APPARATUS AND METHOD FOR GENERATING SPEECH SYNTHESIS IMAGE

Abstract

An apparatus for generating a speech synthesis image according to a disclosed embodiment is an apparatus for generating a speech synthesis image based on machine learning, the apparatus including a first global geometric transformation predictor configured to be trained to receive each of a source image and a target image including the same person, and predict a global geometric transformation for a global motion of the person between the source image and the target image based on the source image and the target image, a local feature tensor predictor configured to be trained to predict a feature tensor for a local motion of the person based on input target image-related information, and an image generator configured to be trained to reconstruct the target image based on the global geometric transformation, the source image, and the feature tensor for the local motion.

Claims

1. An apparatus for generating a speech synthesis image based on machine learning, the apparatus comprising: a first global geometric transformation predictor configured to be trained to receive each of a source image and a target image including the same person, and predict a global geometric transformation for a global motion of the person between the source image and the target image, based on the source image and the target image; a local feature tensor predictor configured to be trained to predict a feature tensor for a local motion of the person, based on input target image-related information; and an image generator configured to be trained to reconstruct the target image, based on the global geometric transformation, the source image, and the feature tensor for the local motion.

2. The apparatus according to claim 1, wherein the global motion is a motion of the person with an amount greater than or equal to a preset threshold amount of motion, and the local motion is a motion of a face when the person is speaking.

3. The apparatus according to claim 2, wherein the first global geometric transformation predictor is further configured to extract a geometric transformation into a source image heat map from a preset reference probability distribution, based on the source image, extract a geometric transformation into a target image heat map from the preset reference probability distribution, based on the target image, and calculate the global geometric transformation, based on the geometric transformation into the source image heat map from the reference probability distribution and the geometric transformation into the target image heat map from the reference probability distribution.

4. The apparatus according to claim 3, wherein the source image heat map is a probability distribution map in an image space as to whether each pixel in the source image is a pixel related to the global motion of the person, and the target image heat map is a probability distribution map in the image space as to whether each pixel in the target image is a pixel related to the global motion of the person.

5. The apparatus according to claim 2, wherein the local feature tensor predictor includes a first local feature tensor predictor configured to be trained to predict a speech feature tensor for a local speech motion of the person, based on the input target image-related information, and the local speech motion is a motion related to a speech of the local motion of the person.

6. The apparatus according to claim 5, wherein the local feature tensor predictor further includes a second local feature tensor predictor configured to be trained to predict a non-speech feature tensor for a local non-speech motion of the person, based on the input target image-related information, and the local non-speech motion is a motion not related to the speech of the local motion of the person.

7. The apparatus according to claim 6, wherein the second local feature tensor predictor is trained to receive a target partial image including only a motion not related to speech of the person in the target image, and predict the non-speech feature tensor, based on the target partial image.

8. The apparatus according to claim 2, further comprising an optical flow predictor configured to be trained to calculate an optical flow between the source image and the target image, based on the source image and the global geometric transformation, wherein the image generator is trained to reconstruct the target image, based on the optical flow between the source image and the target image, the source image, and the feature tensor for the local motion.

9. The apparatus according to claim 1, wherein the first global geometric transformation predictor is further configured to calculate a geometric transformation into any i-th (i?{1, 2, . . . , n}) (n is a natural number equal to or greater than 2) frame heat map in an image having n frames from a preset reference probability distribution when the image is input, and calculate a global geometric transformation between two adjacent frames in the image, based on the geometric transformation into the i-th frame heat map from the reference probability distribution.

10. The apparatus according to claim 9, further comprising a second global geometric transformation predictor configured to receive sequential voice signals corresponding to the n frames, and to be trained to predict a global geometric transformation between two adjacent frames in the image from the sequential voice signals.

11. The apparatus according to claim 10, wherein the second global geometric transformation predictor is further configured to adjust a parameter of an artificial neural network to minimize a difference between the global geometric transformation between the two adjacent frames which is predicted in the second global geometric transformation predictor and the global geometric transformation between the two adjacent frames which is calculated in the first global geometric transformation predictor.

12. The apparatus according to claim 11, wherein in a test process for speech synthesis image generation, the second global geometric transformation predictor is further configured to receive sequential voice signals of a person, calculate a global geometric transformation between two adjacent frames in an image corresponding to the sequential voice signals from the sequential voice signals, and calculate a global geometric transformation between a preset target frame and a preset start frame, based on the global geometric transformation between the two adjacent frames, the local feature tensor predictor is further configured to predict the feature tensor for the local motion of the person, based on the input target image-related information, and the image generator is further configured to reconstruct the target frame, based on the global geometric transformation, the source image, and the feature tensor for the local motion.

13. A method for generating a speech synthesis image, based on machine learning that is performed in a computing device including one or more processors and a memory storing one or more programs executed by the one or more processors, the method comprising: training a first global geometric transformation predictor to receive each of a source image and a target image including the same person, and predict a global geometric transformation for a global motion of the person between the source image and the target image, based on the source image and the target image; training a local feature tensor predictor to predict a feature tensor for a local motion of the person, based on input target image-related information; and training an image generator to reconstruct the target image, based on the global geometric transformation, the source image, and the feature tensor for the local motion.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0022] FIG. 1 is a block diagram illustrating a configuration of an apparatus for generating a speech synthesis image according to an embodiment of the present invention.

[0023] FIG. 2 is a block diagram illustrating a configuration of an apparatus for generating a speech synthesis image according to another embodiment of the present invention.

[0024] FIG. 3 is a diagram illustrating a configuration for estimating a global geometric transformation of a person from a sequential voice signal in an embodiment of the present invention.

[0025] FIG. 4 is a block diagram illustrating a configuration of an apparatus for generating a speech synthesis image according to still another embodiment of the present invention.

[0026] FIG. 5 is a block diagram illustratively describing a computing environment including a computing device suitable for use in exemplary embodiments.

