Device and method for creating a filter for classifying an item of content on the basis of a first item of content

Abstract

The present invention relates to a method for creating a filter for classifying an item of content on the basis of a first item of content, the method comprising a) obtaining a set of items of content obtained by varying, within a range of values, the value of at least one parameter associated with the first item of content; b) obtaining at least one item of information for classifying the items of content; c) associating a value representative of the at least one parameter with the classification - repeating steps a) to c) until the variation between two iterations of the value representative of the at least one parameter associated with the classification is less than a determined value, wherein the range of values is determined during each iteration according to the representative value associated with the classification, and wherein the filter comprises the value of the at least one parameter obtained during the last iteration.

Description

device and method for creating a content classification filter, from a first content

[0001] The present invention relates to the field of determining filters for transforming content into one or more contents. The present invention applies particularly but not limited to audio and / or image or video type content.

[0002] In some areas, it may be considered to want to modify voices or appearances to create avatars for example, or to personalize voices of a voice assistant. It may be considered to want to obtain voices or faces for example, which would be classified as "happy", "sad", "angry"...

[0003] There are reverse correlation methods that aim to extract salient features from stimuli that trigger a specific judgment or mental representation in subjects. These methods make it possible to obtain an average stimulus characterizing this mental representation.

[0004] However, known methods have the disadvantage of requiring a lot of stimuli to construct this average stimulus, resulting in a long and tedious protocol, or even the simple impossibility of obtaining this average stimulus.

[0005] Several methods have been proposed to reduce this number. However, they all have the disadvantage of reducing the search space blindly, by depriving themselves of stimuli whose value with respect to the evaluation criterion considered is unknown. As a result, the average stimulus, thus constrained by a set of given stimuli, models a mental representation in a less reliable, or even biased, way.

[0006] The present invention relates to a method for creating a filter for classifying content, from a first content comprisinga) obtaining a set of contents obtained by varying in at least one range of values ​​the value of at least one parameter associated with said first content,b) obtaining at least one item of classification information for said contents,c) associating at least one value representative of said at least one parameter with said classification- repeating steps a) to c) as long as the variation, between two iterations, of the at least one value representative of said at least one parameter associated with said classification is less than a determined value, said at least one range of values ​​being determined during each iteration as a function of the at least one representative value associated with said classification, said filter comprising the at least one value of said at least one parameter obtained during the last iteration.

[0007] According to at least one embodiment, the contents of said set of contents are obtained from said first content by randomly varying said at least one parameter, said variation of said at least one parameter generating a transformation of at least one characteristic of said first content.

[0008] According to at least one embodiment, the representative value of said at least one parameter associated with a given classification is obtained by calculating the algebraic mean of the values ​​of said at least one parameter of the contents classified in said given classification.

[0009] According to at least one embodiment, the range of values ​​determined during the second and subsequent iterations is centered on the representative value of said at least one parameter associated with said at least one classification obtained during the previous iteration.

[0010] According to at least one embodiment, the determined range of values ​​is reduced at each iteration.

[0011] According to at least one embodiment, when at least the value of two parameters is associated with a classification - a representative value is associated for each of said at least two parameters associated with said at least one classification - steps a) to c) are iterated as long as the variation in the value of at least one of said at least two parameters associated with said at least one classification between two iterations is greater than a determined value.

[0012] According to at least one embodiment, the classification information of said contents is obtained from a user.

[0013] According to at least one embodiment, the method comprises applying said filter to said at least one parameter associated with said characteristic of a target content so that said characteristic of said target content is associated with said classification.

[0014] According to at least one embodiment, said content is audio content and said parameters are parameters associated with the prosodic profile of the voice, and preferably representative of a voice pitch, a speaking speed and a sound level.

[0015] According to at least one embodiment, said content is image content and said parameters are parameters associated with the brightness of the image, and preferably representative of gray levels.

[0016] According to at least certain embodiments, said method comprises - the occultation of at least a part of said first content, - obtaining a set of contents obtained by varying in at least one range of values ​​the value of at least one parameter associated with said first content comprising said at least one occulted part.

