KR102504043B1 - Face-to-face Recording Apparatus and Method with Robust Dialogue Voice Separation in Noise Environments - Google Patents

Abstract

Disclosed is a face-to-face recording device and method with a conversational voice separation function that is robust against noise environments.
According to an embodiment of the present invention, a microphone disposed in a predetermined place to record a sound signal generated in the place and a sound signal recorded by the microphone are received, and only noise signals in the sound signal are filtered, and the noise signal is removed. A face-to-face recording system characterized by comprising a face-to-face recording device that separates each of the voice signals within the filtered sound signal and a sound source storage device that receives and stores the voice signals separated by the face-to-face recording device.

Description

Face-to-face Recording Apparatus and Method with Robust Dialogue Voice Separation in Noise Environments}

The present invention relates to a face-to-face recording apparatus and method that has a conversational voice separation function robustly in a noisy environment and can easily tag separated voices.

The contents described in this part merely provide background information on the present embodiment and do not constitute prior art.

Recently, financial product sellers are obligated to explain investment products to customers, but cases of financial and psychological damage due to claims such as not being fully explained by customers are increasing. As an example of financial damage, if financial product sellers, such as banks or securities companies, violate their duty of explanation or commit unfair acts, they must compensate up to 50% of the revenue related to the violation as a punitive penalty.

Therefore, when customers claim damages for violating the duty of explanation (even though financial product sellers fulfilled their duty of explanation), claiming that they did not properly hear about the product, financial product sellers explained the financial product to prevent such damage. should prove As part of this, financial product sellers are introducing a face-to-face recording system, and the related market is growing.

There is a voice recorder in the space where the customer and the (financial product) seller consult, and the recorded voice is stored to prepare for the above situation. At this time, it is efficient to separate ambient noise from the recorded voice and separate and store the voices of the customer and the vendor, respectively. However, in the conventional recording system, there are technical difficulties in performing all of these, resulting in a decrease in the quality of the recorded voice or There were inconveniences such as an increase in the volume of the recorded voice.

An object of an embodiment of the present invention is to provide a face-to-face recording device and method for recording voice, removing noise in a recorded file, separating each sound source, and tagging and storing it.

According to one aspect of the present invention, a microphone disposed in a predetermined place to record a sound signal generated in the place and a sound signal recorded by the microphone are applied, and only the noise signal in the sound signal is filtered, and the noise signal is filtered. A face-to-face recording system characterized in that it includes a face-to-face recording device that separates each of the voice signals from the face-to-face recording device and a sound source storage device that receives and stores the voice signals separated by the face-to-face recording device.

According to one aspect of the present invention, the microphone is characterized in that it can have one or more channels.

According to one aspect of the present invention, the face-to-face recording device is characterized in that the received sound signal is converted into a frequency domain.

According to one aspect of the present invention, the face-to-face recording device includes a database for storing a noise signal and a voice signal, respectively, and learns the noise signal and the voice signal in each database.

According to one aspect of the present invention, the face-to-face recording device is characterized in inferring a noise canceling frequency filter to filter a noise signal based on a learning result and a frequency-converted sound signal.

According to one aspect of the present invention, in a method for a face-to-face recording system to record a sound signal and separate and store only a voice signal, a recording process of recording a sound signal generated by being placed in a predetermined place, and a recording process recorded in the recording process The filtering process of receiving the sound signal and filtering only the noise signal within the sound signal, the separation process of separating each of the audio signals from the sound signal from which the noise signal is filtered by the filtering process, and the audio signals separated by the separation process It provides a voice signal separation method comprising a storage process of receiving and storing.

According to an aspect of the present invention, a domain conversion unit for receiving a sound signal at a preset location and converting a domain of the sound signal into a frequency domain, learning a voice signal and a noise signal, and converting the signal converted by the domain conversion unit into a frequency domain. A filter inference unit for inferring a noise canceling frequency filter from the received and learned result, and a filtering unit for filtering the noise signal within the signal converted by the domain conversion unit using the inferred filter, and the noise signal is removed by the filtering unit. Provided is a face-to-face recording device comprising a voice separation unit separating each voice signal from a sound signal and a tagging unit tagging the location or direction of the source of each separated voice signal.

According to one aspect of the present invention, the domain converter converts the domain of the sound signal into a frequency domain by performing Fast Fourier Transform (FFT).

According to an aspect of the present invention, the sound signal received by the domain conversion unit is a sound signal recorded from one or more channels.