DETAILED DESCRIPTION

[0027] Hereinafter, a specific embodiment of the present invention will be described with reference to the drawings. The following detailed description is provided to aid in a comprehensive understanding of the methods, apparatus and/or systems described herein. However, this is illustrative only, and the present invention is not limited thereto.

[0028] In describing the embodiments of the present disclosure, when it is determined that a detailed description of related known technologies related to the present invention can unnecessarily obscure the subject matter of the present invention, a detailed description thereof will be omitted. In addition, terms to be described later are terms defined in consideration of functions in the present invention, which can vary according to the intention or custom of users or operators. Therefore, the definition should be made based on the contents throughout this specification. The terms used in the detailed description are only for describing embodiments of the present invention, and should not be limiting. Unless explicitly used otherwise, expressions in the singular form include the meaning of the plural form. In this description, expressions such as comprising or including are intended to refer to certain features, numbers, steps, actions, elements, some or combination thereof, and it is not to be construed to exclude the presence or possibility of one or more other features, numbers, steps, actions, elements, some or combinations thereof, other than those described.

[0029] In addition, terms such as the first and second can be used to describe various components, but the components should not be limited by the terms. The above terms can be used for the purpose of distinguishing one component from another component. For example, without departing from the scope of the present invention, a first component can be referred to as a second component, and similarly, the second component can also be referred to as the first component.

[0030] In the embodiments disclosed herein, the speech synthesis image is an image obtained by synthesizing a speech scene of a person through a machine learning model, and may also be referred to as a lip-sync image. The speech synthesis image may be an upper body image including the head and chest of a person, but is not limited thereto.

[0031] In addition, a global motion may refer to a large motion of the person in the image in the overall frame. When the speech synthesis image is an upper body image, a global motion may refer to the motion of the entire upper body of the person in the image (e.g., the motion such as changing the posture of the upper body of the person in the image or turning the head of the person in the image). The global motion is a motion with an amount greater than or equal to a preset threshold amount, and the threshold amount may be set to represent a large motion of the person in the overall frame.

[0032] In addition, a local motion may refer to a facial motion when a person in the image is speaking. That is, the local motion may refer to a change in facial expression, mouth and jaw motions, or the like that appear on the person's face when the person in the image is speaking. The local motion may be a motion with an amount below a threshold amount.

[0033] In the embodiments disclosed herein, at the time of generating a speech synthesis image, generation of the global motion of a person in the image may be controlled by performing separation and estimation from the input image using a geometric transformation bottleneck, and generation of the local motion of the person in the image may be controlled using an input speech voice or the like.

[0034] Specifically, when the full motion of a specific object in the image consists of a combination of motions of N independent elements, N geometric transformations may be required to fully estimate the full motion. Here, when the motions of the N independent elements are quantitatively different (for example, when the areas or volumes of parts of the object in the image to which the motions are applied are different or the sizes of the motions are different, and so on), the error associated with the motion of each element is proportional to the amount of the motion of the element.

[0035] In this case, when the motion of an object is estimated through the machine learning model using K geometric transformations, where K is a number smaller than N (that is, K<N), in a process in which learning about the part with a large error among the motions of the elements is prioritized in the machine learning model, the K geometric transformations are derived to capture the element motion that is the largest.

[0036] Therefore, when a bottleneck is formed by limiting the number of geometric transformations and element transformations constituting the geometric transformation, it is possible to separate and estimate the global motion of the person in the image (that is, large motion such as the motion of the head and torso of the person) from the local motion with a relatively small motion size (that is, facial motion when a person is speaking). The global motion may be a motion of a single element having the largest motion size, or may be a set of motions of a plurality of elements having a motion size greater than or equal to a preset size.

[0037] In an exemplary embodiment, a single geometric transformation consisting of element transformations such as parallel translation, rotation, and scaling may be used to capture the global motion of a person in an image. Here, the parallel translation may be used to capture the overall movement of the upper body of the person in the image. Rotation and horizontal scaling may be used to capture the changes caused by the rotation of the head of the person in the image. The vertical scaling may be used to capture the vertical length change of the entire head and torso caused by the person in the image raising or lowering his or her head.

[0038] For another example, the motion of the head of the person in the image may be captured through a geometric transformation consisting of parallel translation and rotation, and the motion of the torso of the person in the image may be captured through another geometric transformation consisting of only parallel translation.

[0039] For another example, when a speech synthesis image includes only the head and the upper part of the neck of the person, the position of the neck is dependent on the head motion, and thus, the motion of the head of the person in the image may be captured using a single geometric transformation consisting of parallel translation and rotation.

[0040] Meanwhile, the local motion of the person in the image may be divided into two motions. That is, the local motion may be divided into a motion related to speech (e.g., motions of the mouth (including lips) and jaw of the person) and a motion not related to speech (e.g., blinking of the eyes, eyebrow motion, frown, or the like, of the person). Hereinafter, the motion related to speech among the local motions may be referred to as a local speech motion. The motion not related to speech among the local motions may be referred as a local non-speech motion.

[0041] Here, for the local speech motion, an artificial neural network that uses image information (target image, etc.) of a person as input is added to the machine learning model, and a feature tensor for motion related to speech (local speech motion) is allowed to be output from the added artificial neural network, and the output feature tensor may be used as an input when generating a speech synthesis image.

[0042] In addition, for the local non-speech motion, the artificial neural network that uses an image including only motions not related to the speech of the person (for example, image including only the areas around the eyes and eyebrows of the person) (or a feature point showing only the non-speech motion of the person) as an input is added to the machine learning model and a feature tensor for the motion related to non-speech (local non-speech motion) from the added artificial neural network is output, whereby the output feature tensor may be used as an input at the time of generating a speech synthesis image.

[0043] FIG. 1 is a block diagram illustrating a configuration of an apparatus for generating a speech synthesis image according to an embodiment of the present invention.

