KR102044520B1 - Apparatus and method for discriminating voice presence section - Google Patents

Abstract

An apparatus for determining a voice presence interval according to an embodiment of the present invention is a machine for extracting a feature vector extracting a feature vector for each frame of an audio signal and receiving a feature vector to estimate a voice existence probability per frame and a kind of noise per frame. The learning model unit and a discrimination unit for discriminating a frame having a voice presence probability equal to or greater than a predetermined threshold value are included.

Description

Apparatus and method for discriminating voice presence section {APPARATUS AND METHOD FOR DISCRIMINATING VOICE PRESENCE SECTION}

The present invention relates to an apparatus and method for determining a voice presence section, and more particularly, based on a feature vector extracted from a sound signal, estimating a voice presence probability in a sound signal and a kind of noise included in the sound signal. The present invention relates to a voice presence section determination device and method for determining a section.

In the fields of voice communication, voice recognition, etc., discriminating a voice section from an acoustic signal is a necessary preprocessing step, and various techniques for more accurately determining the voice section have been actively studied.

In the meantime, the speech and noise discrimination technique using a statistical model has been mainly studied, but recently, a study of modeling the characteristics of speech and noise using machine learning has been conducted.

Korean Laid-Open Patent Publication No. 10-2011-0038447: Target Signal Detection Apparatus and Method Using Statistical Model

The problem to be solved in the embodiment of the present invention is to provide a technique using a feature vector containing the unique information of the sound signal more efficiently in the speech section discrimination in the sound signal using the machine learning.

In addition, it is intended to provide a machine learning model having characteristics that are robust from the effects of noise by allowing the machine learning model to learn information on the types of noise that are closely related to determining a speech section in an acoustic signal.

However, the technical problem to be achieved by the embodiment of the present invention is not limited to the above-mentioned problem, various technical problems can be derived within the scope apparent to those skilled in the art from the following description.

According to an embodiment of the present invention, an apparatus for determining a voice presence interval may include a feature vector extracting unit extracting a feature vector for each frame of an audio signal, and receiving the feature vector to determine a voice existence probability per frame and a type of noise for each frame. And a discriminating unit for estimating a machine learning model unit for estimating and a frame having the voice presence probability equal to or greater than a predetermined threshold.

In this case, the frame may be obtained by dividing the sound signal by a predetermined time unit.

The machine learning model unit may set a feature vector of a learning sound signal including noise as input data of machine learning, and set the presence or absence of a voice in the learning sound signal and the kind of noise included in the learning sound signal as output data. Is generated by learning a parameter for outputting the output data from the input data, wherein the parameter is a correlation between the presence or absence of a voice in the learning sound signal and the type of noise included in the learning sound signal. Can be learned.

In addition, the type of noise set as the output data may be specified as a class including identification information about non-voice.

In the method for determining the presence of speech according to an embodiment of the present invention, the method may be performed by one or more processors, extracting a feature vector for each frame of an audio signal, receiving the feature vector, and probing the voice presence probability per frame. Estimating a type of noise for each frame and determining a frame having a voice presence probability greater than or equal to a predetermined threshold.

According to an embodiment of the present invention, by extracting a feature vector based on a sound signal expressed in the time domain, a feature vector containing unique information of the sound signal may be used for machine learning.

In addition, by setting not only the presence of speech in the acoustic signal but also the kind of noise included in the acoustic signal as the output data of the machine learning stage, the parameters of the machine learning model are trained so that the correlation between the existence of speech and the type of noise can be known. According to the present invention, a robust machine learning model can be generated from the influence of noise, so that even if a large amount of noise is included in the acoustic signal, the voice presence section can be more accurately determined.

1 is a functional block diagram of an apparatus for determining a voice presence interval according to an embodiment of the present invention.
FIG. 2 is an exemplary diagram for explaining that a feature vector extractor of the apparatus for determining a voice presence interval according to an embodiment of the present invention derives a feature vector from an acoustic signal.
3 is an exemplary diagram for describing a process of learning a machine learning model unit of an apparatus for determining a voice presence interval according to an embodiment of the present invention.
4 is a flowchart illustrating a process of a method for determining a voice presence interval according to an embodiment of the present invention.