According to one aspect of the present invention, the filter reasoning unit may include a voice database for storing a voice signal and a noise database for storing a noise signal.

According to one aspect of the present invention, the filter reasoning unit is characterized in that, when learning a voice signal, it learns a voice signal in which a voice signal exists as many as a number equal to or less than the number of channels in which the voice signal is recorded.

According to one aspect of the present invention, the voice separation unit is characterized in that an implicit sound source separation technology is used.

According to one aspect of the present invention, in a method of separating and tagging only a voice signal from a sound signal by a face-to-face recording device, a conversion process of receiving a sound signal at a predetermined place and converting the domain of the sound signal into a frequency domain and the voice An inference process of learning a signal and a noise signal, receiving a signal converted in the conversion process, and inferring a noise canceling frequency filter from the learned result, and using the filter inferred in the inference process, in the converted signal in the conversion process. A filtering process of filtering noise signals, a separation process of separating each audio signal from a sound signal from which noise signals have been removed through the filtering process, and a tagging process of tagging the location or direction of the origin of each audio signal separated in the separation process. It provides a face-to-face recording method comprising a process.

According to one aspect of the present invention, the conversion process is characterized in that the domain of the sound signal is converted into a frequency domain by performing FFT (Fast Fourier Transform).

As described above, according to one aspect of the present invention, by recording voice, removing noise in the recorded file, separating and tagging each sound source, each sound source can be distinguished and stored, and the sound quality of the stored sound source can be improved. There are advantages to being able to

1 is a diagram showing the configuration of a face-to-face recording system according to an embodiment of the present invention.
2 is a diagram showing the configuration of a face-to-face recording device according to an embodiment of the present invention.
3 is a diagram showing the configuration of a filter reasoning unit according to an embodiment of the present invention.
4 is a diagram showing a histogram analyzed by a tagging unit according to an embodiment of the present invention.
5 is a flowchart illustrating a method of separating and tagging each sound source by recording by a face-to-face recording device according to an embodiment of the present invention.
6 is a diagram illustrating a method of recording a sound signal by a face-to-face recording system according to an embodiment of the present invention, separating and storing only a voice signal.

Since the present invention can make various changes and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail. However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all modifications, equivalents, or substitutes included in the spirit and technical scope of the present invention. Like reference numerals have been used for like elements throughout the description of each figure.

Terms such as first, second, A, and B may be used to describe various components, but the components should not be limited by the terms. These terms are only used for the purpose of distinguishing one component from another. For example, a first element may be termed a second element, and similarly, a second element may be termed a first element, without departing from the scope of the present invention. The term and/or includes a combination of a plurality of related recited items or any one of a plurality of related recited items.

It is understood that when an element is referred to as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, but other elements may exist in the middle. It should be. On the other hand, when an element is referred to as “directly connected” or “directly connected” to another element, it should be understood that no intervening element exists.

Terms used in this application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. It should be understood that terms such as "include" or "having" in this application do not exclude in advance the possibility of existence or addition of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification. .

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related art, and unless explicitly defined in the present application, they should not be interpreted in an ideal or excessively formal meaning. don't

In addition, each configuration, process, process or method included in each embodiment of the present invention may be shared within a range that does not contradict each other technically.

1 is a diagram showing the configuration of a face-to-face recording system according to an embodiment of the present invention.

Referring to FIG. 1 , a face-to-face recording system 100 according to an embodiment of the present invention includes a microphone 110, a face-to-face recording device 120, and a sound source storage device 130.

The face-to-face recording system 100 records the voice of a corresponding person or each person in a space where one person or two or more people exist. At this time, there is a high probability that there will be noise other than the voice of the person or each person due to the nature of the space, and the face-to-face recording system 100 filters noise based on the recorded file to separate and store only high-quality voice. In addition, when there are two or more people in the space, the voice of each person is separated, the location (direction) of the origin of the voice is classified, and the person from whom the separated voice is from is tagged. Accordingly, the sound source storage device 130 may classify and store desired target voices using tags, and may delete or separately manage unnecessary voices. In addition, since the stored voices are also in a state in which noise is filtered, the quality of voices can be improved.

The microphone 110 records a sound signal generated within a person or a space where each person exists. As shown in FIG. 1, the microphone 110 has as many channels as the number of people present in the space, and each channel records sound signals generated in the space. The microphone 110 transmits the recorded recording file to the face-to-face recording device 120.

