KR20230055025A - Device, method and computer program for recognizing voice - Google Patents

Abstract

An apparatus for recognizing speech, comprising: a feature vector extraction unit extracting a feature vector from speech data; a predictive model learning unit that trains a speech recognition model based on the feature vector and a joint vector including a speech presence probability value for regular speech data; and a voice recognition unit for recognizing a voice using the voice recognition model.

Description

Apparatus, method and computer program for recognizing voice {DEVICE, METHOD AND COMPUTER PROGRAM FOR RECOGNIZING VOICE}

The present invention relates to a device, method and computer program for recognizing speech.

Speech recognition technology is a technology that converts an acoustic signal obtained through a sound sensor such as a microphone into words or sentences. A representative voice recognition service based on voice recognition technology is 'Siri', Apple's voice-based personal assistant service. Siri provides various convenience services such as mobile search, schedule management, phone call, memo and music playback based on iPhone users' voice commands.

In general, voice recognition technology recognizes voice by extracting a sound signal, removing noise, and extracting features of the voice signal and comparing them with a voice model database.

Conventional voice recognition technology detects a section where a human voice exists and a section where a human voice does not exist in voice data through voice activity detection. For example, based on the energy in voice data, a section where human voice exists is marked as '1', and a section where human voice does not exist is marked as '0'.

However, since the conventional voice recognition technology only outputs '0' or '1' to indicate the presence or absence of a human voice in voice data, voice recognition performance may be deteriorated when noise is included in voice data.

Korean Registered Patent Publication No. 10-1066472 (registered on September 15, 2011) Korean Registered Patent Publication No. 10-0998567 (registered on November 30, 2010)

The present invention is to solve the above-mentioned problems of the prior art, and to provide a voice recognition apparatus, method, and computer program capable of measuring and digitizing a section in which a human voice may exist in voice data.

In addition, it is intended to provide a voice recognition apparatus, method, and computer program capable of effectively recognizing even a voice signal in which noise is included in voice data or a first consonant is cut off.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problems described above, and other technical problems may exist.

As a means for achieving the above-described technical problem, an embodiment of the present invention is a device for recognizing speech, comprising: a feature vector extractor extracting a feature vector from speech data; a predictive model learning unit that trains a speech recognition model based on the feature vector and a joint vector including a speech presence probability value for regular speech data; and a voice recognition unit for recognizing a voice using the voice recognition model.

Another embodiment of the present invention is a method for recognizing speech, comprising: extracting a feature vector from speech data; learning a speech recognition model based on the feature vector and a joint vector including a speech existence probability value for regular speech data; and recognizing a voice using the voice recognition model.

Another embodiment of the present invention is a computer program stored on a computer-readable recording medium containing a sequence of instructions for recognizing speech, wherein the computer program, when executed by a computing device, extracts a feature vector from speech data and , a computer readable sequence of instructions for learning a speech recognition model based on the feature vector and a joint vector including a speech presence probability value for regular speech data, and recognizing speech using the speech recognition model. A computer program stored on a recording medium may be provided.

The above-described means for solving the problems is only illustrative and should not be construed as limiting the present invention. In addition to the exemplary embodiments described above, there may be additional embodiments described in the drawings and detailed description.

According to any one of the above-described problem solving means of the present invention, it is possible to measure a section in which a human voice may exist in voice data, and quantify a probability that a human voice exists in the corresponding section.

Therefore, it is possible to provide a voice recognition apparatus, method, and computer program capable of effectively recognizing a human voice even in a voice signal in an environment in which noise is included in voice data or the first consonant is cut off in a specific section.

