KR20170140680A - Method and apparatus for generating speaker rocognition model by machine learning - Google Patents

Abstract

The present invention relates to a method and an apparatus for generating a speaker recognition model through machine learning, the method according to an embodiment of the present invention comprising: receiving a voice of a speaker to be learned through machine learning; applying, to the voice of the speaker, a voice conversion filter for converting a voice in accordance with a characteristic of a voice recognition device and an environment in which the voice of the speaker can be received; and acquiring a speaker recognition model by machine learning of the voice of the speaker and the voice of the speaker to which the voice conversion filter has been applied. The method and apparatus for generating a speaker recognition model are capable of identifying the speaker through machine learning even if the voice of the speaker is converted in accordance with the environment and characteristic of the voice recognition device.

Description

TECHNICAL FIELD [0001] The present invention relates to a method and apparatus for generating a speaker recognition model by machine learning,

The present invention relates to a method and apparatus for generating a speaker recognition model through machine learning, and more particularly, to a method and apparatus for generating a speaker recognition model by using a speaker capable of distinguishing a speaker through a speaker's voice without affecting a change in environment for recognizing the speaker's speech, The present invention relates to a method and apparatus for a speaker recognition model through machine learning.

A sound wave is a wave which is generated by vibrating in the form of a longitudinal wave due to a change in pressure of the air having a uniform vibration. Sound waves are divided into audible sound waves, which are the frequency range from 20 Hz to 20,000 Hz, low sound waves which are in the region below 20 Hz, and ultrasonic waves, which are in the region above 20,000 Hz.

The sound wave has a ripple nature and is used in various technical fields. Sound waves are a key technology in underwater navigation devices, location devices, ultrasonic inspection devices, and physiotherapy devices for physical exploration in the presence of water. Sound genetics, which activates or deactivates neuronal cell genes in the brain, and sound holograms that lift or move objects are also techniques using sound waves.

 Furthermore, recently, a speaker recognition technology for recognizing a speaker through speech recognition technology and a processing terminal for recognizing a sound wave of voice and operating a processing terminal, that is, a computer, a smart phone, a tablet PC, etc., has been developed. However, speech and speech recognition technologies have limitations in that they utilize sound waves that are easily distorted by various conditions. Specifically, depending on the environment in which the voice and speaker recognition technology has received the sound waves and the change in the sound waves with respect to the sound according to the speaker's state, the result value processed by the voice recognition technology is different from the intended result . For example, the speech recognition apparatus is influenced by a change in the environment in which the speech is received, so that it is possible to classify the environment, not to classify the speaker, or to fail the speaker recognition due to the speech changed according to the speech utterance habits, health, There is an inconvenience to do.

Thus, there is a need for a method that can process modified sound waves to recognize speech and speech.

[Related Technical Literature]

A method and apparatus for speech recognition (Patent Document 10-2004-0061659)

SUMMARY OF THE INVENTION The present invention has been made in an effort to solve the above problems, and an object of the present invention is to provide a speaker recognition apparatus and a speaker recognition apparatus, which can recognize an environment element that recognizes a speaker's voice, A method and an apparatus for generating a model are provided.

Another object to be solved by the present invention is to provide a speaker recognition system using machine learning capable of classifying a speaker through simulation based on the results of simulations and simulations that apply various environment factors and speaker's speech state to one voice, And to provide a method and an apparatus for generating the same.

The problems of the present invention are not limited to the above-mentioned problems, and other problems not mentioned can be clearly understood by those skilled in the art from the following description.

According to an aspect of the present invention, there is provided a method of generating a speaker recognition model through machine learning, the method comprising: receiving a speech of a speaker to be machine-learned; A step of applying a speech conversion filter for converting a speech to a speech of a speaker according to characteristics of the environment and the speech recognition apparatus and a step of obtaining a speaker recognition model by machine learning a speech of a speaker to which a speech and speech conversion filter of the speaker is applied .

According to another aspect of the present invention, the step of acquiring the speaker recognition model comprises the steps of performing machine learning of the speech of the speaker applying the speech and speech conversion filter of the speaker to obtain a first weight of the speech conversion filter and a second weight of the speech of the speaker Step < / RTI >

According to still another aspect of the present invention, the step of acquiring the first weight and the second weight may be a step of acquiring a first weight through the machine learning and a second weight higher than the first weight.