[0017] According to at least some embodiments, said method comprises - the occultation of at least one part of said first content, - obtaining a set of contents obtained by varying in at least one range of values ​​the value of at least one parameter associated with said first content comprising said at least one occulted part, - the multiple execution of said method on said content, each execution of said method being carried out on a part of said content in which the other parts of said content are occulted, in order to obtain a representative value of said parameter associated with said classification for each part of said content, - said filter being obtained by aggregating the representative values ​​of said parameter obtained for each part.

[0018] The features presented in isolation in the present application in connection with certain embodiments of the method of the present application can be combined with each other according to other embodiments of the present method.

[0019] The present invention also relates to a computer program comprising instructions for executing the steps of the method according to the invention, according to any one of its embodiments, when said program is executed by a computer.

[0020] The present invention also relates to a computer-readable recording medium on which is recorded a computer program comprising instructions for executing the steps of the method according to the invention, according to any of its embodiments.

[0021] The present invention also relates to a device for creating a filter for classifying content, from a first content comprising one or more processors configured together or separately to implement the instructions for executing the steps of the method according to the invention, according to any one of its embodiments. Thus, the present invention also relates to a device for creating a filter for classifying content, from a first content comprising one or more processors configured together or separately to a) obtain a set of contents obtained by varying in at least one range of values ​​the value of at least one parameter associated with said first content, b) obtain at least one classification information of said contents, c) associate at least one value representative of said at least one parameter with said classification - repeat steps a) to c) as long as the variation, between two iterations,of the at least one representative value of said at least one parameter associated with said classification is less than a determined value, said at least one range of values ​​being determined during each iteration as a function of the at least one representative value associated with said classification, said filter comprising the at least one value of said at least one parameter obtained during the last iteration.,

[0022] Other characteristics and advantages of the present invention will emerge from the description given below, with reference to the accompanying drawings which illustrate an exemplary embodiment thereof without any limiting character.

[0023] Represents a method according to certain embodiments of the invention.

[0024] The figure represents a device according to certain embodiments of the invention.

[0025] This disclosure may advantageously be applied to content of different natures and for example image content, video content or audio content, olfactory content, gustatory, tactile, proprioceptive or gestural content, or multimodal content. In a non-limiting manner, the embodiments described below relate to audio content but may apply in an identical manner to content of other natures.

[0026] Larepresents a method according to certain embodiments of the present disclosure.

[0027] The first content is selected. As mentioned above, this content is audio-based, but it could also be of another nature, as mentioned above. Here, we focus on a characteristic of the content that corresponds to voice.

[0028] We can have a content characteristic that is related to the image when the content is visual. We can be interested in two characteristics at the same time, an audio characteristic and an image characteristic, when the content is audiovisual content (including both an audio and a video component).

[0029] This content may be obtained synthetically or selected from a set of audio content. This content is generally qualified as an inversely correlated baseline stimulus. In some embodiments, it may be, for example, a voice qualified as a neutral voice.

[0030] On this neutral voice, during step E1, we modify one or more parameters in order to modify it, or even to distort it. We thus obtain one or more voices from this neutral voice forming a first set of contents.

[0031] The parameter(s) modified on this neutral channel may, for example, be parameters relating to the prosodic profile of the voice.

[0032] The prosodic profile of voice refers to how prosodic elements are used in a speaker's speech. Prosodic elements are speech characteristics that encompass intonation, rhythm, speaking rate, stress, and other aspects related to how words and sentences are pronounced and organized. This profile helps to impart meaning, emotion, and context to speech. Some of the elements that are part of the prosodic profile include the following parameters:- intonation: Intonation refers to the melodic variation in the pitch of the voice during speech. It can indicate emotion (such as joy, anger, sadness), sentence type (declarative, interrogative, imperative), and other nuances of meaning.- rhythm: The rhythm of speech concerns the distribution of time between words and syllables.A steady rhythm can convey a calm or serious tone, while an irregular rhythm can reflect excitement or stress.- Speech rate: Speech rate refers to the speed at which a person speaks. A fast rate can express enthusiasm or impatience, while a slow rate can indicate thoughtfulness or suppressed emotion.- Voice modulation: This encompasses how the voice can change in pitch, volume, and rhythm to express emotion, irony, and uncertainty.- Stress: Stress involves the emphasis placed on certain syllables or words in a sentence. It can help highlight important elements and clarify meaning.- Pause: Pauses in speech are an important aspect of prosody. They can mark boundaries between sentences, allow the listener to process information, and add clarity to communication.