1 is a configuration diagram of a voice recognition device according to an embodiment of the present invention.
2 is an exemplary diagram for explaining a delay time and periodicity of a specific frame according to an embodiment of the present invention.
3 is an exemplary diagram for explaining an example of a delay time value according to an embodiment of the present invention.
4 is an exemplary diagram for explaining a process of deriving a delay time value according to an embodiment of the present invention.
5 is an exemplary diagram for explaining a gamma filter according to an embodiment of the present invention.
6 is a flowchart of a voice recognition method according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail so that those skilled in the art can easily practice the present invention with reference to the accompanying drawings. However, the present invention may be embodied in many different forms and is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

Throughout the specification, when a part is said to be "connected" to another part, this includes not only the case where it is "directly connected" but also the case where it is "electrically connected" with another element interposed therebetween. . In addition, when a part "includes" a certain component, this means that it may further include other components, not excluding other components, unless otherwise stated, and one or more other characteristics. However, it should be understood that it does not preclude the possibility of existence or addition of numbers, steps, operations, components, parts, or combinations thereof.

In this specification, a "unit" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. Further, one unit may be realized using two or more hardware, and two or more units may be realized by one hardware.

In this specification, some of the operations or functions described as being performed by a terminal or device may be performed instead by a server connected to the terminal or device. Likewise, some of the operations or functions described as being performed by the server may also be performed in a terminal or device connected to the corresponding server.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a configuration diagram of a voice recognition device according to an embodiment of the present invention. Referring to FIG. 1 , the voice recognition apparatus 100 may include a feature vector extractor 110 , a predictive model learner 120 and a voice recognizer 130 . However, the above components 110 to 130 are merely examples of components that can be controlled by the voice recognition apparatus 100 .

Each component of the voice recognition apparatus 100 of FIG. 1 is generally connected through a network. For example, as shown in FIG. 1 , the feature vector extractor 110 , the predictive model learner 120 and the speech recognizer 130 may be connected simultaneously or at a time interval.

A network refers to a connection structure capable of exchanging information between nodes such as terminals and servers, such as a local area network (LAN), a wide area network (WAN), and the Internet (WWW: World Wide Web), wired and wireless data communication network, telephone network, and wired and wireless television communication network. Examples of wireless data communication networks include 3G, 4G, 5G, 3rd Generation Partnership Project (3GPP), Long Term Evolution (LTE), World Interoperability for Microwave Access (WIMAX), Wi-Fi, Bluetooth communication, infrared communication, ultrasonic communication, visible light communication (VLC: Visible Light Communication), LiFi, and the like, but are not limited thereto.

The voice recognition apparatus 100 may measure a section in which a human voice may exist in voice data, and quantify a probability that a human voice exists in the corresponding section. For example, the voice recognition apparatus 100 may calculate a probability value of human voice present in a specific frame among a plurality of frames included in voice data.

The voice recognition apparatus 100 can effectively recognize a human voice even in a voice signal in an environment in which noise is included in voice data or a leading consonant is cut off in a specific section. For example, the voice recognition apparatus 100 can clearly recognize a voice signal in which the first consonant 'ㅎ' in 'I went to school' is cut off from the voice data.

Hereinafter, each component of the voice recognition apparatus 100 will be described.

The feature vector extractor 110 may extract feature vectors from voice data. For example, the feature vector extractor 110 may extract a voice-specific feature value, a Mel-Frequency Cepstral Coefficient (MFCC), from voice data.

Specifically, the feature vector extractor 110 may classify voice data for each frame. The feature vector extractor 110 may extract a spectrum by applying a Fast Fourier Transform (FFT) to each frame of voice data. Here, the fast playy transform is an algorithm for converting a signal into a frequency component, and is an algorithm optimized to perform a conventional Discrete Fourier Transform (DFT) more quickly.

The feature vector extractor 110 may calculate a Mel spectrum by applying a Mel filter bank to the extracted spectrum. The feature vector extractor 110 may extract a feature vector, MFCC, by performing cepstrum analysis on the mel spectrum.

The predictive model learning unit 120 may train a voice recognition model based on a feature vector and a joint vector including a voice presence probability value for regular voice data. For example, the predictive model learning unit 120 may generate a joint vector by using a voice-specific feature value extracted from voice data, an MFCC, and a probability value where voice exists, as shown in Equation 1 below.