Advantages and features of the present invention, and methods of achieving them will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but can be implemented in various forms. The present embodiments merely make the disclosure of the present invention complete, and have the ordinary knowledge in the art to which the present invention pertains. It is provided to fully inform the scope of the invention, and the scope of the invention is defined only by the claims.

In describing the embodiments of the present invention, detailed descriptions of well-known functions or configurations will be omitted unless they are actually necessary in describing the embodiments of the present invention. In addition, terms to be described below are terms defined in consideration of functions in the embodiments of the present invention, which may vary according to intentions or customs of users and operators. Therefore, the definition should be made based on the contents throughout the specification.

The functional blocks shown in the figures and described below are only examples of possible implementations. Other functional blocks may be used in other implementations without departing from the spirit and scope of the detailed description. Also, while one or more functional blocks of the present invention are represented by separate blocks, one or more of the functional blocks of the present invention may be a combination of various hardware and software configurations that perform the same function.

In addition, the expression "comprising" certain components merely refers to the existence of the corresponding components as an open expression, and should not be understood as excluding additional components.

Furthermore, when a component is referred to as being connected or connected to another component, it is to be understood that although the component may be directly connected or connected to the other component, there may be other components in between.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

1 is a functional block diagram of an apparatus for determining a voice presence interval 100 according to an embodiment of the present invention.

Referring to FIG. 1, the apparatus for determining a voice presence interval 100 according to an exemplary embodiment of the present invention may include a voice presence probability and an acoustic signal 10 in the acoustic signal 10 based on a feature vector extracted from the acoustic signal 10. Equation of the noise included in the) to determine the presence of the voice in the acoustic signal (10). To this end, the voice presence section determination apparatus 100 according to an embodiment of the present invention includes a feature vector extractor 110, a machine learning model unit 120, and a determiner 130.

The feature vector extractor 110 extracts a feature vector from the sound signal 10. Here, the acoustic signal 10 is a signal representing the sound as a physical numerical value, and the acoustic signal 10 may include voice and noise.

FIG. 2 is an exemplary diagram for explaining that the feature vector extractor 110 of the apparatus for determining the presence of speech according to an embodiment of the present invention derives the feature vector from the sound signal 10.

Referring to FIG. 2, the feature vector extractor 110 according to an embodiment of the present invention, unlike the conventional method of extracting the feature vector by converting the sound signal 10 into a frequency band, is represented in a time band. A feature vector may be extracted from the waveform of the signal 10.

For example, the feature vector extractor 110 may extract a feature vector from a value obtained by sampling and extracting the waveform of the acoustic signal 10. For example, if the sound signal 10 is sampled at a sampling rate of 16 kHz, the physical value of the analog sound signal 10 is recorded 16,000 times for 1 second, and the feature vector is extracted from the recorded value. Can be.

In this case, when the sampling rate is high, since a large number of sampling values are included for one second, the feature vector may be extracted from the sampling values included in the frame 11 in which the acoustic signal 10 is divided into predetermined time units, for example, 0.1 seconds. have.

On the contrary, when the sampling rate is low, when one box shape of the circular waveform shown in FIG. 2 is referred to as one second unit, a feature vector is obtained from values sampled for five seconds using one unit of unit of five seconds. Can be extracted.

At this time, the feature vector extractor 110 extracts the feature vector from the sampled value of the sound signal 10 represented by the time band. The frame 11 of the sound signal 10 is trained through a deep convolutional neural network (DCNN). Feature vectors can be extracted for each. Accordingly, the feature vector containing the unique information of the sound signal 10 can be used more efficiently than the feature vector extracted from the sound signal 10 converted into the frequency domain.

At this time, since the structure of the layer included in the DCNN is a well-known technology, a detailed description thereof will be omitted. Meanwhile, the neural network included in the feature vector extractor 110 described above has been described with reference to DCNN. However, the neural network included in the feature vector extractor 110 is not limited to this example and may extract the feature vector from the acoustic signal 10 through various algorithms.