[0033] These parameters, as mentioned above, are parameters which can therefore allow us to give a mental or emotional representation of the voice by characterizing it for example as “sad”, “happy”, “dynamic”, “playful”, “angry”, “calm”.

[0034] When the content is image content, parameters can be used that relate to brightness, such as the grayscale. Parameters can also be used that relate to the content itself or to what the content includes or represents, i.e., to the object represented by the content when the content represents an object. An object can be understood, for example, as a face, an animal, but also as a landscape. When the content is representative of a face, parameters can relate to the size of the mouth, eyes, nose, skin color, hair, etc.

[0035] According to certain embodiments, when the content is an image content, during step E1, a set of contents is obtained by previously obscuring one or more parts of the content of the basic stimulus and by varying on this content comprising one or more obscured zones, one or more parameters associated with one or more parts of the image which are not obscured. The present invention then makes it possible to determine one or more filters associated with one or more parameters of this face.

[0036] In step E1, by varying one or more of these parameters, we obtain a set comprising a plurality of contents where the voice present in the neutral content is transformed. The number of contents in this first set can, for example, be between 100 and 1000 contents. It can also depend on the number of parameters that we vary.

[0037] In order to obtain this first set of contents, the parameters, or at least some of them, are varied over a very wide range of values, between a maximum value and a minimum value representative of the entire spectrum of variation associated with this value. The pitch of the voice can be varied, on a scale varying, for example, from -4 tones (lower voice) to +4 tones (higher voice).

[0038] The duration of a syllable or a portion of the stimulus can be varied, on a scale ranging, for example, from a factor of 2 (voice twice as fast) to a factor of 0.5 (voice twice as slow).

[0039] The choice of the value range (boundaries and distribution) may depend on human perception (inaudible above or below a certain threshold), or on the credibility that we want to maintain in the voice.

[0040] For example, we can sample the values ​​in this range based on the desired number of contents. The sampling can be different depending on the parameter and / or can be different depending on the proposed classification(s). Indeed, some parameters may be more representative for one classification than for another. In such a case, we can have a finer sampling frequency for some parameters than for others.

[0041] At the end of step E1, a first set is therefore obtained containing a plurality of audio contents obtained from a first audio content, called neutral content. This plurality of contents represents the same audio content as the neutral content on which one or more parameters of the prosodic profile of the voice of the neutral content have been modified, thus making it possible to obtain a plurality of variations of the neutral content, each of the contents obtained therefore being associated with parameters whose value is different at least for certain parameters, from one content to another.

[0042] In step E2, this first set of contents is presented to one or more individuals. The individual(s) classify each of the contents of the set of contents into a category according to a set of proposed classifications or a single classification. We may indeed only want to determine what characterizes a happy (or sad, or other) voice and in this case we can offer an individual a positive or negative response to the question: "is this voice happy?".

[0043] According to certain embodiments, this first set of contents can also be provided as input to a model-based algorithm, this advantageously making it possible to verify the criteria of these algorithmic models.

[0044] Among the proposed classifications, we can notably cite, in a non-limiting manner, the following classifications: - happy, - sad, - cheerful, - angry, - dynamic, - nervous, - calm

[0045] We can offer, via a user interface for example, on a computer, a tablet, a mobile phone, all the content to one or more users. The user then comes to play the content, or even scrolls through it, content by content of the set of content obtained during step E1. For each of the contents, he selects one of the classifications that are proposed to him according to his personal feeling.

[0046] At the end of step E2, we thus obtain a classification of the voice for each of the contents. If the same set of contents is classified by several individuals, an average can be carried out to obtain the overall classification from the individual classifications of the users.

[0047] In step E3, the values ​​of the parameters representing the classifications that have been proposed to the user are determined. To do this, we can take all the content classified according to a classification, extract (or obtain, or determine, or read) the associated parameter(s).

[0048] From these parameters, the values ​​can be aggregated and a value that is determined to be representative from all of these values ​​is chosen. According to certain embodiments, for example, an average can be made per parameter and, for example, an algebraic average of the parameter values ​​can be made. According to certain embodiments, a distribution of the values ​​can be plotted, and the one that has been associated most often can be kept.

[0049] This gives a representative value for each parameter for the classification(s).