<Equation 1>

는 조인트 벡터이고,

는 특징 벡터이고,

는 음성 존재 확률 벡터일 수 있다. In Equation 1,

is the joint vector,

is a feature vector,

may be a voice presence probability vector.

The predictive model learning unit 120 may calculate a voice presence probability value for each frame included in the voice data based on the autocorrelation value and the cross correlation ratio (CCR) of the voice data.

For example, the predictive model learning unit 120 may calculate an autocorrelation value using a normalized autocorrelation function (NACF). The predictive model learning unit 120 may calculate an autocorrelation value using Equation 2 below.

<Equation 2>

은 입력 신호이고, m은 프레임 인덱스이고,

은 신호의 샘플로서 초(second) 단위의 지연 시간(lag time)일 수 있다. 여기서, 지연 시간은 입력 신호로부터 지난 신호의 패턴 중 입력 신호의 패턴과 가장 유사한 패턴을 가지는 지난 신호와 해당 입력 신호 간의 간격에 해당할 수 있다. 즉, 해당 신호의 주기성을 나타낼 수 있다.In Equation 2,

is the input signal, m is the frame index,

is a sample of a signal and may be a lag time in units of seconds. Here, the delay time may correspond to an interval between a previous signal having a pattern most similar to the pattern of the input signal among patterns of the previous signal from the input signal and the corresponding input signal. That is, the periodicity of the corresponding signal may be indicated.

)에서

만큼 시간 이동된 신호(

)를 곱한 후 정규화하여 자기 상관값을 산출할 수 있다. 예측 모델 학습부(120)는 원 신호를 기반으로 지연된 시간, 즉, 주기성에 따라 자기 상관값을 산출하여 음성 데이터의 주기 성분을 추출할 수 있다. For example, the predictive model learning unit 120, referring to Equation 2, the original signal (

)at

A signal time-shifted by (

) and then normalized to calculate the autocorrelation value. The predictive model learning unit 120 may extract a periodic component of voice data by calculating an autocorrelation value according to a delayed time, that is, periodicity, based on the original signal.

For example, the delay time in Equation 2 may be expressed in seconds (sec) as shown in Equation 3 below.

<Equation 3>

,

,

및

는 기본 주파수(

)가 존재할 수 있는 구간성을 의미할 수 있다. 예를 들어, 기본 주파수가 존재할 수 있는 구간은 일반적으로, 80~500Hz이고, 이 때, 해당 주파수는 수학식 4와 같이 200~32의 범위를 갖는 시간 지연(time-lag) 샘플 인덱스로 나타낼 수 있다.In Equation 3,

and

is the fundamental frequency (

) may mean interval properties that may exist. For example, the period in which the basic frequency may exist is generally 80 to 500 Hz, and in this case, the frequency can be represented by a time-lag sample index having a range of 200 to 32 as shown in Equation 4. there is.

<Equation 4>

The predictive model learner 120 may derive a fundamental frequency based on the autocorrelation function. For example, the fundamental frequency can be expressed as in Equation 5 below.

<Equation 5>

The predictive model learner 120 may calculate a similarity between the first frame included in the voice data and the second frame that is a previous frame of the first frame as a cross-correlation ratio. For example, the predictive model learning unit 120 may calculate the cross-correlation ratio using Equation 6 below.

<Equation 6>

은 입력 신호이고,

는 감마 필터(gammation filter)와의 합성곱을 통해 산출될 수 있다. 예를 들어, 하기 수학식 7을 이용할 수 있다.In Equation 6,

is the input signal,

may be calculated through convolution with a gamma filter. For example, Equation 7 below can be used.

<Equation 7>

In Equations 6 and 7, n is a sample and c is a channel, and the frequency for the frame index m can be divided for each band. For example, the predictive model learner 120 may calculate a ratio between the sum of samples of the first frame and the second frame specific channel using Equations 6 and 7.