According to still another aspect of the present invention, the step of applying the voice conversion filter may be a step of applying a voice conversion filter to the speaker's voice in which at least one of the equalizer, the reverb, the speed of the voice and the height of the voice are changed.

According to still another aspect of the present invention, the step of acquiring the speaker recognition model includes the steps of generating a database storing the speech of the speaker to which the speech conversion filter is applied, And learning the speech characteristics of the speaker.

According to still another aspect of the present invention, the method may further include recognizing the speaker by applying the speaker recognition model to the speaker's voice when receiving the speaker's voice.

According to an aspect of the present invention, there is provided an apparatus for generating a speaker recognition model through machine learning, including a communication unit for receiving a voice of a speaker to be machine-operated, an environment in which a voice of the speaker can be received, And a processor for applying a voice conversion filter for converting a voice according to a characteristic of the recognizing device to the voice of the speaker, wherein the processor obtains a speaker recognition model by machine-learning the voice of the speaker to which the voice and voice conversion filter of the speaker is applied .

According to another aspect of the present invention, the processor may apply a speech conversion filter to each of the speech of the speaker by changing at least one of an equalizer, a reverb, a speed of speech, and a pitch of speech.

According to another aspect of the present invention, there is provided a speech recognition apparatus, further comprising a storage unit for storing speech of a speaker to which a speech conversion filter is applied, wherein the storage unit generates a database of speech of the speaker, It is possible to learn the speech characteristic of the speaker which changes in accordance with the conversion filter.

According to another aspect of the present invention, the processor may machine the speech of the speaker applying the speech and speech translation filter of the speaker to obtain a first weight of the speech translation filter and a second weight of the speech of the speaker.

According to another aspect of the present invention, a processor generates a speaker recognition model based on a result of machine learning, and when receiving a speaker's voice, the speaker can recognize the speaker by applying the speaker recognition model to the speaker's voice.

According to an aspect of the present invention, there is provided a computer-readable recording medium storing instructions for providing a method of generating a speaker recognition model through machine learning, the method comprising: receiving a voice of a speaker to be machine- A voice conversion filter for converting a voice is applied to a voice of a speaker according to the environment in which the voice of the speaker can be received and the characteristics of the voice recognition device and the voice of the speaker to which the voice and voice conversion filter of the speaker is applied is machine- Thereby acquiring a recognition model.

The details of other embodiments are included in the detailed description and drawings.

The present invention has an effect of recognizing a speaker by recognizing a speaker's voice irrespective of an environment element that recognizes the speaker's voice, that is, a speech utterance location, a speech recognition apparatus, and a speaker's speech state.

The present invention has the effect of distinguishing a speaker through machine learning which is performed based on simulation and simulation results of applying various environmental factors and speaker's speech state to one voice.

The effects according to the present invention are not limited by the contents exemplified above, and more various effects are included in the specification.

1 shows a schematic configuration of an apparatus for generating a speaker recognition model according to an embodiment of the present invention.
FIG. 2 illustrates a procedure for generating a speaker recognition model through machine learning according to a speaker recognition model generation method according to an embodiment of the present invention.
FIG. 3 is an exemplary diagram illustrating a process of classifying speakers by applying weights obtained through machine learning according to an embodiment of the present invention.
4A and 4B show graphs of speech of a speaker converted according to a speech conversion filter according to an embodiment of the present invention.
FIGS. 5A to 5C are exemplary views illustrating a process of classifying a speaker by machine learning based on a speech graph of a speaker to which a speech conversion filter is applied according to an embodiment of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention and the manner of achieving them will become apparent with reference to the embodiments described in detail below with reference to the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Is provided to fully convey the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims.

The shapes, sizes, ratios, angles, numbers, and the like disclosed in the drawings for describing the embodiments of the present invention are illustrative, and thus the present invention is not limited thereto. In the following description, well-known functions or constructions are not described in detail since they would obscure the invention in unnecessary detail. Where the terms 'comprises', 'having', 'done', and the like are used herein, other parts may be added as long as '~ only' is not used. Unless the context clearly dictates otherwise, including the plural unless the context clearly dictates otherwise.

In interpreting the constituent elements, it is construed to include the error range even if there is no separate description.

Although the first, second, etc. are used to describe various components, these components are not limited by these terms. These terms are used only to distinguish one component from another. Therefore, the first component mentioned below may be the second component within the technical spirit of the present invention.