The face-to-face recording device 120 receives the sound signal recorded by the microphone 110, filters the noise signal, and classifies and tags the audio signal generated by each person. The face-to-face recording device 120 primarily infers a filter to separate (filter) the noise signal and the human voice signal based on big data learning, and separates the noise signal and the human voice signal. The face-to-face recording device 120 separates each voice signal from a secondly separated human voice signal using sound source separation when a plurality of people generate voice at a corresponding place. At this time, the face-to-face recording device 120 separates the location (direction) of the origin of the audio signal after separating the audio signal, and distinguishes where the audio signal came from or from whom. As mentioned in the background technology, when the face-to-face recording system 100 is used in a space for explaining financial products, the locations of customers, vendors, and other personnel are usually fixed. Therefore, if the location of the origin of the voice signal can be known, it is possible to classify and tag whose voice the corresponding voice signal belongs to. Since the face-to-face recording device 120 separates the noise signal and the human voice signal based on big data learning (deep learning), it can simply and accurately separate the noise signal and the human voice signal. In addition, since the face-to-face recording device 120 uses the implicit sound source separation technology for the audio signal from which the noise signal is separated, the audio signal in the state where the noise signal is separated is accurately separated without the need to learn the voice signals of each person individually. You can do it. A detailed structure of the face-to-face recording device 120 will be described later with reference to FIGS. 2 and 3 .

The sound source storage device 130 stores voice signals that are separated from the face-to-face recording device 120 and tagged. Since each person is tagged and transmitted to the sound source storage device 130 while being separated, the sound source storage device 130 can separately store voice signals of each person. As a result, management and search may be more easily performed. In addition, since the noise signal is separated from the sound signal recorded from the microphone 110, the volume of the audio signal is relatively reduced. Accordingly, the storage capacity of the sound source storage device 130 may be more marginal.

2 is a diagram showing the configuration of a face-to-face recording device according to an embodiment of the present invention.

Referring to FIG. 2 , the face-to-face recording device 120 according to an embodiment of the present invention includes a domain conversion unit 210, a filter reasoning unit 220, a filtering unit 230, a voice separation unit 240, and a tagging unit. (250).

The domain converter 210 receives the sound signal recorded by the microphone 110 from the microphone 110 and converts it into a frequency domain. When receiving sound signals from the microphone 110, the domain conversion unit 210 receives sound signals separately when the microphone 110 records sound signals in multiple channels. Since the sound signal received by the domain converter 210 is time domain data, the domain converter 210 converts the input sound signal into a frequency domain. The domain conversion unit 210 may perform FFT (Fast Fourier Transform) in performing the above-described operation.

The filter reasoning unit 220 learns the stored voice signal and noise signal, and uses the sound signal converted by the domain conversion unit 210 to infer an optimal filter to separate the noise signal from the input sound signal. The filter reasoning unit 220 learns various noise signals that may occur in various spaces and various human voice signals. The filter reasoning unit 220 receives the sound signal converted by the domain conversion unit 210, and infers an optimal filter for the received data based on a result of learning using the received data. A detailed structure of the filter reasoning unit 220 is shown in FIG. 3 .

3 is a diagram showing the configuration of a filter reasoning unit according to an embodiment of the present invention.

Referring to FIG. 3 , the filter inference unit 120 according to an embodiment of the present invention includes a voice database 310, a noise database 315, a multi-speech generator 320, a deep learning unit 330, and an inference unit. (340).

The voice database 310 stores various types of voice signals that the face-to-face recording device 120 wants to acquire separately from noise signals. As in the foregoing example, if it is desired to separate voice signals of persons or persons located in a place where the microphone 110 is (or will be placed), the voice database 310 is a deep learning unit 330 to learn, various It stores numerous voice signals of humans.

The noise database 315 stores various types of noise signals that the face-to-face recording device 120 wants to separate from the voice signal. As in the foregoing example, the noise database 315 stores various types of noise signals that may occur at a place where the microphone 110 is or will be placed, for the deep learning unit 330 to learn.

The multi-speech generation unit 320 generates multi-speech signals to be learned by the deep learning unit 330 among the voice signals stored in the voice database 310 . The multi-voice generating unit 320 generates multi-voice in which voice signals exist as many as the number of microphone channels installed in the recording device or less. For example, considering the case where the microphone 110 records sound signals through two channels in a place where two people are located, the multi-voice generator 320 has one single voice signal or two voice signals at the same time. The deep learning unit 330 generates an amount sufficient to learn by alternating multiple voice signals at random. When the deep learning unit 330 learns a voice signal suitable for the number of people actually present in the place, it can learn and infer a filter that will separate only the noise signal from the sound signal accurately recorded by the microphone 110.