[0044] Referring to FIG. 1, a speech synthesis image generating apparatus 100 may include a first global geometric transformation predictor 102, a first local feature tensor predictor 104, an optical flow predictor 106, and an image generator 108. The speech synthesis image generating apparatus 100 may be implemented based on a machine learning technology. Hereinafter, a training process for generating a speech synthesis image will be mainly described.

[0045] The first global geometric transformation predictor 102 may receive each of a source image I.sub.s and a target image I.sub.d. Here, the source image I.sub.s and the target image I.sub.d are a pair of images including the same person, and the speech synthesis image generating apparatus 100 may include an artificial neural network for generating the target image I.sub.d as the speech synthesis image by using the source image I.sub.s as an input.

[0046] The source image I.sub.s and the target image I.sub.d may be a video part of an image (that is, including video and audio) in which a person is speaking. The source image I.sub.s and the target image I.sub.d may be images including the face and upper body of the person, but are not limited thereto.

[0047] The first global geometric transformation predictor 102 may calculate a geometric transformation (hereinafter, may be referred to as a global geometric transformation) for a global motion between the source image I.sub.s and the target image I.sub.d. That is, the first global geometric transformation predictor 102 may calculate a global geometric transformation capable of expressing a difference in the global motion of the person between the source image I.sub.s and the target image I.sub.d (that is, a large motion such as the motion of the head and torso of the person). Hereinafter, the first global geometric transformation predictor 102 is described as calculating the global geometric transformation into the source image I.sub.s from the target image I.sub.d by way of example, but is not limited thereto, and may also calculate the global geometric transformation into the target image I.sub.d from the source image I.sub.s.

[0048] Specifically, the first global geometric transformation predictor 102 may receive each of the source image I.sub.s and the target image I.sub.d, and may extract heat maps for the source image I.sub.s and the target image I.sub.d. That is, the first global geometric transformation predictor 102 may extract a heat map H.sub.s (source image heat map) for the source image I.sub.s from the source image I.sub.s. The first global geometric transformation predictor 102 may extract a heat map Ha (target image heat map) for the target image I.sub.d from the target image I.sub.d.

[0049] In an exemplary embodiment, the first global geometric transformation predictor 102 may be constituted by an artificial neural network based on a convolutional neural network (CNN), but is not limited thereto. The first global geometric transformation predictor 102 may extract each of the source image heat map H.sub.s and the target image heat map Ha through Equation 1 below.

[00001]$\begin{array}{cc}{H}_s=F^{\mathrm{global}}\left(I_s\right)& \left(\mathrm{Equation}1\right)\end{array}$$H_d=F^{\mathrm{global}}\left(I_d\right)$ [0050] F.sup.global: Artificial neural network constituting the first global geometric transformation predictor 102

[0051] Here, each of the source image heat map H.sub.s and the target image heat map Ha may be a map represented by a probability distribution in an image space. That is, the source image heat map H.sub.s may be a probability distribution map in the image space as to whether each pixel in the source image I.sub.s is a pixel related to the global motion of the person. The target image heat map Ha may be a probability distribution map in the image space as to whether each pixel in the target image I.sub.d is a pixel related to the global motion of the person. In order to achieve the above, an output end of the first global geometric transformation predictor 102 may include a 2D softmax layer.

[0052] The first global geometric transformation predictor 102 may calculate each of a probability mean ?.sub.s of the source image heat map H.sub.s and a probability mean ?.sub.d of the target image heat map H.sub.d through Equation 2.

[00002]$\begin{array}{cc}{?}_s={?}_z{H}_s\left(z\right)z& \left(\mathrm{Equation}2\right)\end{array}$${?}_d={?}_z{H}_d\left(z\right)z$ [0053] z: Position value of a pixel in the image space (coordinates) [0054] H.sub.s(z): Probability value of the pixel at a position z in the source image heatmap H.sub.s [0055] H.sub.d(z): Probability value of the pixel at a position z in the target image heatmap H.sub.d

[0056] The first global geometric transformation predictor 102 may calculate a covariance matrix of the source image heat map H.sub.s based on the probability mean us of the source image heat map H.sub.s, and may calculate a covariance matrix of the target image heat map Ha based on the probability mean ?.sub.d of the target image heat map H.sub.d. The first global geometric transformation predictor 102 may calculate a covariance matrix K.sub.s of the source image heat map H.sub.s and a covariance matrix K.sub.d of the target image heat map H.sub.d through Equation 3.

[00003]$\begin{array}{cc}{K}_s={?}_z{H}_s\left(z\right)\left(z-{?}_s\right){\left(z-{?}_s\right)}^T& \left(\mathrm{Equation}3\right)\end{array}$$K_d={?}_z{H}_d\left(z\right)\left(z-{?}_d\right){\left(z-{?}_d\right)}^T$ [0057] T: Transpose matrix

[0058] Here, the covariance matrix K.sub.s of the source image heat map H.sub.s and the covariance matrix K.sub.d of the target image heat map H.sub.d may be decomposed as in Equation 4 below through singular value decomposition, respectively.

[00004]$\begin{array}{cc}{K}_s=U_s{?}_s{V}_s& \left(\mathrm{Equation}4\right)\end{array}$$K_d=U_d{?}_d{V}_d$ [0059] U.sub.s, U.sub.d: Unitary matrices [0060] ?.sub.s, ?.sub.d: Diagonal matrices [0061] V.sub.s, V.sub.d: Unitary matrices

[0062] When the covariance matrix K.sub.s of the image heat map H.sub.s and the covariance matrix K.sub.d of the target image heat map H.sub.d are each m?n matrices, U.sub.s and U.sub.d may be unitary matrices having a size of m?m, ?.sub.s and ?.sub.d may be diagonal matrices having a size of m?n, and V.sub.s and V.sub.d may be unitary matrices having a size of n?n.