Like reference numerals refer to like elements throughout the specification unless otherwise specified.

It is to be understood that each of the features of the various embodiments of the present invention may be combined or combined with each other partially or entirely and technically various interlocking and driving is possible as will be appreciated by those skilled in the art, It may be possible to cooperate with each other in association.

1 shows a schematic configuration of an apparatus for generating a speaker recognition model according to an embodiment of the present invention.

Referring to FIG. 1, a speaker recognition model generation apparatus 100 includes a communication unit 110, a processor 120, and a storage unit 130.

The communication unit 110 can receive a voice of a speaker to be machine-operated from the voice recognition apparatus. At this time, the speech recognition device may be, but is not limited to, a microphone, a tablet PC with a built-in microphone or a microphone, a smart phone, a laptop, a desktop, and the like. The voice of the speaker received by the communication unit 110 may be a voice having an extension such as m4a, acc, mkv, mp3, but is not limited thereto.

The communication unit 110 may transmit the voice of the received speaker to the processor 120 and may receive the voice of the speaker to which the voice conversion filter is applied from the processor 120 in order to apply the voice conversion filter to the voice of the speaker.

The processor 120 applies a speech conversion filter to the speech of the speaker. In addition, the processor 120 obtains a speaker recognition model by machine-running the speaker's voice using the speaker's voice and voice conversion filter.

The storage unit 130 may store data necessary for machine learning, such as a speaker's voice, a voice conversion filter, and a voice of a speaker to which a voice conversion filter is applied. Furthermore, the storage unit 130 generates a database of the speaker's voice. Also, the storage unit 130 may store the speaker recognition model generated by the processor 120.

The communication unit 110 may transmit the speech of the speaker to be stored in the storage unit 130 to the storage unit 130, and the speech of the speaker to which the speech conversion filter is applied. The communication unit 110 may receive the voice of the speaker stored in the storage unit 130, the voice conversion filter, and the voice of the speaker to which the voice conversion filter is applied from the storage unit 130.

Accordingly, the speaker recognition model generation apparatus 100 receives a voice of a speaker to be machine-operated from the communication unit 110, applies various voice conversion filters to the speaker's voice through the processor 140, Lt; / RTI > In addition, the speaker recognition model generation apparatus 100 generates a database of the speaker's voice through the storage unit 130, and machine-runs the speaker's voice to which the speaker's voice and voice conversion filter is applied through the processor 140 A speaker recognition model is created. Therefore, the speaker recognition model generation apparatus 100 can correctly recognize the speaker's voice without being influenced by the environment in which the speaker's speech is received and the characteristics of the speech recognition apparatus, based on the speaker recognition model.

Although the speaker recognition model generating apparatus 100 and the sub-components shown in FIG. 1 are each shown as one module, the present invention is not limited thereto, and one module may be divided into two or more modules, or two or more modules may be integrated into one module .

Hereinafter, FIG. 2 will be referred to for a more detailed description of a method for generating a speaker recognition model through machine learning in the speaker recognition model generation apparatus 100. [

FIG. 2 illustrates a procedure for generating a speaker recognition model through machine learning according to a speaker recognition model generation method according to an embodiment of the present invention. For convenience of explanation, the components will be described with reference to FIG. 1 and reference numerals.

The communication unit 110 receives a voice of a speaker to be machine-operated (S210).

Specifically, the communication unit 110 receives a voice for one speaker to be machine-operated. Here, the machine learning is a method in which the processing apparatus learns itself on the basis of data given from the outside on the basis of a learning model. For example, when the apparatus 100 for recognizing a speaker recognition model identifies a speaker on the basis of a speaker's voice To learn by themselves. Although embodiments have been described in which the voice of the speaker received by the communication unit 110 is a voice of one speaker, in some embodiments, there may be a plurality of voice for a plurality of speakers and a plurality of voices for a plurality of speakers.

Subsequently, the processor 120 applies the speech conversion filter for converting the speech according to the environment in which the speech of the speaker can be received and the characteristics of the speech recognition apparatus, to the speech of the speaker (S220).