The deep learning unit 330 receives noise signals from the noise database 315 and speech signals from the speech database 310 or the multivoice generator 320 and performs learning to generate a noise canceling frequency filter. The deep learning unit 330 learns data stored in the databases 310 and 315. The deep learning unit 330 learns the data stored in the noise database 315 to recognize information about the noise signal, and the multiple voice generation unit ( 320) to recognize information about the voice signals by learning the multiple voice signals generated. The deep learning unit 330 learns to generate a noise canceling frequency filter capable of filtering only the noise signal in the sound signal mixed with both through this learning.

The reasoning unit 340 receives the frequency value converted by the domain conversion unit 210 based on the learning result of the deep learning unit 330 and infers a noise canceling frequency filter. The inference unit 340 receives a frequency value converted by the domain conversion unit 210 from the learning result of the deep learning unit 330 . At this time, when the domain conversion unit converts each of the sound signals recorded from a plurality of channels into the frequency domain, the frequency value received by the reasoning unit 340 is converted from the sound signal input to each channel. Instead, one value obtained by averaging values converted from sound signals input to each channel is received. When the reasoning unit 340 receives the values converted from the sound signals input to each channel instead of receiving the average value, the spatial characteristics are different, so when separating the voice signal after removing the noise signal, the location (direction) of the origin of the voice It is difficult to infer The frequency filter can be commonly used only when the inference unit 340 receives the average value of the values converted by the domain conversion unit 210 . Accordingly, the spatial characteristics of each voice signal from which noise has been removed become the same. Accordingly, the inference unit 340 receives an average value of values converted from sound signals input to each channel, and the inference unit 340 infers a noise canceling frequency filter based on the received values.

Referring back to FIG. 2 , the filtering unit 230 uses the filter reasoning unit 220, in particular, the filter inferred by the reasoning unit 340, to convert the noise signal within the sound signal input to the domain conversion unit 210. filter The filtering unit 230 multiplies the inferred (noise removal frequency) filter by the frequency domain value converted by the domain conversion unit 210 to remove only the noise signal from the recorded sound signal. At this time, when each sound signal recorded from a plurality of channels is input to the domain converter 210, the filtering unit 230 filters each sound signal frequency-converted value. Accordingly, among the sound signals recorded in each channel, a voice (frequency domain value) from which noise signals are removed is derived.

The voice separation unit 240 separates voice signals using an implicit sound source separation technique. When there are a plurality of audio signals, the audio separation unit 240 separates each audio signal. For example, there may be a situation in which voice signals S ₁ (t, k) and S ₂ (t, k) are recorded on microphone channels 1 and 2 in an arbitrary space in the case of frequency k index and time t. If X ₁ (t, k) and X ₂ (t, k) are the audio signals recorded on microphone channels 1 and 2, respectively, X ₁ (t, k) and X ₂ (t, k) satisfy the following .

X ₁ (t, k)=H ₁₁ (t, k)S ₁ (t, k)+H ₁₂ (t, k)S ₂ (t, k)

X ₂ (t, k)=H ₂₁ (t, k)S ₁ (t, k)+H ₂₂ (t, k)S ₂ (t, k)

Here, H ₁₁ (t, k) is the transfer conversion complex weight between microphone channel 1 and the first voice signal, H ₁₂ (t, k) is the transfer conversion complex weight between microphone channel 1 and the second voice signal, H ₂₁ (t, k) denotes a transfer conversion complex weight between microphone channel 2 and the first voice signal, and H ₂₂ (t, k) denotes a transfer conversion complex weight between microphone channel 2 and the second voice signal.

The audio separator 240 estimates and separates S ₁ (t, k) and S ₂ (t, k) from the noise-removed recorded audio signals X ₁ (t, k) and X ₂ (t, k). . According to the above formula,

is achieved By the above formula,

is satisfied

를 만족하는 음원 분리행렬W(k)가 존재한다고 가정할 경우, 음성 분리부(240)는 S₁(t, k)와 S₂(t, k)가 서로 독립적(상관관계가 존재하지 않음)인 값을 갖는 W(k)를 연산(X₁(t, k)와 X₂(t, k)는 알고 있는 값이기 때문에)하여 추정한다.At this time,

Assuming that there exists a sound source separation matrix W(k) _{that satisfies} W(k) having a value of is calculated (Since X ₁ (t, k) and X ₂ (t, k) are known values), it is estimated.

