundings, reduce the calculation load and cost, and realize speaker recognition with high accuracy.

  In particular, the feature quantity calculation unit 4 which is a feature quantity calculation means estimates and removes noise other than the interference sound from the input sound, and calculates the voice feature quantity based on the input sound, so that the accuracy is higher. Speaker recognition can be realized.

  Further, the operation unit 5 accepts an operation of inputting report information for self-reporting that the user is the user, and the speech collation unit 9 as a speaker recognition unit receives the report information input by the operation unit 5. The interference sound generation unit 1 that is the interference sound generation means changes the interference sound according to the result of the speaker recognition by the speech collation unit 9, and thus generates the interference for each user. The sound can be changed. As a result, the disturbing sound can be changed and output according to the user's preference or the like.

  The present invention is not limited to the specific hardware configuration shown in the embodiment as described above, and can be realized by software. That is, the function of each unit included in the speaker recognition devices 100 and 101 can be realized by software. In this case, the speaker recognition devices 100 and 101 include a CPU (not shown) that controls each unit in a centralized manner. The CPU is connected to a ROM storing a BIOS and various programs, a RAM (not shown) that stores various data in a rewritable manner, and the like. The CPU executes processing for realizing various functions based on a program stored in the ROM.

It is a block diagram which shows schematic structure of the speaker recognition apparatus of 1st embodiment of this invention. It is a schematic diagram which shows the waveform of the audio | voice at the time of the audio | voice registration in a quiet environment. It is a schematic diagram which shows the waveform of the audio | voice under interference sound. It is a schematic diagram which shows the waveform of the sound under the disturbance sound after an adaptive filter process. It is a schematic diagram which shows the waveform of the audio | voice at the time of the audio | voice registration under disturbance sound. It is a schematic diagram which shows the waveform of the audio | voice at the time of the audio | voice collation under interference sound. It is a block diagram which shows schematic structure of the speaker recognition apparatus of 2nd embodiment of this invention.

Explanation of symbols

1 Interference sound generation means (interference sound generator)
2 Interfering sound output section 3 Voice input section 4 Feature quantity calculation means (feature quantity calculation section)
5 Operation part 6 State switching means (state switching part)
7 Voice registration means (voice registration part)
8 storage unit (standard pattern DB)
9 Voice verification means (voice verification unit)
10 Interference sound change means (interference sound change part)
100 speaker recognition device 101 speaker recognition device

Claims (3)

An interference sound generating means for generating the interference sound;
An interference sound output unit for outputting the interference sound generated by the interference sound generation means to an external space;
A voice input unit for inputting the user's voice;
Volume of the interference sound to be output by the disturbing sound output unit by the ratio of the volume of the sound input unit included in the input sound input from the output volume of the audio of the user to the external space the interference sound Disturbing sound changing means for changing
An operation unit for receiving an operation by the user;
A storage unit for storing voice information of the user;
A state switching unit that switches between a registration state for registering the voice of the user and a collation state for collating the voice of the user in response to an operation of the user with respect to the operation unit;
Feature quantity calculating means for calculating a voice feature quantity based on the input sound;
Voice registration means for registering the voice of the user as the voice information in the storage unit using the voice feature quantity calculated by the feature quantity calculation means in the registration state switched by the state switching means; ,
In the collation state switched by the state switching unit, collation of the user's voice is performed using the voice information stored in the storage unit and the voice feature amount calculated by the feature amount calculation unit. Voice collation means to perform;
A speaker recognition device comprising: The feature amount calculating means estimates and removes noise other than the interference sound from the input sound, and calculates the speech feature amount based on the input sound.
The speaker recognition device according to claim 1. The operation unit accepts an operation of inputting report information for self-reporting that the user is the user,
Means for causing the disturbing sound generating means to generate a preset disturbing sound corresponding to a speaker specified by the input report information;
The speaker recognition device according to claim 1 or 2.

Images