[0063] The first global geometric transformation predictor 102 may calculate a geometric transformation into the source image heat map H.sub.s from a preset reference probability distribution H.sub.r based on the unitary matrix U.sub.s and the diagonal matrix ?.sub.s according to the singular value decomposition of the covariance matrix K.sub.s of the source image heat map H.sub.s, and the probability mean us of the source image heat map H.sub.s. Here, the preset reference probability distribution H.sub.r may be a probability distribution in which a probability mean is 0, the covariance matrix is an identity matrix, and the main axis is aligned with an image axis.

[0064] The first global geometric transformation predictor 102 may calculate a geometric transformation A.sub.s?r.sup.global into the source image heat map H.sub.s from the preset reference probability distribution H.sub.r through Equation 5 below.

[00005]$\begin{array}{cc}{A}_{s?r}^{\mathrm{global}}=\left[U_s{{?}_s}^{\frac{1}{2}},{?}_s\right]& \left(\mathrm{Equation}5\right)\end{array}$

[0065] Further, the first global geometric transformation predictor 102 may calculate a geometric transformation into the target image heat map H.sub.d from the preset reference probability distribution H.sub.r based on the unitary matrix U.sub.d and the diagonal matrix ?.sub.d according to the singular value decomposition of the covariance matrix K.sub.d of the target image heat map H.sub.d, and the probability mean ?.sub.d of the target image heat map H.sub.d. The first global geometric transformation predictor 102 may calculate a geometric transformation global A.sub.d?r.sup.global into the target image heat map H.sub.d from the preset reference probability distribution H.sub.r through Equation 6 below.

[00006]$\begin{array}{cc}{A}_{d?r}^{\mathrm{global}}=\left[U_d{.Math.}_d^{\frac{1}{2}},{?}_d\right]& \left(\mathrm{Equation}6\right)\end{array}$

[0066] Meanwhile, it has been described here that the artificial neural network of the first global geometric transformation predictor 102 receives the source image I.sub.s and the target image I.sub.d to extract the source image heat map H.sub.s and the target image heat map H.sub.d, respectively, and the subsequent process is performed through calculations, but the embodiment is not limited thereto, and the artificial neural network of the first global geometric transformation predictor 102 may receive the source image I.sub.s and the target image I.sub.d to extract the geometric transformation A.sub.s?r.sup.global into the source image heat map H.sub.s from the preset reference probability distribution H.sub.r and the geometric transformation A.sub.d?r.sup.global into the target image heat map H.sub.d from the preset reference probability distribution H.sub.r, respectively.

[0067] The first global geometric transformation predictor 102 may calculate a global geometric transformation into the source image I.sub.s from the target image I.sub.s based on the geometric transformation A.sub.s?r.sup.global into the source image heat map H.sub.s from the reference probability distribution H.sub.r and the geometric transformation A.sub.d?r.sup.global into the target image heat map H.sub.d from the reference probability distribution H.sub.r. The first global geometric transformation predictor 102 may calculate a global geometric transformation A.sub.s?d.sup.global into the source image I.sub.s from the target image I.sub.d through Equation 7 below.

[00007]$\begin{array}{cc}{A}_{s?d}^{\mathrm{global}}=A_{s?r}^{\mathrm{global}}{A}_{d?r}^{\mathrm{globa}{l}^{-1}}& \left(\mathrm{Equation}7\right)\end{array}$

[0068] Meanwhile, it has been described here that the neural network of the first global geometric transformation predictor 102 receives the source image I.sub.s and the target image I.sub.d to extract the source image heat map H.sub.s and the target image heat map H.sub.d, respectively (that is, calculate the global geometric transformation based on the heat map), but the embodiment is not limited thereto, and a method of directly estimating the global geometric transformation from the source image I.sub.s and target image I.sub.d without a heatmap by using an artificial neural network such as an encoder-predictor structure may also be used.

[0069] The first local feature tensor predictor 104 may include an artificial neural network for estimating the local motion of the person in the speech synthesis image. In an exemplary embodiment, the artificial neural network may be trained to estimate a local speech motion of the person (motion related to speech, such as motions of the mouth and jaw of the person) from input target image-related information.

[0070] Specifically, the first local feature tensor predictor 104 may receive a target image I.sub.d. Here, the target image I.sub.d may be the same as that input to the first global geometric transformation predictor 102. Here, it is described that the target image-related information is the target image I.sub.d, but is not limited thereto. The target image-related information may include feature point information extracted from the target image I.sub.d or a partial image of the target image I.sub.d.

[0071] The first local feature tensor predictor 104 may estimate a feature tensor for the local motion of the person from the target image I.sub.d. That is, the first local feature tensor predictor 104 may estimate a feature tensor (hereinafter, may be referred to as a speech feature tensor) capable of expressing the local speech motion based on the target image I.sub.d when a person utters a voice.

[0072] The first local feature tensor predictor 104 may include an artificial neural network for estimating the speech feature tensor from the target image I.sub.d. The first local feature tensor predictor 104 may estimate a speech feature tensor z.sub.d from the target image I.sub.d through Equation 8 below. The shape and size of the speech feature tensor z.sub.d may be appropriately set according to the degree required for controlling a speech action.

[00008]$\begin{array}{cc}{z}_d=F^{\mathrm{local}}\left(I_d\right)& \left(\mathrm{Equation}8\right)\end{array}$ [0073] F.sup.local: artificial neural network constituting the first local feature tensor predictor 104

[0074] Meanwhile, here, the description is made in such a way that the speech feature tensor is estimated by using the target image as an input, but is not limited thereto. The speech feature tensor may be estimated by using a feature point extracted from the target image as an input.

[0075] In addition, the first local feature tensor predictor 104 may estimate the speech feature tensor by using a partial image (e.g., an image in which parts of the target image are masked or cropped, etc.), which contains only a part of the source image I.sub.s and the target image I.sub.d rather than the entire image thereof, as an input.

[0076] The optical flow predictor 106 may calculate an optical flow representing a motion (or change amount) in units of pixels between the source image I.sub.s and the target image I.sub.d by using global geometric transformation calculated by the first global geometric transformation predictor 102 and the source image I.sub.s as inputs.