The processor 120 applies a voice conversion filter to the voice of the speaker received by the communication unit 110. [ Here, the voice conversion filter is a filter that can convert at least one of the equalizer, the reverb, the speed of voice, and the height of voice to the voice of the speaker. That is, the processor 120 can perform an equalizer conversion, a reverb conversion, a speed change of voice, and a voice height change through a voice conversion filter. Here, the equalizer conversion means to convert the overall frequency of the speech of the speaker. In addition, the reverb conversion means to convert the thickness and depth of the speaker's voice. The speed conversion of speech means a fast conversion of the speech output of the speaker. In addition, the conversion of the voice to the height means conversion of the voice of the speaker to the tone.

Accordingly, the processor 120 may apply a voice conversion filter to the speaker's voice in which at least one of the equalizer, the reverberation, the speed of the voice, and the height of the voice is changed. The voice conversion filter is applied to the voice of the speaker to convert it according to the features of the various environments and the voice recognition device with the voice of one speaker without receiving the voice of the speaker to convert according to the features of the voice recognition device The speaker's voice can be simulated. A graph in which a voice conversion filter is applied to a voice of a specific speaker will be described later with reference to FIG.

In addition, the storage unit 130 may generate a database storing speech of the speaker to which the speech conversion filter is applied.

Specifically, the storage unit 130 may generate a database storing speech of the speaker whose at least one of the equalizer, the reverb, the speed of the voice, and the height of the voice is changed. The storage unit 130 may also include a voice of a speaker to which the voice conversion filter is not applied to generate a database.

Next, the processor 120 obtains a speech recognition model by machine-running the speech of the speaker to which the speech and speech conversion filter of the speaker is applied (S230).

In some embodiments, the processor 120 may machine-run to obtain a first weight for the speech translation filter and a second weight for speech of the speaker higher than the first weight based on the speech of the speaker. In this case, the first weight is a weight obtained by expressing the speech of the speaker to which the speech conversion filter is applied, that is, the speech of the converted speaker, smaller than the speech of the speaker to which the speech is not converted. The second weight is a weight obtained by recognizing the speaker's voice more clearly than the voice conversion filter. At this time, the processor 120 applies a voice conversion filter to the voice of the speaker to express the voice of the converted speaker according to various environments and features of the voice recognition apparatus.

For example, the processor 120 may obtain a second weight higher than the first weight, i.e., a first weight of the speech transform filter is 0.1, and a second weight of the speech of the speaker is 0.9. Accordingly, the processor 120 can obtain a low weight from the voice to which the voice conversion filter is applied through the machine learning, that is, the voice of the converted speaker, and obtain a high weight from the voice of the untranslated speaker, The speaker can recognize the speaker even if it receives the voice of the converted speaker. The process of dividing the speaker's voice by applying the first weight and the second weight, which are obtained through specific machine learning, will be described with reference to FIG. Thus, the processor 120 creates a speaker recognition model based on the results of the machine learning. When the communication unit 110 receives the speaker's voice, the processor 120 applies the speaker recognition model to the speaker's voice to recognize the speaker even if the speaker's voice is converted according to the characteristics of the environment and the voice recognition apparatus.

Accordingly, the speaker recognition model generation apparatus 100 generates a speaker recognition model by machine-learning based on the speaker's voice and the voice of the various speakers converted, so that even if the speaker's voice is converted according to the environment or the characteristics of the apparatus, Can be recognized.

FIG. 3 is an exemplary diagram illustrating a process of classifying speakers by applying weights obtained through machine learning according to an embodiment of the present invention.

Referring to FIG. 3, the speech conversion filter graph 310 includes a speech conversion filter graph 311 applied to speech of a speaker and a weighted speech conversion filter graph 312 to which weights obtained through machine learning are applied.

Referring to FIG. 3, a speech conversion filter graph 311 applied to speech of a speaker is a speech graph of a speaker to which a speech conversion filter is applied. The speech conversion filter graph 311 applied to the speech of the speaker indicates the variation of the speech characteristics of the speaker with the change of the frequency to which the speech conversion filter is applied. The weighted speech conversion filter graph 312 is a speech conversion filter graph of the speaker to which the weight is applied to the speech conversion filter graph 311 applied to the speech of the speaker. For example, when a weight of 0.1 obtained through machine learning is applied to the speech conversion filter graph 311 applied to the speaker's voice, the size value of the speech conversion filter graph 311 applied to the speaker's voice for each frequency 0.1 times.

Referring to FIG. 3, the speaker's voice graph 320 includes a speaker's original voice graph 321 and a weighted speech graph 322 to which weights obtained through machine learning are applied.