[0050] We can thus, by applying this parameter value, for the parameter(s), to the neutral, initial audio content, obtain a so-called average content, or also called average stimulus by classification.

[0051] Then, during a step E4, a variation in the value of this or these parameters is determined for each classification. This step is not implemented for the set of parameters obtained following the classification of the first set of contents since the variation determined concerns the variation obtained following the presentation to the user of successive sets of contents, as described later.

[0052] Following the determination of a representative value for each parameter for the classification(s), obtained in step E3, a new set of parameter values ​​is determined in step E5. This new set of parameter values ​​comprises a range of values ​​depending on the representative value obtained in step E3 and preferably centered on this value.

[0053] For example, when the content is audio content, during a first iteration, one can, during step E1, generate voice pitch values ​​with a uniform distribution between -3 and +3 tones relative to the voice of the basic stimulus. After a first iteration, one can determine that the parameter(s) associated with the voice pitch is concentrated in an area between +2 and +3 tones relative to the parameter relating to the voice pitch of the basic stimulus. During the following iteration, the values ​​of the parameter relating to the voice pitch are therefore chosen in this area.

[0054] According to some embodiments, it is also possible to simply move the range, without necessarily reducing it (for example during the first iterations), and to choose for example an area between +2 and +8 tones, for the purpose of exploration. According to some embodiments, it may be considered to enlarge the range if at the first iteration no particular area seems to emerge.

[0055] According to certain embodiments, it may be envisaged, during a new iteration, to explore at least two non-contiguous ranges of values, when for example at least two values ​​can be associated with a parameter to represent a classification. This may for example occur when the representative value is not obtained by algebraic mean but comes from a distribution of values. It could be envisaged that a distribution of values ​​highlights two or more representative values. For example, if a concentration is observed both between 0 and +1 tone, and between +2 and +3 tones (but nothing on the other hand between +1 and +2 tones), it may be decided, during the following iteration, to generate two ranges of values ​​in these two zones.

[0056] According to certain embodiments, if after a certain number of iterations it is observed that the values ​​of the parameters are still concentrated around +3 tones, it may be decided to move to a normal rather than uniform distribution, centered on +3 tones, with a standard deviation of + / - 1 tone for example.

[0057] A new set of contents is then obtained by applying the values ​​of the parameters in this range (or these ranges) of values ​​to the initial neutral content, step E1. As mentioned previously, a certain number of values ​​can be sampled for each parameter.

[0058] Thus, unlike the methods of the prior art, the method is adaptive and progressive since an average stimulus is constructed, representative of a classification, by refining it iteratively. Starting from an average stimulus obtained during a first iteration, the values ​​of the parameters which provoke reactions in the users are evaluated, these parameter values ​​are refined by creating new content from parameter values ​​around this average value.

[0059] We then repeat steps E2 and E3 on this new set of content.

[0060] In step E4, the representative value of the parameter(s) is compared for the classification(s) obtained from this new set and the variation is evaluated.

[0061] For the parameter(s), it is determined whether the variation is less than a determined value. This determined value may be different for at least some of the parameters.

[0062] If the variation is less than this determined value, the value determined in step E3 is considered to be representative of the value of the parameter for classification, otherwise we move on to step E5 again. Thus, we move on to step E5 as long as at least one of the variations of a parameter is greater than the value determined for the parameter. If during step E4, the variation associated with one or more parameters is less than the value determined for this parameter or these parameters, then during the following step E5, these parameters may not vary (they may be fixed at the last value obtained during step E3) and only the parameter or parameters whose variation is greater than the value determined for this or these parameters are sampled in a new range.

[0063] When the variation of all the parameters is less than the determined value, we move on to step E6 in which we associate the representative value of each parameter obtained during the last passage of step E3, with each classification. We can thus obtain a filter associated with each classification comprising a value for each of the parameters representative of the classification by associating the value of said at least one parameter obtained for the last set of contents with said at least one classification (in order to obtain a representative value of said at least one parameter for said classification).