[0077] In an exemplary embodiment, when the first global geometric transformation predictor 102 calculates a global geometric transformation A.sub.s?d.sup.global into the source image I.sub.s from the target image I.sub.d, the optical flow predictor 106 may calculate the optical flow from the target image I.sub.d to the source image I.sub.s based on the global geometric transformation A.sub.s?d.sup.global and the source image I.sub.s, which will be described below. However, the embodiment is not limited thereto, and the optical flow from the source image I.sub.s to the target image I.sub.d may be calculated.

[0078] Specifically, the optical flow predictor 106 may transform the source image I.sub.s by applying the global geometric transformation A.sub.s?d.sup.global into the source image I.sub.s by using an image warping operator. In this case, the optical flow predictor 106 may transform the source image I.sub.s through Equation 9 below.

[00009]$\begin{array}{cc}{I}_{s^{}}=\mathrm{warp}\left(I_s,A_{s?d}^{\mathrm{global}}\right)& \left(\mathrm{Equation}9\right)\end{array}$ [0079] I.sub.s: Transformed source image I.sub.s [0080] warp( ): Operator for image warping

[0081] In Equation 9, for the operator for image warping, a backward warping operation may be used that calculates coordinates of the source image I.sub.s corresponding to coordinates of the transformed source image I.sub.s by applying each of global geometric transformations A.sub.s?d.sup.global to the coordinates of the transformed source image I.sub.s, and estimates pixel values of the transformed source image I.sub.s from pixel values of the source image I.sub.s using interpolation.

[0082] The optical flow predictor 106 may calculate a weighted probability distribution map for estimating the optical flow based on the transformed source image I.sub.s. In this case, the optical flow predictor 106 may calculate a weighted probability distribution map P having two classes for each pixel by inputting the transformed source image I.sub.s to the artificial neural network. This may be expressed by Equation 10 below.

[00010]$\begin{array}{cc}P=F^{\mathrm{flow}}\left(I_{s^{}},I_s\right)& \left(\mathrm{Equation}10\right)\end{array}$ [0083] F.sup.flow: Artificial neural network for estimating the optical flow in the optical flow predictor 106

[0084] Here, the artificial neural network may include a one-dimensional softmax layer at an output end to calculate the weighted probability distribution map P.

[0085] Meanwhile, here, the transformed source image I.sub.s is used as an input to the artificial neural network F.sup.flow, but the embodiment is not limited thereto and the weighted probability distribution map P may be calculated by using a feature tensor extracted from the source image I.sub.s as an input to the artificial neural network F.sup.flow.

[0086] The optical flow predictor 106 may calculate the optical flow from the target image I.sub.d to the source image I.sub.s for each pixel by linearly combining the global geometric transformation A.sub.s?d.sup.global and an identity geometric transformation A.sub.s?d.sup.global identity by using a weighed probability distribution value corresponding to each of pixel positions of the transformed source image I.sub.s. Here, the identity geometric transformation may be a transformation in which both before and after transformation are identical, for example, a geometric transformation consisting of 0 for parallel translation, the identity matrix for rotation transformation, and 1 for scaling.

[0087] The optical flow predictor 106 may calculate the optical flow from the target image I.sub.d to the source image I.sub.s for each pixel through Equation 11.

[00011]$\begin{array}{cc}{f}_{s?d}\left(z\right)=P_0\left(z\right)A_{s?d}^{global}+P_1\left(z\right)A_{s?d}^{\mathrm{identity}}& \left(\mathrm{Equation}11\right)\end{array}$ [0088] z: Pixel position [0089] f.sub.s?d(z): Optical flow from target image I.sub.d to source image I.sub.s at a pixel z position [0090] P(z): Weighted probability distribution value of the pixel z position

[0091] The image generator 108 may reconstruct and generate the target image I.sub.d based on the optical flow between the source image I.sub.s and the target image I.sub.d calculated by the optical flow predictor 106, the source image I.sub.s, and the speech feature tensor z.sub.d.

[0092] In an exemplary embodiment, when the optical flow predictor 106 calculates the optical flow f.sub.s?d(z) from the target image I.sub.d to the source image I.sub.s for each pixel, the image generator 108 may reconstruct the target image I.sub.d based on the optical flow f.sub.s?d(z) from the target image I.sub.d to the source image I.sub.s for each pixel, the source image I.sub.s, and the speech feature tensor z.sub.d.

[0093] Specifically, the image generator 108 may extract a feature tensor by inputting the source image I.sub.s into an artificial neural network (e.g., an encoder). In this case, the artificial neural network may encode the source image I.sub.s and extract a feature tensor from the source image I.sub.s.

[0094] The image generator 108 may transform a feature tensor ?(I.sub.s) of the source image I.sub.s by using the optical flow f.sub.s?d(z) from the target image I.sub.d to the source image I.sub.s for each pixel. The image generator 108 may transform the feature tensor ?(I.sub.s) of the source image I.sub.s through Equation 12 below.

[00012]$\begin{array}{cc}{?\left(I_s\right)}^{}=\mathrm{warp}\left(?\left(I_s\right),f_{s?d}\right)& \left(\mathrm{Equation}12\right)\end{array}$

[0095] Here, as the operator warp( ) for image warping, a backward warping operator may be used.

[0096] The image generator 108 may reconstruct the target image I.sub.d by inputting a transformed feature tensor ?(I.sub.s) of the source image to an artificial neural network (e.g., a decoder). The image generator 108 may learn the artificial neural network to minimize the difference between the reconstructed target image I.sub.d and the actual target image I.sub.d. The image generator 108 may generate a reconstructed target image ?.sub.d through Equation 13 below.