)를 포함한 k개의 근접한 채널(수학식 6에서

)을 도출할 수 있다. The prediction model learner 120 may derive the number of channels according to the fundamental frequency of the first frame. For example, the predictive model learning unit 120 calculates the fundamental frequency using Equation 2 (

), including k adjacent channels (in Equation 6)

) can be derived.

)의 배수의 주파수 채널(

)들을 도출할 수 있고, 이 때, 범위는 80Hz ~ 1000Hz이고, 각 채널은 Center-frequency와 Band-width로 구성될 수 있다.For example, the predictive model learning unit 120 uses Equation 2, and in Equation 6, the fundamental frequency (

) frequency channels of multiples (

) can be derived, and at this time, the range is 80 Hz to 1000 Hz, and each channel can be composed of a center-frequency and a bandwidth.

= {120, 240, 360, 480, 600, 720, 840, 960}을 도출할 수 있고, 성인 여성 및 어린이의 경우, 현재 프레임의 주파수가 300Hz일 때, k개의 근접한 채널로,

= {300, 600, 900}을 도출할 수 있다.Specifically, the fundamental frequency of an adult male is generally about 120 Hz, and the fundamental frequency of an adult female and child is about 300 Hz. Therefore, in the case of an adult male, when the frequency of the current frame is 120 Hz, the predictive model learning unit 120 uses k adjacent channels,

= {120, 240, 360, 480, 600, 720, 840, 960} can be derived, and in the case of adult women and children, when the frequency of the current frame is 300 Hz, with k adjacent channels,

= {300, 600, 900} can be derived.

Referring to Equations 6 and 7, the predictive model learning unit 120 calculates the sum of at least one sample corresponding to the first frame and the second frame based on the number of channels derived as a similarity between the first frame and the second frame. A ratio between the sum of at least one sample corresponding to may be calculated. For example, the predictive model learning unit 120 may calculate the sum of signals of all frames by extracting only multiples of the fundamental frequency.

For example, the predictive model learner 120 may quantify the similarity between the first frame and the second frame using Equation 6. Through this, the continuity of harmonic components of the fundamental frequency of the unvoiced frame and the voiced frame can be improved.

The predictive model learning unit 120 may calculate a voice presence probability value using an absolute value of an autocorrelation value. The predictive model learner 120 may calculate a similarity function indicating similarity between the first frame and the second frame, which is a previous frame of the first frame, based on the absolute value of the autocorrelation value and the cross-correlation ratio. For example, the predictive model learning unit 120 may calculate the similarity between the first frame and the second frame using Equation 8 below.

<Equation 8>

은 유사성 함수이고,

는 수학식 2에서 산출한 자기 상관값에 절대값을 적용한 값일 수 있다. 예를 들어, 예측 모델 학습부(120)는 하기 수학식 9를 이용하여 자기 상관값의 절대값, 즉, 기본 주파수의 상관값에 절대값을 산출할 수 있다.In Equation 8,

is the similarity function,

May be a value obtained by applying an absolute value to the autocorrelation value calculated in Equation 2. For example, the predictive model learner 120 may calculate an absolute value of an autocorrelation value, that is, an absolute value of a correlation value of a fundamental frequency using Equation 9 below.

<Equation 9>

의 값이 클수록 관계성이 높은 것으로 판단할 수 있다. 즉, 예측 모델 학습부(120)는 특정 프레임에서 도출한 기본 주파수가 제대로 도출한 기본 주파수인지 수치를 통해 확인할 수 있다.Referring to Equation 9, the predictive model learning unit 120 may measure the degree of association by applying an absolute value to the autocorrelation value. for example,

The higher the value of is, the higher the relationship can be determined. That is, the predictive model learning unit 120 may check through numerical values whether the fundamental frequency derived from a specific frame is properly derived.

Referring to Equations 8 and 9, the prediction model learning unit 120 calculates the similarity between the first frame and the second frame by dividing the absolute value of the second frame by the product of the absolute value of the first frame and the cross-correlation ratio. can do.