[0064] When the content is image content and prior to the application of the parameters, step E1, one or more areas of the image have been occluded during the first iteration, it is possible to occlude, during one or more subsequent iterations, one or more areas already occluded during the first iteration and to modify the parameters thereof by varying them in a new range depending on the representative value or values ​​obtained during the previous iteration. According to certain embodiments, it is also possible, during subsequent iterations, to occlude one or more areas not occluded during the previous iteration. Advantageously, according to this latter mode, it is thus possible to determine the area or areas of an image which make it possible to determine a classification reliably. For example, when the image is representative of a face, it is possible to occlude all parts of the face, except the right eye and the nose, during the first iteration.During the second iteration or during subsequent iterations, all areas of the face except the forehead and the mouth can be occluded. The classification can consist of gender, "male" or "female" and thus, during each iteration, a representative value of one or more parameters can be obtained for the unoccluded area(s) and thus, refined as iterations progress to obtain a representative value of a parameter representative of a classification for the unoccluded parts of the face. For example, the method can be executed in a loop, by reiterating it from a basic stimulus in which one or more parts are occluded. A representative value of the parameter representative of the classification is thus obtained for each unoccluded part. An advantage of this embodiment is to vary the parameter value(s) per part of the content.This can thus make it possible to determine which part of the content is the most relevant for determining the classification. Indeed, during the classification step E2, it can be envisaged, in several embodiments (including the mode in which a part of the content is not hidden), not to be able to determine a classification during step E2. For example, when the classification of step E2 consists of choosing a classification from a plurality of classifications, for example proposed to a user, the latter may possibly not decide on one of the stimuli or several, by choosing for example “not classifiable”.In this case, when parts are hidden, if for one or more non-hidden parts, the classification results in several “unclassifiable” classifications, during successive iterations of step E2 with new parameter values, this means that this or these non-hidden parts are not relevant to obtain a classification.

[0065] Advantageously, the size of the occluded regions and their number can also be dynamically adapted. Furthermore, one of the advantages of the present invention can also be to determine which areas are relevant for obtaining a classification, their size and their position in the image.

[0066] The embodiment in which certain parts of the image are hidden can allow several parameters to be varied individually, while still arriving at the end of the process when all parts have not been hidden. It is thus possible to obtain several parameters representative of a classification, one or more parameters being able to be associated with each part of the content.

[0067] For example, we can obtain one or more parameters for the eyes, one or more parameters for the mouth and so on for the different components of a face. From the representative value of each part associated for example with a "happy" classification, we then determine a filter to obtain a face classified as "happy" composed of the representative parameters of the mouth, the eyes, etc. for the "happy" classification.

[0068] In certain embodiments, from the parameters obtained in isolation, following these successive occultations, it is possible to combine a filter with a mouth classified as happy, with eyes classified as "angry"... It is therefore possible to advantageously obtain, following these successive occultations, a filter aggregating different classifications for different parts of a content.

[0069] When the content is audio content, for example, a person's voice can be isolated from a sound background. It is thus possible to obtain parameters representative of a sound background according to a certain classification and of a voice according to a certain classification, by executing the process on the voice and then on the sound background, following the isolation of each component and the execution of the process on each component. It is thus easy to obtain a filter representative of a voice of a first classification with a sound background belonging to a second classification. For example, it is possible to obtain a filter comprising the parameters that make it possible to obtain an angry voice on a happy sound background.

[0070] When the content is audio video content, we can also classify the audio by removing the video and then the other way around, by performing the process according to the different embodiments mentioned above on the audio and then on the video.

[0071] Thus, according to at least certain embodiments, said method comprises - the occultation of at least one part of said first content, - obtaining a set of contents obtained by varying in at least one range of values ​​the value of at least one parameter associated with said first content comprising said at least one occulted part, - the multiple execution of said method on said content, each execution of said method being carried out on a part of said content in which the other parts of said content are occulted, in order to obtain a representative value of said parameter associated with said classification for each part of said content, - said filter being obtained by aggregating the representative values ​​of said parameter obtained for each part.

[0072] The filter may be a filter representative of a single classification, but may thus be representative of a classification by part of the content, as mentioned above, and therefore not representative of a global classification but more customizable, by creating the filter from a first classification for a first parameter (or several) associated with a first part of the content, from a second classification for a second parameter (or several) associated with a second part and so on for the plurality of parts of the content on which the method is applied, according to the different embodiments.

[0073] Thus, from any content, one can obtain one or more other contents according to a desired classification. For example, we take any voice, or our own voice and we want to transform it into a "happy" voice. We then apply the parameter(s) as determined previously, corresponding to the happy filter, to our voice and we obtain a transformation of our voice into a happy voice.

