Patent Application
20180336891
The present invention relates to a technique suitable for application to devices that perform automatic speech translation, and the like. More specifically, the present invention relates to a technique for automatically generating auditory information (translated speech) in a second language as well as visual information (edited video, reproduction of motion by a robot, and the like) for the listener, from input first language auditory information (speech) and input visual information (motion of the speaker, and the like).
Against the background of recent significant progress in techniques such as speech recognition, machine translation, and speech synthesis, speech translation systems, which are a combination of these techniques, have been put into practical use. In such systems, an input in a first language is converted into a text in the first language by speech recognition technique. Further, the text in the first language is translated into a text in a second language by machine translation, and then is converted into a speech in the second language by a speech synthesis module corresponding to the second language. The practical application of this technique will eliminate the language barrier, thus allowing people to freely communicate with foreigners.
At the same time, in addition to auditory information from the ears, visual information from the eyes such as facial expression and gesture can greatly contribute to the transmission of meaning. For example, a gesture such as “pointing” can greatly contribute to the understanding of meaning. Thus, the motion of the speaker is transmitted to the listener through an image or a robot, to achieve more natural communication. For example, Patent Literature 1 has proposed reproduction of the motion of the speaker through a robot.
However, in a speech translation system, when a first language speech is translated into a second language speech, it is difficult to guarantee that a word with the same meaning always comes at the same time (the time relative to the beginning of the speech). Thus, mismatch (hereinafter, referred to as “time lag”) occurs between the visual information from the eyes of the listener and the auditory information from the ears of the listener, which may significantly impair understanding of the meaning.
The conventional method (Patent Literature 2) proposes a method in which the start time and end time of the second language speech are adjusted to the first language speech, which makes it possible to synchronize between the speech and the image at the start and end times. However, the problem of local time lag between visual information and auditory information remains unsolved. In particular, in the case of translation of Japanese and English, the time lag due to the difference in the order of words is significant, and this may lead to misinterpretation.
For example, it is assumed that a speaker points (gesture 1) a can 1 (first spoken can) while speaking “Put this can into this can.” in a first language (English), and then makes the motion to pointing (gesture 2) a can 2 (second spoken can). At this time, the temporal correspondence between the sound of the can 1 and the gesture 1, and the sound of the can 2 and the gesture 2 significantly contributes to the listener's understanding of the meaning. However, when it is translated into “Kono kan ni, kono kan wo irete kudasai.” for a native speaker of a second language (Japanese), the order of the can 1 (first spoken can) and the can 2 (second spoken can) is reversed from the first language, so that the temporal correspondence is between the sound of the can 2 and the gesture 1, and the sound of the can 1 and the gesture 2. As a result, the meaning is reversed from what the speaker intended to say.
Patent Literature 1: Japanese Unexamined Patent Application Publication No. 2004-230479
Patent Literature 2: Japanese Unexamined Patent Application Publication No. 2002-123282
Patent Literature 3: Japanese Unexamined Patent Application Publication No. Hei 6-253197
The aim of the present invention is to eliminate the influence of the traditional automatic speech translation system on the understanding of the meaning due to translated second language speech, reproduced motion of a person who speaks a first language (speaker), and “time lag” in information that both sides have.
In order to solve the above problem, an aspect of the present invention is a method for synchronizing visual information and auditory information, including: extracting the visual information included in an image; recognizing the auditory information in a first language that is included in a speech in the first language; associating the visual information with the auditory information in the first language; translating the auditory information in the first language into visual information in a second language; and editing at least one of the visual information and the auditory information in the second language so as to associate the visual information with the auditory information in the second language.
A specific example is given for easy understanding, in which the visual information is, for example, gesture (pointing, motion) and facial expression, and the auditory information is, for example, speech content. The image is, for example, a video of the speaker that is captured by a camera.