The machine learning model unit 120 receives the feature vector and estimates the speech existence probability of the sound signal 10 that is the basis of the feature vector and the kind of noise included in the sound signal 10 that is the basis of the feature vector. In this case, the machine learning model unit 120 may estimate a voice presence probability and a kind of noise for each frame 11 of the acoustic signal 10. To this end, the machine learning model unit 120 may be trained as shown in FIG. 3.

3 is an exemplary diagram for describing a process of learning the machine learning model unit 120 of the apparatus for determining the presence of voice according to an embodiment of the present invention.

Referring to FIG. 3, the machine learning model unit 120 may be generated by using a learning sound signal including noise as learning data of machine learning. For example, a machine learning model is set by setting a feature vector of a learning sound signal including noise as input data of machine learning, and setting the presence or absence of a voice in the learning sound signal and the kind of noise included in the learning sound signal as output data. After learning the unit 120, the machine learning model unit 120 may output a speech presence probability and a kind of noise of the corresponding sound signal 10 from the feature vector of the sound signal 10.

In this case, the type of noise set as output data may be specified as a class including identification information about non-voice. For example, by specifying whether a sound other than the voice included in the learning sound signal is a car sound, an airplane sound, or a noise, a class capable of identifying specific information is based on the learning data labeling the learning sound signal. The machine learning model unit 120 may be trained.

Therefore, the machine learning model unit 120 is trained to learn whether a voice exists in the learning sound signal 10, and at the same time, to specify which sounds are noises other than the voice, so that the machine learning model unit 120 is provided. The various sounds included in the acoustic signal 10 can be better distinguished from the voice and the sound other than the voice.

That is, the machine learning model unit 120 is trained to output not only the voice presence probability but also the result of the kind of noise, so that the parameters of the machine learning model are the presence or absence of the voice of the learning acoustic signal 10 and the like. Since the correlation with the type of noise included in the learning sound signal 10 is learned, the machine learning model unit 120 may more accurately estimate the voice presence probability from the specific sound signal 10.

Therefore, the machine learning model unit 120 has robust characteristics from the influence of noise by learning information on the types of noise having a close relationship in determining a voice presence probability.

The determination unit 130 determines the frame 11 whose speech existence probability estimated by the machine learning model unit 120 is equal to or greater than a predetermined threshold. For example, when the predetermined threshold is 75%, if the machine learning model unit 120 estimates that the voice existence probability is 80% for the specific frame 11, the determination unit 130 generates a voice in the frame 11. Can be determined to exist.

In addition, the determination unit 130 may extract only the voice signal through a filtering process of removing noise from the frame 11 determined that the voice exists. Since the process of separating the noise from the acoustic signal 10 is a known technique, a detailed description thereof will be omitted.

Meanwhile, the feature vector extracting unit 110, the machine learning model unit 120, and the determining unit 130 included in the above-described embodiment include a memory including instructions programmed to perform these functions, and a microcomputer that performs these instructions. It may be implemented by a computing device including a processor.

4 is a flowchart illustrating a process of a method for determining a voice presence interval according to an embodiment of the present invention. Each step of the voice presence section determination method according to FIG. 4 may be performed by the voice presence section determination apparatus 100 described with reference to FIGS. 1 to 3.

First, the feature vector extractor 110 extracts a feature vector for each frame 11 of the sound signal 10 (S410). Next, the machine learning model unit 120 receives the feature vector and estimates the speech existence probability per frame 11 and the type of noise for each frame 11 (S420). Thereafter, the determination unit 130 determines a frame 11 having a voice presence probability greater than or equal to a predetermined threshold (S430).

On the other hand, since the process for the components that are the subject of each step described above to perform the step has been described with reference to Figures 1 to 3, duplicate description is omitted.

According to the above-described embodiment, by extracting the feature vector based on the sound signal 10 expressed in the time domain, the feature vector containing the unique information of the sound signal 10 can be used for machine learning more efficiently.