[0074] Thus, content can be transformed into other content and the present invention can generate or synthesize artificial voices representative of certain emotions. For example, the present invention can be used in the medical field by transforming the voice of depressed patients into a joyful voice, to make them aware of their own voice. This method can also make it possible to personalize the voice used by voice assistants. It can also be envisaged to regionalize the voice according to the regions of each country. Thus, the classification can be associated with a region, according to the countries. It can also be associated with a country. A set of content including a voice speaking English can be selected and the parameters can be modified to determine classifications associated with the country such as English spoken in the United States, English spoken in Scotland, in Ireland, or French spoken in France, spoken in Belgium, spoken in Canada.

[0075] The present invention can also be used to generate voices in animated films, for example, or even to modify real voices during film editing.

[0076] The present invention can also allow avatars to be generated by playing on facial expressions, by applying the filter according to the type of classification desired. Of course, the classifications can be very fine and for example we can have various degrees in the "happy" classification, each degree being able to represent its own classification.

[0077] Beyond emotions, the scope of the method based on inverse correlation can extend to all dimensions of expressiveness (intentions, personality traits, etc.) or even to lower-level perceptual information (perception of brightness of an image, perception of spatialization of a sound, etc.).

[0078] Classifications can represent expressivities of a voice (intentions, personality traits, emotions) and can be subject to classification filters to be applied to a neutral voice.

[0079] The present invention can therefore be applied to transform real voices or artificial voices.

[0080] The figure represents an example of a device 10 that may implement the present invention according to at least certain embodiments.

[0081] The device 10 comprises in particular a processor 1, a random access memory 2, a read only memory 3 and a non-volatile memory 4. They also comprise communication means 5, and / or a user interface 6.

[0082] The read-only memory 3 constitutes a recording medium in accordance with at least one embodiment of the present disclosure, readable by the processor 1 and on which is recorded a computer program PROG in accordance with at least one embodiment of the present disclosure, comprising instructions for executing steps of the method according to at least one embodiment of the present disclosure. The program PROG defines functional modules of the device.

[0083] The present invention therefore relates to a method for creating a content transformation filter, or content classification filter, said filter comprising at least one parameter representative of at least one classification of a characteristic of the content. By characteristic of the content, we can understand voice, image, etc. The method comprising: - obtaining a first set of contents obtained by varying at least the value of some of said parameters of an initial content, - determining at least one value of said at least one parameter, obtained from information representative of said at least one classification of said contents of said first set, - successively determining at least one new set of contents and determining at least one value of said at least one parameter representative of said at least one classification from said new set of contents,as long as the variation in the value of said parameter representative of said at least one classification between a set of contents and the following set of contents is less than a determined value, said new set of contents being obtained by varying the value of said at least one parameter representative of said at least one classification in a range of values ​​determined from the value of said at least one parameter determined for said previous set of contents, - obtaining said filter by associating the value of said at least one parameter obtained for the last set of contents with said at least one classification. This for example in order to obtain a value representative of said at least one parameter for said classification.,

[0084] According to certain embodiments, the contents of said first set of contents are obtained from generic content by randomly varying said at least one parameter, said variation of said at least one parameter generating a transformation of said characteristic.

[0085] According to certain embodiments, the determination of at least one value of said at least one parameter representative of a given classification is obtained by calculating the algebraic average of the values ​​of said at least one parameter associated with the characteristic of all the contents classified in said given classification.

[0086] According to certain embodiments, the determined range of values ​​is centered on the representative value of said at least one parameter of said at least one classification.

[0087] In some embodiments, the determined range of values ​​is reduced for a new set of content relative to a previous set of content.

[0088] According to certain embodiments, when at least two parameters representative of a classification of a characteristic of a content are determined, - a representative value is determined for each of the at least two parameters representative of said at least one classification, a new set of contents and at least one value of said at least two parameters representative of said at least one classification are determined from said new set of contents as long as the variation in the value of said at least two parameters representative of said at least one classification between a set of contents and the following set of contents is less than a determined value, said new set of contents being obtained by varying the value of said at least two parameters representative of said at least one classification in a determined range of values, for each of said at least two parameters,from the value of said at least two parameters determined for said previous set of contents, - the association of the value of said at least two parameters obtained for the last set with said at least one classification in order to obtain a representative value for each of said at least two parameters for said classification.,

[0089] According to certain embodiments, the information representative of classification of said contents of said first set and of said at least one new set is obtained from a user.