The voice separator 240 may separate sound sources S ₁ (t, k) and S ₂ (t, k) by estimating the sound source separation matrix W(k) using implicit sound source separation.

The tagging unit 250 analyzes the direction of the voice using the sound source separation matrix estimated by the voice separation unit 240 and tags the location (direction) of the origin of the voice signal. The tagging unit 250 calculates the phase angle of the sound source separation matrix as follows.

As described above, the tagging unit 250 calculates phase angles of complex weights of the same sound source in the sound source separation matrix.

The tagging unit 250 analyzes the histogram based on the calculated phase angle result. An example of the histogram analyzed by the tagging unit 250 is shown in FIG. 4 .

4 is a diagram showing a histogram analyzed by a tagging unit according to an embodiment of the present invention.

4 shows a histogram in the case where the microphone 110 records sound signals through two channels at a place where two people are located. In the graph of FIG. 4, the x-axis means the angle and the y-axis means the frequency.

If the voice signal originates from the left side of the reference value (for example, the central axis of the microphone 110), it can be seen that the frequency is significantly higher at an angle smaller than the reference value, as shown in FIG. 4A.

Conversely, if the voice signal originates from the right side with respect to the reference value (for example, the central axis of the microphone 110), it shows a significantly higher frequency at an angle greater than the reference value as shown in FIG. 4B. It can be seen.

Since the locations of customers, sales companies, and other personnel are usually determined, the tagging unit 250 tags whose voice each voice signal belongs to based on the location (direction) of the source of the voice signal.

Since the reasoning unit 340 receives the average value of the frequency domain of the sound signal input to each channel and infers the noise canceling frequency filter, the spatial characteristics do not change. Accordingly, separation of the sound source in the voice separation unit 240 or tagging in the tagging unit 250 is performed without difficulty.

The voice file tagged by the tagging unit 250 is stored in the sound source storage device 130 .

5 is a flowchart illustrating a method of separating and tagging each sound source by recording by a face-to-face recording device according to an embodiment of the present invention.

The domain conversion unit 210 receives sound signals recorded in each channel of the microphone and converts each into a frequency domain (S510).

The filter reasoning unit 220 learns noise signals and voice signals of a number less than or equal to the number of channels (S520).

The filter reasoning unit 220 receives the average value of each sound signal converted to the frequency domain and infers a noise canceling frequency filter from the learned result (S530).

The filtering unit 230 filters the noise signal within each sound signal converted to the frequency domain based on the inferred filter (S540).

The voice separation unit 240 separates each voice signal within the noise-removed sound signal (S550).

The tagging unit 250 tags the location or direction of the source of each separated voice signal (S560).

6 is a diagram illustrating a method of recording a sound signal by a face-to-face recording system according to an embodiment of the present invention, separating and storing only a voice signal.

The microphone 110 is placed in a preset location and records the generated sound signal (S610).

The face-to-face recording device 120 receives sound signals recorded in each channel of the microphone and converts each into a frequency domain (S620).

The face-to-face recording device 120 learns noise signals and voice signals equal to or smaller than the number of channels (S630).

The face-to-face recording device 120 receives the average value of each sound signal converted to the frequency domain and infers a noise canceling frequency filter from the learned result (S640).

The face-to-face recording device 120 filters the noise signal within each sound signal converted to the frequency domain based on the inferred filter (S650).

The face-to-face recording device 120 separates each voice signal from the noise-removed sound signal (S660).

The face-to-face recording device 120 tags the location or direction of the source of each separated audio signal (S670).

The sound source storage device 130 receives and stores the audio signals tagged by the face-to-face recording device 120 (S680).

5 and 6 describe that each process is sequentially executed, but this is merely an example of the technical idea of one embodiment of the present invention. In other words, those skilled in the art to which an embodiment of the present invention pertains may change and execute the order described in each drawing or perform one or more processes of each process without departing from the essential characteristics of an embodiment of the present invention. 5 and 6 are not limited to a time-series order, since various modifications and variations can be applied by executing them in parallel.