[00013]$\begin{array}{cc}{\stackrel{?}{I}}_d=F_{\mathrm{dec}}^{\mathrm{generator}}\left({?\left(I_s\right)}^{}\right)& \left(\mathrm{Equation}13\right)\end{array}$ [0097] F.sub.dec.sup.generator: Artificial neural network constituting the decoder of the image generator 108

[0098] In this case, the image generator 108 may modulate the weight of the artificial neural network F.sub.dec.sup.generator through Equations 14 and 15 below to control the local motion of the person with the speech feature tensor z.sub.d.

[0099] That is, since the optical flow f.sub.s?d(z) for each pixel is generated through the global geometric transformation of the person, the weight of the artificial neural network F.sub.dec.sup.generator may be modulated using the speech feature tensor z.sub.d to reflect the local motion of the person when the target image is reconstructed.

[00014]$\begin{array}{cc}s=\mathrm{fc}\left(z_d\right)& \left(\mathrm{Equation}14\right)\end{array}$ [0100] Fc( ): Fully-connected layer formed at the output terminal of F.sub.dec.sup.generator [0101] s: Scale value estimated through the fully-connected layer from the speech feature tensor z.sub.d

[00015]$\begin{array}{cc}{w}^{}=\frac{s.Math.w}{?\left(s.Math.w\right)}& \left(\mathrm{Equation}15\right)\end{array}$ [0102] w: Weight of F.sub.dec.sup.generator [0103] w: Modulated weight of F.sub.dec.sup.generator [0104] ?: Variance

[0105] That is, the image generator 108 may modulate the weight w of the artificial neural network F.sub.dec.sup.generator dec through a normalizing method through variance using a scale value estimated from the speech feature tensor z.sub.d through the fully-connected layer. Meanwhile, here, it has been described that the optical flow predictor 106 calculates the optical flow f.sub.s?d(z) for each pixel and the image generator 108 reconstructs the target image using the calculated flow, but the embodiment is not limited thereto, but without the process of calculating the optical flow for each pixel, the image generator 108 may receive the global geometric transformation A.sub.s?d.sup.global, the source image I.sub.s, and the speech feature tensor z.sub.d and reconstruct the target image based on the received ones.

[0106] Furthermore, here, it has been described that the local motion of the person is reflected by modulating the weight w of the artificial neural network F.sub.dec.sup.generator using the speech feature tensor z.sub.d, the embodiment is not limited thereto, and the local motion of the person may be reflected in the target image reconstructed using the speech feature tensor z.sub.d as the input to the artificial neural network F.sub.dec.sup.generator. That is, the target image may be reconstructed by inputting not only the transformed feature tensor ?(I.sub.s) of the source image but also the speech feature tensor z.sub.d into the artificial neural network F.sub.dec.sup.generator.

[0107] Meanwhile, when the training of the speech synthesis image generating apparatus 100 is completed, an arbitrary target image may be reconstructed from the source image by inputting the source image and an arbitrary target image to the first global geometric transformation predictor 102, and inputting the arbitrary target image into the first local feature tensor predictor 104.

[0108] According to the disclosed embodiment, the global motion and the local motion of a person in an image are separately estimated at the time of generating a speech synthesis image, thereby making it possible to reduce the overall volume of the machine learning model for generating a speech synthesis image and reduce the number of computations used therefor.

[0109] FIG. 2 is a block diagram illustrating a configuration of an apparatus for generating a speech synthesis image according to another embodiment of the present invention.

[0110] Referring to FIG. 2, the speech synthesis image generating apparatus 100 may include a first global geometric transformation predictor 102, a first local feature tensor predictor 104, an optical flow predictor 106, an image generator 108, and a second local feature tensor predictor 110.

[0111] That is, the speech synthesis image generating apparatus 100 illustrated in FIG. 2 may be an apparatus further including the second local feature tensor predictor 110 than in the embodiment illustrated in FIG. 1. Hereinafter, parts different from the embodiment shown in FIG. 1 will be mainly described.

[0112] Here, as described with reference to FIG. 1, the first local feature tensor predictor 104 may be provided to calculate a speech feature tensor for a local speech motion of a person from an input target image. That is, the first local feature tensor predictor 104 may be provided to calculate the speech feature tensor capable of expressing a difference in the local speech motion of a person between a source image and a target image based on the input target image.

[0113] The second local feature tensor predictor 110 may include an artificial neural network for estimating the local non-speech motion of the person in the speech synthesis image. In an exemplary embodiment, the artificial neural network may be trained to estimate the local non-speech motion of the person (e.g., blinking of the eyes, eyebrow motion, frown, or the like of the person) from an input partial image (or feature point).

[0114] The second local feature tensor predictor 110 may receive a partial image including only a motion related to non-speech of the person. In an exemplary embodiment, the second local feature tensor predictor 110 may receive a target partial image I.sub.d.sup.eyes including only the parts around the eyes and eyebrows of the person from the target image.

[0115] Here, the target partial image I.sub.d.sup.eyes may be an image where a mask that covers parts of the target image except the parts around the eyes and eyebrows of the person in the target image is used, or an image in which only the parts around the eyes and eyebrows in the target image are picked up.

[0116] Meanwhile, here, it has been described that the target partial image I.sub.d.sup.eyes is input to the second local feature tensor predictor 110, but the embodiment is not limited thereto, and a feature point of the target partial image I.sub.d.sup.eyes may be input to the second local feature tensor predictor 110.

[0117] Meanwhile, here, it has been described that the target partial image I.sub.d.sup.eyes is input to the second local feature tensor predictor 110, but the embodiment is not limited thereto, and the source partial image and the target partial image may be input to the second local feature tensor predictor 110. Here, the source partial image may be a partial image including only the parts around the eyes and eyebrows of the person in the source image.

[0118] When information corresponding to the global motion of the person is present in input data (partial image or feature point) including only motion related to non-speech of the person, the second local feature tensor predictor 110 may remove information corresponding to the global motion of the person from the input data. For example, when the input data is a partial image, the second local feature tensor predictor 110 may fix the position and size of a motion part related to non-speech of the person in the partial image and remove information corresponding to the global motion of the person. Further, when the input data is a feature point, the second local feature tensor predictor 110 may remove a value corresponding to the global motion of the person from feature point coordinates and leave only the motion value related to the non-speech of the person.