Another aspect of the present invention is an information processing device that inputs input image data including first visual information as well as input speech data in a first language that includes the first visual information, and outputs output speech data including second visual information corresponding to the first visual information as well as output speech data in a second language that includes second auditory information corresponding to the first auditory information. This device includes: a detection unit for detecting the first visual information from the input image data; a recognition unit for recognizing the first auditory information from the input speech data; an association unit for associating the first visual information with the first auditory information; a translation unit for converting the input speech data in the first language into output speech data in the second language; and an optimization unit for controlling the output timing of at least one of the second visual information and the second auditory information so as to reduce the output timing difference between the second visual information and the second auditory information, both of which correspond to the first visual information and the first auditory information that are associated with each other.
The output visual data can be data that is finally recognized by the user's eyes, and may be video (for example, obtained by editing the input image data) or other data (for example, motion command data to specify the robot motion).
Still another aspect of the present invention is a device that automatically translates an input of speech in a first language into speech in a second language. In addition to the base configuration, the device includes: means for obtaining the body motion of the speaker, together with the input speech in the first language; means for detecting the correspondence between the obtained body motion and the input speech in the first language; means for detecting the correspondence between the speech in the second language, which is translated from the speech in the first language, and the obtained body motion of the speaker; means for reproducing the second language speech while reproducing the body motion of the speaker; means for evaluating the difference on the time axis between the second language and the reproduced body motion of the speaker; a plurality of means such as editing of the reproduced body motion of the speaker, editing of the translated second language text, and editing of the synthesized second language speech, with respect to the time axis; means for quantitatively evaluating the influence on the ease of understanding of the spoken context by editing of the reproduced body motion of the speaker, editing of the translated second language text, and editing of the synthesized second language speech; and means for selecting the most understandable means by using the evaluation means described above, after minimizing the time lag.
According to the present invention, it is possible to reduce the “time lag” between the translated speech and the reproduced motion of the speaker and to achieve natural communication.
Hereinafter, preferred embodiments will be described with reference to the accompanying drawings. However, the present invention is not construed to be limited to the embodiments described below. Those skilled in the art can easily understand that the specific configuration can be modified without departing from the spirit or scope of the present invention.
In the configuration of the present invention described below, the same reference numerals are used in common for the same parts or parts having the same functions among different figures, and overlapping description may be omitted or simplified.
The expressions such as “first”, “second”, and “third” used in this specification or other applicable documents are intended to identify the components and do not necessarily limit the number or order of components. Further, a number for identifying a component is used for each context and the number used in one context does not necessarily represent the same configuration in other contexts. Further, the configuration identified by a certain number does not interfere with having the function of the component identified by another number.
The position, size, shape, range, and the like of each of the configurations shown in the figures or described herein may not represent the actual position, size, shape, range, and the like to facilitate the understanding of the invention. Thus, the present invention is not necessarily limited to the position, size, shape, range, and the like disclosed in the figures and the specification.
In the following embodiments, functions such as calculation and control are assumed to perform a given process in corporation with other hardware when a program stored in a storage device is executed by a processor. The means for achieving the program that the computer or other devices executes, or its function may be referred to as “unit”, “function”, “means”, “part”, “module” or the like. Further, in the embodiments, the function equivalent to the function configured in software can also be configured in hardware such as FPGA (Field Programmable Gate Array) and ASIC (Application Specific Integrated Circuit). Such aspects are also included in the scope of the present invention.
The present embodiment will describe the basic configuration of the present invention under the assumption that the recorded (or remotely transferred) second language speech and the image data are converted into a second language speech and an image. For example, by assuming a remote class or teaching, the description assumes the case in which the image can be edited.
In the present embodiment, the basic functionality is achieved that convers a first language speech and image into a second language speech and image having the same meaning as the input by a speech translation device (system) of the present invention. In the following description, each unit of internal processing is called oo device, oo unit, or oo module. However, it is possible to implement it as a system that is implemented as software program or as a form of program, instead of implementing it as hardware.
Of the overall configuration shown in
When synchronized image data and speech data (first language) are input, the speech translation device 100 of this embodiment identifies the correspondence between gesture (pointing, and the like) extracted from the image, which is important for communication, and word (first language) spoken at the same timing. Then, the speech translation device 100 determines the most appropriate editing method (image editing, text editing, or speech editing) from the same translated text (second language), and minimizes the time lag between the gesture and the corresponding word (second language) while allowing naturalness and ease of understanding of image, text, and speech to be maintained as much as possible.