Referring to FIG. 3, the original voice graph 321 of the speaker is a voice graph of a speaker received for machine learning. The original speech graph 321 of the speaker indicates the variation of the speech characteristic of the speaker according to the change of the frequency. The weighted speech graph 322 is the original speech graph of the speaker to which the weight is applied to the original speech graph 321 of the speaker. For example, when a weight of 0.9 is applied to the original speech graph 321 of the speaker, the size of the original speech graph 321 of the speaker is 0.9 times for each frequency.

Referring to FIG. 3, the speaker recognition graph 330 is a speech graph of a speaker recognized through the sum of the weighted speech conversion filter graph 312 and the weighted speech graph 322. The speaker recognition graph 330 is similar to the original speech graph 321 of the speaker because the weights applied to the weighted speech graph 322 are greater than the weights applied to the weighted speech translation filter graph 312. [ That is, the speech conversion according to the environment of the speaker or the characteristic of the speech recognition apparatus is reduced and the speech characteristic of the original speaker is strengthened, so that the speaker recognition model generation apparatus 100 is less affected by the speech conversion, Voice can be recognized.

Accordingly, the speaker recognition model generation apparatus 100 applies a low weight to the speech of the converted speaker according to the environment and the speech recognition apparatus through machine learning, applies a high weight to the speech of the speaker, So that the speaker can be more accurately distinguished even if the speaker's voice is converted.

4A and 4B show graphs of speech of a speaker converted according to a speech conversion filter according to an embodiment of the present invention.

Referring to FIG. 4A, the equalizer filter 410, which is a speech conversion filter, adjusts the dB of the speech included in a specific frequency band in the speech 420 of the speaker. The equalizer filter 410 may reduce the size of the voice by adjusting the dB of the voice to -, and may increase the size of the voice by adjusting it to +. For example, the equalizer filter 410 may be set to 0 dB (411) for the 55 Hz, 77 Hz, and 110 Hz bands. That is, the 55 Hz, 77 Hz, and 110 Hz bands are set to 0 dB (411) in the speech 420 of the speaker, so that the sizes of the voices in the 55 Hz, 77 Hz, and 110 Hz bands are not changed. In addition, the equalizer filter 410 can set the audio of the 220 Hz, 311 Hz, 440 Hz, and 622 Hz bands to-dB (412). Therefore, the size of voice in the 55 Hz, 77 Hz, and 110 Hz bands is reduced. In addition, the equalizer filter 410 can set the sound of the 14 kHz band and the 20 kHz band to + dB (413), thereby increasing the size of the sound of the 14 kHz and 20 kHz band.

By applying the equalizer filter 410 to the speaker's voice 420, a first voice 421 of the speaker to which the voice conversion filter is applied and a second voice 440 of the speaker to which the voice conversion filter is applied can be generated. For example, when-dB (412) is set in the entire frequency band of the speaker's voice 420, the first voice 421 of the speaker having the voice size smaller than the speaker's voice 420 can be generated have. In addition, when + dB 413 is set to the entire frequency band of the speaker's voice 420, a second voice 422 of the speaker having a larger voice size than the speaker's voice 420 can be generated.

Accordingly, the speaker recognition model generation apparatus generates and simulates variously converted voices from the speaker's voice 420 by variously adjusting the value of the equalizer filter 410, and performs machine learning based on the simulated various voices do.

Referring to FIG. 4B, the speech speed and the speech level filter 430, which are speech conversion filters, adjust the speed of the speech 440 of the speaker. The speed of voice and the height of the filter 430 can adjust the height of the voice simultaneously by adjusting the speed of the voice. Specifically, the speech speed and pitch filter 430 can adjust the speed multiples of speech 440 of the speaker 440 to change the speech speed and pitch. For example, when the speed ratio 431 of the speech 440 of the speaker is set to 1.103, the speed of the speaker's voice 440 is 1.103 times faster and the speed of the speaker's voice 440 The elevation can also be converted. In addition, the speed of speech and the height filter 430 can adjust the length of the speech 440 of the speaker to change the speed and the height of the speech. For example, when the current length of the speech 440 of the speaker is 1.980 sec and the modification length 432 is 1.795 sec, the length of the speech 440 of the existing speaker is shortened, The elevation can also be converted.