The predictive model learning unit 120 may calculate a voice presence probability value with a value between 0 and 1 by applying a sigmoid function to the similarity function. For example, the predictive model learning unit 120 may apply a sigmoid function to the similarity function using Equation 10 below.

<Equation 10>

Referring to Equation 10, the predictive model learning unit 120 applies a sigmoid function to a similarity function between a first frame and a second frame included in speech data to finally calculate a speech presence probability value for the first frame. can do. For example, the predictive model learning unit 120 may apply a sigmoid function to the similarity function to quantify the probability that a human voice exists in the first frame as a value between 0 and 1.

In this way, the voice recognition apparatus 100 according to the present invention calculates the presence or absence of a human voice in each section of voice data, that is, each frame, as a probability value between 0 and 1, rather than a dichotomy value of 0 or 1. can do. Accordingly, the voice recognition apparatus 100 can effectively recognize a human voice even in a voice signal in an environment in which noise is included in voice data or a leading consonant is cut off in a specific frame.

The voice recognition unit 130 may recognize voice using a voice recognition model. For example, the voice recognition unit 130 may recognize a human voice included in the corresponding voice data by using a voice recognition model learned based on a feature vector of the voice data and a voice presence probability value.

Also, the speech recognition unit 130 may recognize speech included in speech data and detect sentences by combining a speech recognition model learned as a joint vector and a language model. For example, the voice recognition unit 130 may use Equation 11 below.

<Equation 11>

는 조인트 벡터이고,

는 조인트 벡터로 학습된 음향(Acoustic) 모델이고,

는 언어 모델이고,

는 입력된 음성 데이터의 최대사후확률을 갖는 문장일 수 있다. 수학식 11을 참조하면, 음성 인식부(130)는 음성 인식 모델 및 언어 모델을 통해 음성 데이터에 포함된 문장을 인식하고 검출할 수 있다.In Equation 11,

is the joint vector,

is an acoustic model learned as a joint vector,

is the language model,

may be a sentence having the maximum posterior probability of input speech data. Referring to Equation 11, the voice recognition unit 130 may recognize and detect a sentence included in voice data through a voice recognition model and a language model.

2 is an exemplary diagram for explaining a delay time and periodicity of a specific frame according to an embodiment of the present invention.

를 이용할 수 있다. Referring back to Equation 1, the voice recognition apparatus 100 may generate a joint vector using a feature vector extracted from voice data and a voice presence probability vector. In Equation 1, the speech recognition apparatus 100 includes a 3-dimensional speech presence probability vector,

is available.

For example, the voice presence probability vector of the first frame may include a voice presence probability value of the first frame, a delay time value 210 of the first frame, and a delta delay time value 220 of the first frame.

The voice recognition apparatus 100 may calculate the delay time value 210 for a specific frame, for example, the first frame, using Equation 12 below.

<Equation 12>

In addition, the voice recognition apparatus 100 may calculate the delta delay time value 220 for a specific frame, for example, the first frame, using Equation 13 below.

<Equation 13>

3 is an exemplary diagram for explaining an example of a delay time value according to an embodiment of the present invention. In (a) of FIG. 3, the signal of the first frame and the signal of the time delayed by 2 from the signal of the first frame show the most similar pattern, and (b) shows the signal of the first frame and the signal of the first frame. A signal with a delayed time of 5 shows the most similar pattern.

In (c) of FIG. 3, the signal of the first frame and the signal of the time delayed by 10 from the signal of the first frame show the most similar pattern, and (d) shows the signal of the first frame and the signal of the first frame. A signal with a delayed time of 18 shows the most similar pattern.

Therefore, (a) of FIG. 3 may have a delay time value of 2, (b) may have a delay time value of 5, (c) may have a delay time value of 10, and (d) may have a delay time value of 18. . That is, the periodicity of the example shown in (a) of FIG. 3 is 2, the periodicity of the example shown in (b) is 5, the periodicity of the example shown in (c) is 10, and the periodicity of the example shown in (d) is An exemplary periodicity may be 18.