In order to achieve this speech translation process, the speech translation device 100 has the following processing units that general speech translation systems have: the speech recognition unit (first language) 101, the text translation unit (translating from first language to second language) 102, the speech synthesis unit (second language) 1061, and the speech output unit 107. In addition to these processing units, the speech translation device 100 also includes a gesture detection unit 104, an information synchronization unit 105 between gesture and speech (first language), an information synchronization unit 103 between text (second language) and speech (first language), an optimization unit 106 of information synchronization between image and speech (second language), and an image output unit 108.
Further, as detailed shown in
The server 1003 includes, as general configuration, an input device 1005, an output device 1006, a processing device 1007, and a storage device 1008. The modules described in
Hereinafter, the processes performed by each of the processing units configuring this embodiment are first described with reference to
Further, the following description focuses on an example of translating speech from Japanese to English and from English to Japanese. However, it is also possible to translate between other languages. In this case, the internal processing (speech recognition, text translation, speech synthesis) should use program and data corresponding to each particular language.
Further, in the following description, “pointing” is described as an example but it can be expanded to all visual information that can be detected from an image. In particular, visual information that changes over time, such as hand and body gestures of the speaker and changing background, is considered to be important. Hereinafter, in the present invention, just describing gesture means all visual information.
Further, each of the processing units that configure this embodiment is described as automatic processing by a program. However, it is also possible to consider manual work by a worker in place of automatic program processing. For example, it is assumed that a gesture extraction unit allows extraction by hands of a worker.
The gesture detection unit 104 extracts visual information, which is important for communication, from the input image. It may be possible to perform a pattern recognition process on the input image. However, the importance of visual information is greatly affected by the content of the speech, so that a more accurate extraction is possible when extracting visual information by reference to the text obtained by speech recognition, rather than when extracting visual information only from the input image. The description exemplifies a remote class or teaching in which it is considered that the motion of “pointing” is important, so that extraction of the “pointing” motion is performed.
In an example, with respect to the input speech “Put this can into this can.” (English), two pointing motions are detected from the image. The first motion (pointing 1) represents can 1 (corresponding to the can which is spoken first), and the second motion (pointing 2) represents can 2 (corresponding to the can which is subsequently spoken). Hereinafter, for ease of illustration, the first spoken can is expressed as can (1) which is assumed to correspond to kan 1 in the speech (text) in the second language, and the second spoken can is expressed as can (2) which is assumed to correspond to kan 2 in the speech (text) in the second language. Note that the numbers attached to the words, kan and can, are given to the specification and drawings for the purpose of illustration, and please keep in mind that the numbers that differentiate between kan and can are not actually spoken.
The information synchronization unit 105 between gesture and speech (first language) determines the correspondence between the gesture extracted by the gesture detection unit 104 and the corresponding word (first language).
An example is shown in
The information synchronization unit 103 between text (second language) and speech (first language) synchronizes the information between the second language text and the first language speech, by using the correspondence information between the first language text obtained by the speech recognition unit 101 and the second language text obtained by the text translation unit 102. It is assumed that the speech recognition unit 101 and the text translation unit 102 use various known techniques, and thus detailed description thereof will be omitted.
As shown in the example in
The optimization unit 106 of information synchronization between image and speech (second language) first generates a second language speech from the translated second language text by the speech synthesis unit (second language) 1061.
It is assumed that the time lag can be evaluated from the difference between the start/end time of the gesture and the start/end time of the corresponding second language speech. If all time lags are equal to or less than a certain threshold, the time-lag evaluation module 1062 outputs the evaluation results to the image output unit 108 and the speech output unit 107. However, if any of the time lags exceeds the threshold, the time-lag evaluation module 1062 outputs the evaluation results to the editing module 1064.