As the speech speed and pitch filter 430 is applied to speech 440 of the speaker, a third speech 441 of the speaker to which the speech conversion filter is applied and a fourth speech 442 of the speaker to which the speech conversion filter is applied . For example, if the speaker's voice 430 is applied to one or more speed multiples 431, a third voice 441 of the speaker can be generated which is faster than the speaker's voice 440. The generated third voice 441 is output at a higher speed than the voice 440 of the speaker and has a higher frequency than the voice 440 of the speaker. Accordingly, the sound height of the generated third sound 441 becomes higher than the sound height of the sound 440 of the speaker. In addition, applying a rate multiple 431 less than 1 to the speaker's voice 430 may produce a fourth voice 442 of the speaker, which is a voice at a slower rate than the speaker's voice 440. The generated fourth voice 442 is output at a slower rate than the voice 440 of the speaker and has a lower frequency than the voice 440 of the speaker. Accordingly, the sound height of the generated fourth sound 442 becomes lower than the sound height of the speaker sound 440. Therefore, the speaker recognition model generation apparatus applies the speech speed and the value of the height filter 430 to the speech 440 of the speaker differently to simulate a speech having a variety of speech speeds and heights. Based on the simulated speech, Run.

Accordingly, the speaker recognition model generation apparatus can apply the equalizer filter, the speed of speech, and the height filter to the speech of the speaker to simulate the characteristics of various speech recognition apparatuses and various environments in which the speech of the speaker can be received, Machine running can be performed.

FIGS. 5A to 5C are exemplary views illustrating a process of classifying a speaker by machine learning based on a speech graph of a speaker to which a speech conversion filter is applied according to an embodiment of the present invention.

Referring to FIG. 5A, the first speech conversion applied speech graph 510 includes a first frequency change graph 511 according to a time of speech of a speaker to which a speech transform filter is applied, a speaker 511 shown in the first frequency change graph 511, The first acoustic feature 512 of the voice. In addition, the first speech translation applied speech graph 510 includes a first speech speech trend line 513 according to the speech translation filter shown in the first frequency change graph 511. At this time, the first speaker's voice trend line 513 is based on the environment in which the speaker's voice is received and the voice conversion filter for the characteristics of the voice recognition apparatus.

Referring to FIG. 5B, the second speech conversion applied speech graph 520 includes a second frequency change graph 521 according to the time of the speech of the speaker to which the speech transform filter is applied and a second frequency change graph 521 And a second acoustic feature 522 of voice. In addition, the second speech translation applied speech graph 520 includes the second speaker's speech tendency line 523 according to the speech translation filter shown in the second frequency change graph 521. At this time, the voice talker line 523 of the first speaker is based on the environment in which the speaker's voice is received and the voice conversion filter for the characteristics of the voice recognition apparatus.

5A and 5B, the voice of the speaker applied to the voice graph 510 applied to the first voice conversion filter and the voice graph 520 applied to the second voice conversion filter are the same. Therefore, even if the same speaker is uttered and the voice including the same word is used, a different graph can be derived according to the voice conversion filter. That is, even though the same voice is used, the acoustic characteristics can be calculated differently according to the voice conversion filter.

Accordingly, the speaker recognition model generation apparatus applies a high weight to the first acoustic feature 512 and the second acoustic feature 522, so that even if the speaker's voice is converted through various voice conversion filters, So that the machine can run on the basis of the speech characteristics of the speaker.

That is, the speaker recognition model generation apparatus lowers the speaker's voice trend line according to the voice conversion filter and gives a high weight to the acoustic characteristic of the speaker's voice, thereby minimizing the voice change of the speaker according to the voice conversion filter, Running can be performed.

Accordingly, the speaker recognition model generation apparatus performs machine learning, so that the speaker can be recognized through the characteristics of the speaker's voice.

The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. The software module may reside in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, which is capable of reading information from, and writing information to, the storage medium. Alternatively, the storage medium may be integral with the processor. The processor and the storage medium may reside within an application specific integrated circuit (ASIC). The ASIC may reside within the user terminal. Alternatively, the processor and the storage medium may reside as discrete components in a user terminal.