4 is an exemplary diagram for explaining a process of deriving a delay time value according to an embodiment of the present invention. Referring to FIG. 4 , a delay time value may be derived from a value 410 having the highest correlation among an input signal and a signal past the input signal.

For example, the voice recognition apparatus 100 may derive a delay time value based on a signal having the highest similarity between a signal of the first frame and a signal of a delay time from the first frame.

That is, the voice recognition apparatus 100 may derive the periodicity of the first frame based on the interval between the signal of the first frame and the signal having the highest similarity among the signals of the delayed time from the first frame.

5 is an exemplary diagram for explaining a gamma filter according to an embodiment of the present invention. Referring to FIG. 5 , the voice recognition apparatus 100 may apply a gamma filter to voice data ((a) of FIG. 5 ) to obtain data separated into time domain and frequency domain ((b) of FIG. 5 ). there is.

Referring back to Equation 6, the voice recognition apparatus 100 may derive the number of channels according to the fundamental frequency, and calculate the cross-correlation ratio based on the sum of samples corresponding to the derived number of channels. In this case, the speech recognition apparatus 100 may perform convolution by applying a gamma filter to samples corresponding to the number of channels and the input signal.

)을 도출할 수 있고, 10개의 샘플의 합에 기초할 수 있다. 다른 예를 들어, 80~1000Hz이고, 도출된 기본 주파수가 200일 때, 음성 인식 장치(100)는 기본 주파수를 포함한 5개의 채널(수학식 6에서

)을 도출할 수 있고, 5개의 샘플의 합에 기초할 수 있다. 이 때, 음성 인식 장치(100)는 감마 필터를 적용하여, 해당 채널, 프레임에 존재하는 감마 성분들에 대한 전체 합을 산출할 수 있다.For example, when 80 to 1000 Hz and the derived fundamental frequency is 100, the voice recognition apparatus 100 transmits 10 channels including the basic frequency (in Equation 6).

) can be derived, and can be based on the sum of 10 samples. As another example, when 80 to 1000 Hz and the derived fundamental frequency is 200, the voice recognition apparatus 100 has 5 channels including the basic frequency (in Equation 6).

) can be derived, and can be based on the sum of five samples. At this time, the speech recognition apparatus 100 may apply a gamma filter to calculate the total sum of gamma components existing in the corresponding channel and frame.

6 is a flowchart of a voice recognition method according to an embodiment of the present invention. The method for recognizing a voice shown in FIG. 6 includes steps processed time-sequentially according to the embodiment shown in FIGS. 1 to 5 . Therefore, even if the content is omitted below, it is also applied to the method of recognizing a voice in the voice recognition apparatus according to the exemplary embodiment shown in FIGS. 1 to 5 .

In step S610, the voice recognition device may extract a feature vector from voice data.

In step S620, the speech recognition apparatus may train a speech recognition model based on a feature vector and a joint vector including a speech existence probability value for regular speech data.

In step S630, the voice recognition device may recognize voice using the voice recognition model.

In the above description, steps S610 to S630 may be further divided into additional steps or combined into fewer steps, depending on an implementation example of the present invention. Also, some steps may be omitted as needed, and the order of steps may be switched.

The method of recognizing a voice in the voice recognition apparatus described with reference to FIGS. 1 to 6 is implemented in the form of a computer program stored in a computer-readable recording medium executed by a computer or a recording medium including instructions executable by a computer. It can be. Also, the method of recognizing a voice in the voice recognition apparatus described with reference to FIGS. 1 to 6 may be implemented in the form of a computer program stored in a computer-readable recording medium executed by a computer.

Computer readable media can be any available media that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. Also, computer readable recording media may include computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

The above description of the present invention is for illustrative purposes, and those skilled in the art can understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, the embodiments described above should be understood as illustrative in all respects and not limiting. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined form.

The scope of the present invention is indicated by the following claims rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and equivalent concepts should be construed as being included in the scope of the present invention. do.