An example of the evaluation method is as follows:
Time lag 1=|start time 1 (gesture 1)−start time 2 (second language speech)|+|end time 1 (gesture 1)−end time 2 (second language speech)|
Time lag 2=|start time 2 (gesture 2)−start time 1 (second language speech)|+|start time 2 (gesture 2)−start time 1 (second language speech)|
The mark “|*|” means the absolute value of *. Here, when the start time 1, end time 1, start time 2, and end time 2 of the gesture are respectively 100 ms, 300 ms, 400 ms, and 600 ms, and when the start time 1, end time 1, start time 2, and end time 2 of the second language are respectively 0 ms, 200 ms, 200 ms, and 400 ms, the following equations are obtained:
Time lag 1=|100 ms−200 ms|+|300 ms−400 ms|=200 ms
Time lag 2=|400 ms−0 ms|+1600 ms−200 ms|=800 ms
When the predetermined threshold of the “time lag” is 50 ms, both the “time lag” (=200 ms) of the gesture 1 and the “time lag” (=800 ms) of the gesture 2, both of which are calculated by the above equation, are greater than the threshold. Thus, the time-lag evaluation module 106 determines that it is necessary to edit them, and outputs to the editing module 1064.
The editing module 1064 can edit (or control the timing of) the second language speech and the input image by using at least one of the methods of editing with a plurality of modules, such as an image editing module 1064a, a text editing module 1064b, and a speech editing module 1064c.
The image editing module 1064a allows acceleration and slowing down of the image, and changing the order of the scenes. Here, image editing can use all kinds of image editing techniques such as temporarily holding image, changing the reproduction order of video, and reproduction by CG using computer graphics (CG).
The text editing module 1064b may allow change in the order of words as well as conversion of a certain word into a synonym with different length. Here, text editing can use all kinds of techniques for changing text.
The editing module 1064a may allow partial acceleration and slowing down of the image, or inserting a pause. Here, speech editing can use all kinds of speech synthesis techniques for changing speech. Here, it is possible to use the speech synthesis unite (second language) 1061 as the speech editing module 1064c. In other words, it is possible to prevent the degradation of naturalness due to speech editing, by resynthesizing the speech, not because of editing the synthesized speech.
As described above, by using the editing functions, the editing module 1064 edits the image and speech so as to further reduce the time lag between the second language speech and the gesture.
Each of the modules can be used in combination or alone. Note that each module can be configured to adjust the rate of output speech so that the whole length of the input speech (first language) and the whole length of the output speech (second language) are approximately equal before the process by the optimization unit, and then perform editing within this time range.
The speech/image naturalness evaluation module 1065 evaluates the naturalness for each of a plurality of time lag elimination methods (image editing, text editing, speech editing, and the like), and selects the method with high naturalness.
With respect to the naturalness evaluation of the edited image, as described in Patent Literature 3, it is possible to evaluate the naturalness of the image by evaluating the spatio-temporal continuity of the edited image as well as the smoothness of spatio-temporal variation in the edited image.
The naturalness evaluation of the edited text evaluates the naturalness from several viewpoints such as grammar, written style, and written style. For example, it may be consider an automated Japanese essay scoring system proposed in Non-Patent Literature 1.
The naturalness evaluation of the edited text can allow an objective evaluation by using the continuity of the spectrum of the speech, as well as the generation probability calculation by a statistical model such as Hidden Markov Model (HMM). For example, discontinuity due to text editing can be evaluated by using the amount of change in speech parameter (Non-Patent Literature 2). On the other hand, with respect to speech distortion due to speech editing such as prosody modification, it may be possible to consider a method for evaluating the naturalness by calculating the probability of generating the edited text by using HMM (Non-Patent Literature 3) that expresses the process of speech generation by the probability. As described above, the method for evaluating the naturalness can use various well-known techniques and there is no particular limitation.
In an example, the text editing method is selected because it is evaluated that the naturalness is higher in changing the word order after text editing than in image editing. However, in the case of only performing text editing, although the time lag from the gesture 1 is eliminated, speech editing (pause insertion and speech rate adjustment) is required to make the time lag from the gesture 2 being smaller than the threshold.
Thus, in an example, when performing speech editing (
As described above, the editing methods with excellent naturalness are selected and combined. As for selection criteria, there are method of placing emphasis on the naturalness of the image, method of placing emphasis on the naturalness of the speech, and method of balancing the both, and an arbitrary method can be set depending on the intended use and situation.