Meanwhile, the processes shown in FIGS. 5 and 6 can be implemented as computer readable codes on a computer readable recording medium. A computer-readable recording medium includes all types of recording devices in which data readable by a computer system is stored. That is, computer-readable recording media include storage media such as magnetic storage media (eg, ROM, floppy disk, hard disk, etc.) and optical reading media (eg, CD-ROM, DVD, etc.). In addition, the computer-readable recording medium may be distributed to computer systems connected through a network to store and execute computer-readable codes in a distributed manner.

The above description is merely an example of the technical idea of the present embodiment, and various modifications and variations can be made to those skilled in the art without departing from the essential characteristics of the present embodiment. Therefore, the present embodiments are not intended to limit the technical idea of the present embodiment, but to explain, and the scope of the technical idea of the present embodiment is not limited by these embodiments. The scope of protection of this embodiment should be construed according to the claims below, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of rights of this embodiment.

100: face-to-face recording system
110: microphone
120: face-to-face recording device
130: sound source storage device
210: domain conversion unit
220: filter reasoning unit
230: filtering unit
240: voice separator
250: tagging unit
310 Voice database
315 noise database
320: multiple voice generation unit
330: deep learning unit
340: reasoning unit

Claims (14)

a microphone disposed in a predetermined place, having as many channels as the number of people present in the place, and recording a sound signal generated in the place using each channel;
a face-to-face recording device receiving the sound signal recorded by the microphone, filtering only the noise signal within the sound signal, and separating each of the audio signals from the sound signal from which the noise signal is filtered; and
A sound source storage device for receiving and storing voice signals separated by the face-to-face recording device,
The face-to-face recording device,
a domain conversion unit that receives the sound signal recorded by the microphone and converts it into a frequency domain;
a noise database for storing a plurality of types of noise signals;
a voice database for storing a plurality of types of voice signals;
a multi-speech generating unit generating a multi-speech signal to be learned from among the voice signals stored in the voice database;
a deep learning unit receiving noise signals from the noise database and voice signals from the voice database or the multi-speech generation unit and learning to generate a noise canceling frequency filter; and
Based on the learning result of the deep learning unit, an inference unit for inferring a noise canceling frequency filter by receiving one value obtained by averaging values converted from sound signals input to each channel by the domain conversion unit. recording system. According to claim 1,
the microphone,
A face-to-face recording system, characterized in that it can have one or more channels. According to claim 1,
The face-to-face recording device,
A face-to-face recording system characterized by converting an authorized sound signal into a frequency domain. According to claim 3,
The face-to-face recording device,
A face-to-face recording system comprising a database storing each of the noise signal and the voice signal, and learning the noise signal and the voice signal in each database. According to claim 4,
The face-to-face recording device,
A face-to-face recording system characterized by inferring a noise canceling frequency filter to filter a noise signal based on a learning result and a frequency-converted sound signal. delete a domain converter for receiving a sound signal at a preset location and converting a domain of the sound signal into a frequency domain;
a filter inference unit that learns the voice signal and the noise signal, receives the signal converted by the domain converter, and infers a noise canceling frequency filter from the learned result;
a filtering unit filtering a noise signal within the signal converted by the domain conversion unit using the inferred filter;
a voice separator separating each voice signal from the sound signal from which the noise signal has been removed by the filtering unit; and
A tagging unit for tagging the location or direction of the source of each of the separated voice signals,
The filter reasoning unit,
a noise database for storing a plurality of types of noise signals;
a voice database for storing a plurality of types of voice signals;
a multi-speech generating unit generating a multi-speech signal to be learned from among the voice signals stored in the voice database;
a deep learning unit receiving noise signals from the noise database and voice signals from the voice database or the multi-speech generation unit and learning to generate a noise canceling frequency filter; and
Based on the learning result of the deep learning unit, an inference unit for inferring a noise canceling frequency filter by receiving one value obtained by averaging values converted from sound signals input to each channel by the domain conversion unit. recording device. According to claim 7,
The domain conversion unit,
A face-to-face recording device characterized by converting the domain of a sound signal into a frequency domain by performing FFT (Fast Fourier Transform). According to claim 7,
The sound signal received by the domain conversion unit,
A face-to-face recording device, characterized in that the sound signal recorded from one or more channels. delete According to claim 9,
The filter reasoning unit,
In learning the voice signal, the face-to-face recording device, characterized in that for learning the voice signal in which the voice signal exists as many as the number of channels in which the sound signal is recorded. According to claim 7,
The voice separator,
A face-to-face recording device characterized by using implicit sound source separation technology. delete delete

Images