[0090] According to certain embodiments, the method comprises applying said filter to said at least one parameter associated with said characteristic of a target content so that said characteristic of said target content is associated with said classification.

[0091] According to certain embodiments, said characteristic is an audio characteristic and said parameters are parameters associated with the prosodic profile of the voice, and preferably representative of a voice pitch, a speaking rate and a sound level.

[0092] According to certain embodiments, said characteristic is a visual characteristic and said parameters are parameters associated with the brightness of the image, and preferably representative of gray levels.

[0093] The present invention also relates, according to certain aspects, to a device for creating a filter for transforming content, said filter comprising at least one parameter representative of at least one classification of a characteristic of the content, said device comprising one or more processors configured together or separately to: - obtain a first set of contents obtained by varying at least the value of some of said parameters of an initial content, - determine at least one value of said at least one parameter, obtained from information representative of said at least one classification of said contents of said first set, - successively determine at least one new set of contents and determine at least one value of said at least one parameter representative of said at least one classification from said new set of contents,as long as the variation in the value of said parameter representative of said at least one classification between a set of contents and the following set of contents is less than a determined value, said new set of contents being obtained by varying the value of said at least one parameter representative of said at least one classification in a range of values ​​determined from the value of said at least one parameter determined for said previous set of contents, - obtaining said filter by associating the value of said at least one parameter obtained for the last set of contents with said at least one classification. This can advantageously make it possible to obtain a value representative of said at least one parameter for said classification.,

Claims

Method for creating a filter for classifying content, from a first content comprisinga) obtaining (E1) a set of contents obtained by varying in at least one range of values the value of at least one parameter associated with said first content,b) obtaining (E2) at least one item of classification information for said contents,c) associating (E6) at least one value representative of said at least one parameter with said classification- repeating steps a) to c) as long as the variation (E4), between two iterations, of the at least one value representative of said at least one parameter associated with said classification is less than a determined value, said at least one range of values being determined (E5) during each iteration as a function of the at least one representative value associated (E3) with said classification, said filter comprising the at least one value of said at least one parameter obtained during the last iteration. Method according to claim 1 wherein the contents of said set of contents are obtained from said first content by randomly varying said at least one parameter, said variation of said at least one parameter generating a transformation of at least one characteristic of said first content. Method according to one of the preceding claims in which the at least one representative value of said at least one parameter associated with a given classification is obtained by calculating the algebraic average of the values of said at least one parameter of the contents classified in said given classification. Method according to one of the preceding claims in which the at least one range of values determined during the second and subsequent iterations is centered on the at least one value representative of said at least one parameter associated with said at least one classification obtained in the previous iteration. Method according to claim 4 in which the determined range of values is reduced at each iteration. Method according to one of the preceding claims in which, when associating at least the value of two parameters with a classification - associating a representative value for each of said at least two parameters associated with said at least one classification - iterating steps a) to c) as long as the variation in the value of at least one of said at least two parameters associated with said at least one classification between two iterations is greater than a determined value. Method according to one of the preceding claims in which the classification information of said contents is obtained from a user. Method according to one of the preceding claims comprising applying said filter to said at least one parameter associated with said characteristic of a target content so that said characteristic of said target content is associated with said classification. Method according to one of the preceding claims wherein- when said content is audio content, said parameters are parameters associated with the prosodic profile of the voice, and preferably representative of a voice pitch, a speaking speed and a sound level and / or- when said content is image content, said parameters are parameters associated with the brightness of the image, and preferably representative of gray levels. Device for creating a filter for classifying content, from a first content comprising one or more processors configured together or separately for a) obtaining a set of contents obtained by varying in at least one range of values the value of at least one parameter associated with said first content, b) obtaining at least one classification information of said contents, c) associating at least one value representative of said at least one parameter with said classification - repeating steps a) to c) as long as the variation, between two iterations, of the at least one value representative of said at least one parameter associated with said classification is less than a determined value, said at least one range of values being determined during each iteration as a function of the at least one representative value associated with said classification, said filter comprising the at least one value of said at least one parameter obtained during the last iteration.

Images