Although the embodiments of the present invention have been described in detail with reference to the accompanying drawings, it is to be understood that the present invention is not limited to those embodiments and various changes and modifications may be made without departing from the scope of the present invention. . Therefore, the embodiments disclosed in the present invention are intended to illustrate rather than limit the scope of the present invention, and the scope of the technical idea of the present invention is not limited by these embodiments. Therefore, it should be understood that the above-described embodiments are illustrative in all aspects and not restrictive. The scope of protection of the present invention should be construed according to the following claims, and all technical ideas within the scope of equivalents should be construed as falling within the scope of the present invention.

100: Speaker recognition model generation device
110:
120: Processor
130:
310: Voice Conversion Filter Graph
311: Voice Conversion Filter Graph
311: Voice conversion filter graph applied to speaker's voice
312: Weighted speech conversion filter graph
320: Speaker's voice graph
321: Speaker's original voice graph
322: weighted speech graph
330: Speaker recognition graph
410: Equalizer filter
411: 0 dB
412: - dB
412: + dB
420: Speaker's Voice
421: First voice
422: second voice
430: Speech speed and elevation filter
431: Speed multiplier
432: Change length
440: Speaker's voice
441: Third voice
442: fourth voice
510: First speech conversion filter applied speech graph
511: First frequency change graph
512: First Acoustic Feature
513: Speech trend line of the first speaker
520: Second speech conversion filter applied speech graph
521: Second frequency change graph
522: Second Acoustic Features
523: Second Acoustic Features

Claims (12)

Receiving a voice of a speaker to be machine learning;
Applying a speech conversion filter to the speech of the speaker to convert speech according to the environment in which the speech of the speaker can be received and the characteristics of the speech recognition apparatus; And
And machine learning a voice of the speaker to which the voice of the speaker and the voice of the speaker to which the voice conversion filter is applied is acquired. The method according to claim 1,
Wherein the step of acquiring the speaker recognition model comprises:
And performing a machine learning of the speech of the speaker to which the speech of the speaker and the speech of the speaker to which the speech conversion filter is applied to obtain a first weight of the speech conversion filter and a second weight of the speech of the speaker, How to create a model. 3. The method of claim 2,
Wherein the obtaining of the first weight and the second weight comprises:
And acquiring the first weight and the second weight higher than the first weight through machine learning. The method according to claim 1,
Wherein applying the speech translation filter comprises:
Applying the speech conversion filter to the speech of the speaker to change at least one of an equalizer, a reverb, a speed of the speech, and a height of the speech, respectively. The method according to claim 1,
Wherein the step of acquiring the speaker recognition model comprises:
Generating a database storing speech of the speaker to which the speech conversion filter is applied; And
Learning the speech characteristics of the speaker corresponding to the speech conversion filter based on the speech of the speaker stored in the database. The method according to claim 1,
When the speech of the speaker is received,
Further comprising the step of recognizing the speaker by applying the speaker recognition model to the speech of the speaker. A communication unit for receiving a voice of a speaker to be machine-operated; And
And a processor for applying a speech conversion filter to the speech of the speaker to convert the speech according to the environment in which the speech of the speaker can be received and the characteristics of the speech recognition apparatus,
The processor comprising:
And a speaker recognition model is obtained by machine-running the speech of the speaker to which the speech of the speaker and the speech of the speaker to which the speech conversion filter is applied. 8. The method of claim 7,
The processor comprising:
Wherein the voice conversion filter is adapted to apply at least one of an equalizer, a reverb, a speed of the voice and a height of the voice to the voice of the speaker. 8. The method of claim 7,
Further comprising a storage unit for storing speech of the speaker to which the speech conversion filter is applied,
Wherein,
Generates a database of the speech of the speaker,
The processor comprising:
And learning the speech characteristics of the speaker corresponding to the speech conversion filter based on the speech of the speaker stored in the database. 8. The method of claim 7,
The processor comprising:
And machine-runs the speech of the speaker applying the speech of the speaker and the speech conversion filter to obtain a first weight of the speech conversion filter and a second weight of the speech of the speaker. 8. The method of claim 7,
The processor comprising:
When the speech of the speaker is received,
And recognizing the speaker by applying the speaker recognition model to the speech of the speaker. A voice of a speaker to be machine-operated is received,
Applying a speech conversion filter for converting speech according to the environment in which the speech of the speaker can be received and the characteristics of the speech recognition apparatus to the speech of the speaker,
And a speaker recognition model is acquired by machine-running the speech of the speaker to which the speech of the speaker and the speech conversion filter are applied.

Images