Note that, as shown in the configuration in
The configuration of the embodiment described above can be configured with a single computer as shown in
This embodiment describes the basic configuration of the present invention, under the assumption that the speech translation is performed through a robot. In other words, the configuration compliments the speech by the motion of a robot, in place of image output in the first configuration. In this embodiment, when the speech and image of a first language speaker is input to a speech translation device (robot) of the present invention, the robot translates the first language speech into a second language speech while achieving the basic functionality to reproduce the motion of the speaker.
When synchronized image data and speech data (first language) are input, a speech translation device 800 of this embodiment identifies the correspondence between a gesture (pointing, and the like) extracted from the image, which is important for communication, and a word (first language) spoken at the same timing. Then, the speech translation device 800 determines the most appropriate editing method (robot motion command editing, text editing, speech editing, or the like), and minimizes the time lag between the reproduced gesture of the robot and the corresponding word (second language), while allowing the naturalness and ease of understanding of the robot motion, the text, and the synthesized speech to be maintained as much as possible.
In order to achieve this speech translation process, the characteristic configuration of the second embodiment compared to the configuration of the first embodiment is that the speech translation device 800 includes the optimization unit of information synchronization between robot motion and speech (second language) 806, and a robot motion control unit 808. The optimization unit of information synchronization between robot motion and speech (second language) 806 includes an editing module 8064 in place of the editing module 1064 of the first embodiment, as well as a time-lag evaluation module 8062 between the second language speech and the gesture (of the robot) in place of the time-lag evaluation module 1062 between the second language speech and the gesture (of the image).
The editing module 8064 includes a motion editing module 8064a that edits the motion of the robot, in place of the image editing module 1064a. The evaluation of reduction in the naturalness of the speech/robot motion due to the editing module 8064 is performed by a speech/motion naturalness evaluation module 8065.
As for the description of the process of this embodiment, only the difference from the first embodiment is described. The description of the same processing units as in the first embodiment will be omitted.
In the second embodiment, it is possible to change the motion of the robot by editing the motion command for controlling the motion of the robot, in place of the image editing of the first embodiment. (Hereinafter, “motion command editing” and “robot motion editing” are the same meaning.)
In the optimization unit of information synchronization between robot motion and speech (second language) 806, it is possible to evaluate the time lag between the second language speech by speech synchronization and the robot body motion (gesture) according to the robot motion command, by the time-lag evaluation module between second language speech and gesture 8062.
Time lag 1=|start time 1 (gesture 1)−start time 2 (second language speech)|+|end time 1 (gesture 1)−end time 2 (second language speech)|
Time lag 2=|start time 2 (gesture 2)−start time 1 (second language speech)|+|start time 2 (gesture 2)−start time 1 (second language speech)|
The mark “*” means the absolute value of *. Here, when the start time 1, end time 1, start time 2, and end time 2 of the gesture are respectively 100 ms, 300 ms, 400 ms, and 600 ms, and when the start time 1, end time 1, start time 2, and end time 2 of the second language are respectively 0 ms, 200 ms, 200 ms, and 400 ms, the following equations are obtained:
Time lag 1=|100 ms−200 ms|+1300 ms−400 ms|=200 ms
Time lag 2=|400 ms−0 ms|+1600 ms−200 ms|=800 ms
When the predetermined threshold of “time lag” is 50 ms similarly to the first embodiment, both the “time lag” (200 ms) of gesture 1 and the “time lag” (800 ms) of gesture 2, which are calculated with the above respective equations, are greater than the threshold. Thus, it is determined that editing is required, and the data is output to the editing module 8064.
In the editing module 8064, unlike text editing in the first embodiment, the operation command editing module 8064a includes the function for editing the robot motion according to the motion command editing. Here, methods such as a text editing module 8064b and a speech editing module 8064c are the same as those in the first embodiment, and the description thereof will be omitted.
The motion command editing module 8064a may allow acceleration and slowing down of the image, and changing the order of the motions. With this configuration, the motion command editing module 8064a edits the robot motion to further reduce the time lag between the second language speech and the gesture of the robot.