100: voice recognition device
110: feature vector extraction unit
120: prediction model learning unit
130: voice recognition unit

Claims (19)

In the device for recognizing voice,
a feature vector extractor extracting feature vectors from voice data;
a predictive model learning unit that trains a speech recognition model based on the feature vector and a joint vector including a speech presence probability value for regular speech data; and
A voice recognition unit for recognizing a voice using the voice recognition model
To include, voice recognition device.
According to claim 1,
The predictive model learning unit,
and calculating a voice presence probability value for each frame included in the voice data based on an autocorrelation value and a cross correlation ratio (CCR) of the voice data.
According to claim 2,
The predictive model learning unit,
and a similarity between a first frame included in the voice data and a second frame that is a previous frame of the first frame is calculated as the cross-correlation ratio.
According to claim 3,
The predictive model learning unit,
And deriving the number of channels according to the basic frequency of the first frame.
According to claim 4,
The predictive model learning unit,
A ratio between the sum of at least one sample corresponding to the first frame and the sum of at least one sample corresponding to the second frame based on the derived number of channels as a similarity between the first frame and the second frame A voice recognition device that calculates.
According to claim 2,
The predictive model learning unit,
and calculating the voice existence probability value using an absolute value of the autocorrelation value.
According to claim 6,
The predictive model learning unit,
and calculating a similarity function indicating a similarity between a first frame and a second frame that is a previous frame of the first frame based on the absolute value of the autocorrelation value and the cross-correlation ratio.
According to claim 7,
The predictive model learning unit,
And calculating the voice presence probability value as a value between 0 and 1 by applying a sigmoid function to the similarity function.
According to claim 4,
The predictive model learning unit,
And deriving the fundamental frequency based on the autocorrelation function.
In the voice recognition method,
extracting feature vectors from voice data;
learning a speech recognition model based on the feature vector and a joint vector including a speech existence probability value for regular speech data; and
Recognizing a voice using the voice recognition model
To include, voice recognition method.
According to claim 10,
The step of learning the voice recognition model,
Calculating a voice presence probability value for each frame included in the voice data based on an autocorrelation value and a cross correlation ratio (CCR) of the voice data
To include, voice recognition method.
According to claim 11,
The step of learning the voice recognition model,
Calculating a similarity between a first frame included in the voice data and a second frame that is a previous frame of the first frame as the cross-correlation ratio
To further include, the voice recognition method.
According to claim 12,
The step of learning the voice recognition model,
Deriving the number of channels according to the basic frequency of the first frame
To further include, the voice recognition method.
According to claim 13,
The step of learning the voice recognition model,
A ratio between the sum of at least one sample corresponding to the first frame and the sum of at least one sample corresponding to the second frame based on the derived number of channels as a similarity between the first frame and the second frame steps to calculate
To further include, the voice recognition method.
According to claim 11,
The step of learning the voice recognition model,
Calculating the negative presence probability value using the absolute value of the autocorrelation value
To further include, the voice recognition method.
According to claim 15,
The step of learning the voice recognition model,
Calculating a similarity function indicating similarity between a first frame and a second frame that is a previous frame of the first frame based on the absolute value of the autocorrelation value and the cross-correlation ratio
To further include, the voice recognition method.
17. The method of claim 16,
The step of learning the voice recognition model,
Calculating the negative presence probability value as a value between 0 and 1 by applying a sigmoid function to the similarity function
To further include, the voice recognition method.
According to claim 13,
The step of learning the voice recognition model,
Deriving the fundamental frequency based on the autocorrelation function
To further include, the voice recognition method.
A computer program stored in a computer readable recording medium containing a sequence of commands for recognizing voice,
When the computer program is executed by a computing device,
extract feature vectors from speech data;
Learning a speech recognition model based on the feature vector and a joint vector including a speech presence probability value for regular speech data;
A computer program stored on a computer-readable recording medium comprising a sequence of commands for recognizing a voice using the voice recognition model.

Images