[0119] The second local feature tensor predictor 110 may estimate a non-speech feature tensor for a local non-speech motion of the person from the target partial image I.sub.d.sup.eyes. The second local feature tensor predictor 110 may estimate a non-speech feature tensor z.sub.d2 from the target partial image I.sub.d.sup.eyes through Equation 16 below.

[00016]$\begin{array}{cc}{z}_{d2}=F^{\mathrm{local}2}\left(I_d^{eyes}\right)& \left(\mathrm{Equation}16\right)\end{array}$ [0120] F.sup.local2: Artificial neural network constituting the second local feature tensor predictor 110

[0121] The optical flow predictor 106 may calculate an optical flow representing a motion (or change amount) in units of pixels between the source image I.sub.s and the target image I.sub.d by using a global geometric transformation calculated by the first global geometric transformation predictor 102 and the source image I.sub.s as inputs This is the same as the embodiment shown in FIG. 1, and thus a detailed description thereof will be omitted.

[0122] The image generator 108 may reconstruct and generate the target image I.sub.d based on the optical flow between the source image I.sub.s and the target image I.sub.d calculated by the optical flow predictor 106, the source image I.sub.s, the speech feature tensor z.sub.d, and the non-speech feature tensor z.sub.d2.

[0123] The image generator 108 may transform the feature tensor ?(I.sub.s) of the source image I.sub.s by using the optical flow f.sub.s?d(z) from the target image I.sub.d to the source image I.sub.s for each pixel, and may reconstruct the target image I.sub.d by inputting the transformed feature tensor ?(I.sub.s) of the source image to the artificial neural network F.sub.dec.sup.generator. In this case, the image generator 108 may modulate the weight of the artificial neural network F.sub.dec.sup.generator by using the speech feature tensor z.sub.d and the non-speech feature tensor z.sub.d2. This may be done in the same manner as in Equations 14 and 15, and thus a detailed description thereof will be omitted.

[0124] Here, it has been described that both the first local feature tensor predictor 104 and the second local feature tensor predictor 110 are included, but the embodiment is not limited thereto, and when the non-speech feature tensor for the local non-speech motion is estimated, the first local feature tensor predictor 104 may be omitted.

[0125] Meanwhile, in the disclosed embodiment, the relative change amount of the global geometric transformation of the person may be learned using a voice signal sequence. That is, a separate artificial neural network that uses a voice signal sequence (sequential voice signal) as an input may be added, and the artificial neural network may be trained to estimate the relative change amount of the global geometric transformation calculated by the first global geometric transformation predictor 102 shown in FIG. 1 from the input voice signal sequence.

[0126] FIG. 3 is a diagram illustrating a configuration for estimating the global geometric transformation of a person from the sequential voice signal in an embodiment of the present invention. Referring to FIG. 3, the speech synthesis image generating apparatus 100 may further include a second global geometric transformation predictor 112 in addition to the first global geometric transformation predictor 102.

[0127] Here, when the first global geometric transformation predictor 102 is in a trained state, and an image I.sub.i, (1?i?n) having n frames is input to the first global geometric transformation predictor 102, the first global geometric transformation predictor 102 may calculate a geometric transformation A.sub.i?r.sup.global into the i-th frame heat map from the preset reference probability distribution H.sub.r.

[0128] In addition, the first global geometric transformation predictor 102 may calculate a global geometric transformation A.sub.i+1?i.sup.global between two adjacent frames based on the geometric transformation A; global into the i-th frame heat map (i?{1 . . . , n}) from the preset reference probability distribution H.sub.r. Here, the first global geometric transformation predictor 102 may calculate the global geometric transformation A.sub.i+1?i.sup.global between two frames through Equation 17 below.

[00017]$\begin{array}{cc}{A}_{i+1?i}^{\mathrm{global}}=A_{i+1?r}^{\mathrm{global}}{A}_{i?r}^{{\mathrm{global}}^{-1}}& \left(\mathrm{Equation}17\right)\end{array}$

[0129] Meanwhile, in the training stage of the second global geometric transformation predictor 112, the second global geometric transformation predictor 112 may receive a sequential voice signal M.sub.i (1?i?n) corresponding to an image I.sub.i (1?i?n) having n frames. The second global geometric transformation predictor 112 may include an artificial neural network F.sup.seq that is trained to estimate the global geometric transformation A.sub.i+1?i.sup.global between two frames of the corresponding image from the input sequential voice signal M.sub.i.

[0130] In this case, the second global geometric transformation predictor 112 may use the global geometric transformation A.sub.i+1?i.sup.global between two frames calculated by the first global geometric transformation predictor 102 as a correct answer value, and may train the artificial neural network F.sup.seq (that is, adjust the parameter or weight of the artificial neural network F.sup.seq) to minimize the difference between the global geometric transformation between the two frames output from the artificial neural network F.sub.seq and the correct answer value.

[0131] The second global geometric transformation predictor 112 may estimate the global geometric transformation A.sub.i+1?i.sup.global between two frames of the corresponding image from the input sequential voice signal M.sub.i through Equation 18 below:

[00018]$\begin{array}{cc}\left\{A_{2?1}^{\mathrm{global}},A_{3?2}^{\mathrm{global}},.Math.,A_{n?n-1}^{\mathrm{global}}\right\}=F^{seq}\left(\left\{M_1,M_2,.Math.,M_n\right\}\right)& \left(\mathrm{Equation}18\right)\end{array}$

[0132] As described above, when the training of the second global geometric transformation predictor 112 is completed, the global geometric transformation of the person may be predicted using the source image and the sequential voice signal as inputs. In this case, the global geometric transformation of the person is predicted through the second global geometric transformation predictor 112 instead of the first global geometric transformation predictor 102. The configuration of a speech synthesis image generating apparatus 100 for achieving the above is shown in FIG. 4. That is, FIG. 4 is a block diagram illustrating a configuration of an apparatus for generating a speech synthesis image according to still another embodiment of the present invention.