Further, as shown in
The speech/motion naturalness evaluation module 8065 evaluates the naturalness for each of a plurality of methods (motion command editing, text editing, speech editing, and the like) that eliminate the “time lag. Then, the speech/motion naturalness evaluation module 8065 selects the method with the highest naturalness. In an example, as shown in
However, there is a possibility that the motion may not be obtained at a desired timing even after the motion command editing due to the limitations of the robot itself (rate of motion, and the like).
This embodiment describes the basic configuration of the present invention, under the assumption that speech translation is performed through a mobile terminal (mobile phone, tablet, or the like).
This embodiment achieves the basic functionality for converting first language speech and image into second language speech and image having the same meaning as the input, by using a speech translation device (system) according to the present invention. In the following description, each unit of internal processing is called oo device. However, it is possible to implement it as a system that is implemented as software program or as a form of program, instead of implementing it by a device as hardware.
The speaker A has a tablet A 10010. The speech spoken by the speaker A in language A is speech-recognized and displayed in a display area 10020 of the tablet (sentence 10050, “aabbcc”, on the tablet A). The result of translating the word into language B that the speaker B can understand is displayed in a display area 10040 of a tablet B 10030 that the speaker B has (sentence 10070, “AABBCC”, on the tablet B).
On the other hand, the speech spoken by the speaker B in language B is speech-recognized and displayed on the tablet B 10030 (sentence 10080, “DDEEFF”, on the tablet B). This word is translated to language A and displayed in the tablet A 10010 that the speaker A has (sentence 10060, “ddeeff”, on the tablet A).
Here, it is assumed that the speaker A cancels a word that has been spoken. Various methods can be considered for specifying cancellation. For example, the speaker A sweeps the cancellation part of the display area 10020 on the tablet A. Here, when the speaker A sweeps the sentence 10050 “aabbcc” on the tablet A, the sentence 10070 on the tablet B also has a sign indicating that the particular word has been cancelled. For example, it is considered that the x mark is displayed or the word itself is deleted.
Further, it could also be a case of wanting to cancel only some words in one sentence. At this time, for example, the speaker A sweeps the cancellation part of the display area 10020 on the tablet A. Here, the speaker A sweeps only “bb” of the sentence 10050 “aabbcc” on the tablet A. Then, a sign indicating that the particular words have been cancelled is displayed also on the corresponding display “BB” of the sentence 10070 on the tablet B. For example, actions such as displaying the X mark on “BB” or deleting “BB” may take place.
The above process can be performed by the tablet. However, when the tablets A and B are connected by a network, for example, the tablet A can be associated with the input-side terminal 1001 and the tablet B can be associated with the output-side terminal in the configuration in
As described above, according to the embodiment described in
a step of inputting a first speech in a first language into a terminal A;
a step of displaying the first speech to the terminal A (by speech recognition, or the like) as a first text;
a step of transferring the first text from the terminal A to a terminal B (through a network, or the like);
a step of displaying the first text on the terminal B as a second text in a second language (by machine translation or other translation techniques) (note that machine translation can be performed by the server within the network or by the terminal A or B);
a step of specifying an arbitrary part of the first text displayed on the terminal A;
a step of notifying the terminal B from the terminal A about the specification of the arbitrary part; and
a step of changing the display of the part corresponding to the specified arbitrary part of the first text, in the second text displayed on the terminal B.
It should be noted that the present invention is not limited to the above exemplary embodiments and includes various variations. For example, part of the configuration of an embodiment can be replaced with the configuration of another embodiment, and the configuration of an embodiment can be added to the configuration of another embodiment. Further, the addition, deletion, and substitution of the configuration of another embodiment can be made for parts of the configuration of each embodiment.
The present invention can be used in communication tools such as for automatic translation.
Speech translation device 100, Speech recognition unit 101, Text translation unit 102, Information synchronization unit 103 between text (second language) and speech (first language), Gesture detection unit 104, Information synchronization unit 105 between gesture and speech (first language), Speech synthesis unit (second language) 1061, Optimization unit 106 of information synchronization between image and speech (second language), Speech output unit 107, Image output unit 108
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2015/080574 | 10/29/2015 | WO | 00 |