In addition, the type of noise included in the acoustic signal 10 as well as the presence or absence of the voice in the acoustic signal 10 is set as output data of the machine learning step so that the parameters of the machine learning model are correlated with the voice existence probability and the type of noise. By learning to know, since a robust machine learning model can be generated from the influence of noise, even if a large amount of noise is included in the acoustic signal 10, it is possible to more accurately determine the presence of speech.

Embodiments of the present invention described above may be implemented through various means. For example, embodiments of the present invention may be implemented by hardware, firmware, software, or a combination thereof.

In the case of a hardware implementation, a method according to embodiments of the present invention may include one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), and Programmable Logic Devices (PLDs). It may be implemented by field programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, and the like.

In the case of an implementation by firmware or software, the method according to the embodiments of the present invention may be implemented in the form of a module, a procedure, or a function that performs the functions or operations described above. The computer program in which the software code or the like is recorded may be stored in a computer readable recording medium or a memory unit and driven by a processor. The memory unit may be located inside or outside the processor, and may exchange data with the processor by various known means.

As such, those skilled in the art will appreciate that the present invention can be implemented in other specific forms without changing the technical spirit or essential features thereof. Therefore, the above-described embodiments are to be understood as illustrative in all respects and not as restrictive. The scope of the present invention is shown by the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalents should be construed as being included in the scope of the present invention. .

10: sound signal
11: frame
100: voice presence section determination device
110: feature vector extracting unit
120: machine learning model unit
130: determination unit

Claims (6)

A feature vector extracting unit extracting a feature vector from a value obtained by sampling the original waveform of the sound signal expressed in the time band for each frame;
A machine learning model unit which receives the feature vector and estimates the speech existence probability per frame and the type of noise per frame; And
A determination unit for determining a frame having the voice presence probability equal to or greater than a predetermined threshold value,
The machine learning model unit,
A feature vector of a plurality of learning sound signals associated with each of a plurality of kinds of noises is set as input data of machine learning, and whether a voice exists for each of the plurality of learning sound signals and noise included in each of the plurality of learning sound signals. Is set by setting a type of as an output data, and learning a parameter for outputting the output data from the input data,
The parameter is
A correlation between the presence or absence of a voice of the learning sound signal and the type of noise included in the learning sound signal,
The type of noise set as the output data is
Specified as a class containing identifying information for non-voice
Voice presence section determination device. The method of claim 1,
The frame,
The sound signal divided by a predetermined time unit
Voice presence section determination device. delete delete In the voice presence interval determination method, the method is performed by one or more processors, and the method includes:
Extracting a feature vector from a value obtained by sampling the original waveform of the sound signal expressed in the time band for each frame;
Estimating the speech existence probability per frame and the type of noise for each frame by receiving the feature vector; And
Determining a frame in which the voice presence probability is equal to or greater than a predetermined threshold value;
The estimating step,
The feature vector of the learning sound signal including noise is set as input data of machine learning, the presence or absence of a voice in the learning sound signal and the kind of noise included in the learning sound signal are set as output data. Generating by learning a parameter for outputting said output data,
The parameter is
The correlation between the presence or absence of a voice of the learning sound signal and the type of the noise included in the learning sound signal is learned.
The kind of noise set as the output data is
Specified as a class containing identifying information for non-voice
Voice presence section determination method. Extracting a feature vector from a value obtained by sampling the original waveform of the sound signal expressed in the time band for each frame;
Estimating the speech existence probability per frame and the type of noise for each frame by receiving the feature vector; And
Causing the processor to perform a step of determining a frame at which the speech presence probability is greater than or equal to a predetermined threshold,
The estimating step,
The feature vector of the learning sound signal including noise is set as input data of machine learning, the presence or absence of voice in the learning sound signal and the kind of noise included in the learning sound signal are set as output data. Generating by learning a parameter for outputting said output data,
The parameter is
The correlation between the presence or absence of the voice of the learning sound signal and the type of the noise included in the learning sound signal is learned,
The type of noise set as the output data is
Contains instructions to be specified as a class containing identifying information for nonvoice
Computer-readable recording medium in which the program is recorded.

Images