[0133] Referring to FIG. 4, a speech synthesis image generating apparatus 100 may include a second global geometric transformation predictor 112, a first local feature tensor predictor 104, an optical flow predictor 106, and an image generator 108. Here, each of artificial neural networks of the speech synthesis image generating apparatus 100 is in a state in which training is completed, and a test process for generating a speech synthesis image through input data will be described.

[0134] The second global geometric transformation predictor 112 may receive a sequential voice signal of a predetermined person, and estimate a global geometric transformation A.sub.i+1?i.sup.global between two frames of an image corresponding to the sequential voice signal from the received sequential voice signal.

[0135] The second global geometric transformation predictor 112 may calculate a global geometric transformation into a start frame (source image) from a target frame (i-th frame) based on the global geometric transformation A.sub.i+1?i.sup.global between two frames of the image corresponding to the sequential voice signal.

[0136] Here, the start frame may be for providing information about the identity of the person. In this case, in order to provide information about the identity of the person, an embedding vector or the like for the person may be additionally input instead of the start frame or in addition to the start frame.

[0137] Specifically, the second global geometric transformation predictor 112 may calculate a global geometric transformation A.sub.i?1.sup.global into the i-th frame (that is, target frame) from the start frame through Equation 19 below by using the source image as the start frame based on the global geometric transformation. A.sub.i+1?i.sup.global between two frames of the image corresponding to the sequential voice signal.

[00019]$\begin{array}{cc}{A}_{i?1}^{\mathrm{global}}={.Math.}_{j=1}^{i-1}{A}_{j+1?j}^{\mathrm{global}}& \left(\mathrm{Equation}19\right)\end{array}$

[0138] Next, the second global geometric transformation predictor 112 may calculate the global geometric transformation A.sub.1?i.sup.global into the start frame from the i-th frame, which is the target frame, through Equation 20 below.

[00020]$\begin{array}{cc}{A}_{1?i}^{\mathrm{global}}=A_{i?1}^{{\mathrm{global}}^{-1}}& \left(\mathrm{Equation}20\right)\end{array}$

[0139] The first local feature predictor 104 receives a target image. Here, the target image may be an image corresponding to the i-th frame (i.e., target frame). The first local feature tensor predictor 104 may estimate the speech feature tensor z.sub.d for the local motion of a person from the target image.

[0140] The optical flow predictor 106 may receive the start frame and the global geometric transformation A.sub.1?i.sup.global into the start frame from the i-th target frame, which is the target frame, and calculate an optical flow f.sub.1?i from the target frame to the start frame for each pixel from the received ones.

[0141] The image generator 108 may receive each of the start frame, the speech feature tensor z.sub.d, and the optical flow f.sub.1?i from the target frame to the start frame for each pixel, and reconstruct and generate the target frame therefrom.

[0142] As described above, according to the disclosed embodiment, it is possible to estimate the global motion of the person by using the sequential voice signal as an input, and to generate a speech synthesis image based on the estimation.

[0143] Meanwhile, here, the first local feature tensor predictor 104 is illustrated to estimate the local speech motion of the person, but the embodiment is not limited thereto, and the second local feature tensor predictor 110 may be added to additionally estimate the local non-speech motion of the person.

[0144] FIG. 5 is a block diagram illustratively describing a computing environment 10 including a computing device suitable for use in exemplary embodiments. In the illustrated embodiment, respective components can have different functions and capabilities other than those described below; and can include additional components in addition to those described below:

[0145] The illustrated computing environment 10 includes a computing device 12. In an embodiment, the computing device 12 can be the apparatus 100 for generating the speech synthesis image.

[0146] The computing device 12 includes at least one processor 14, a computer-readable storage medium 16, and a communication bus 18. The processor 14 can cause the computing device 12 to operate according to the exemplary embodiment described above. For example, the processor 14 can execute one or more programs stored on the computer-readable storage medium 16. The one or more programs can include one or more computer-executable instructions, which, when executed by the processor 14, can be configured so that the computing device 12 performs operations according to the exemplary embodiment.

[0147] The computer-readable storage medium 16 is configured so that the computer-executable instruction or program code, program data, and/or other suitable forms of information are stored. A program 20 stored in the computer-readable storage medium 16 includes a set of instructions executable by the processor 14. In one embodiment, the computer-readable storage medium 16 can be a memory (volatile memory such as a random access memory, non-volatile memory, or any suitable combination thereof), one or more magnetic disk storage devices, optical disk storage devices, flash memory devices, other types of storage media that are accessible by the computing device 12 and capable of storing desired information, or any suitable combination thereof.

[0148] The communication bus 18 interconnects various other components of the computing device 12, including the processor 14 and the computer-readable storage medium 16.

[0149] The computing device 12 can also include one or more input/output interfaces 22 that provide an interface for one or more input/output devices 24, and one or more network communication interfaces 26. The input/output interface 22 and the network communication interface 26 are connected to the communication bus 18. The input/output device 24 can be connected to other components of the computing device 12 through the input/output interface 22. The exemplary input/output device 24 can include a pointing device (such as a mouse or trackpad), a keyboard, a touch input device (such as a touch pad or touch screen), a speech or sound input device, input devices such as various types of sensor devices and/or photographing devices, and/or output devices such as a display device, a printer, a speaker, and/or a network card. The exemplary input/output device 24 may be included inside the computing device 12 as a component constituting the computing device 12, or may be connected to the computing device 12 as a separate device distinct from the computing device 12.

[0150] Although representative embodiments of the present invention have been described in detail, those skilled in the art to which the present invention pertains will understand that various modifications can be made thereto within the limits that do not depart from the scope of the present invention. Therefore, the scope of rights of the present invention should not be limited to the described embodiments, but should be defined not only by claims set forth below but also by equivalents to the claims.