JP4992295B2 - Information processing device - Google Patents

Description

The present invention relates to an information processing equipment, in particular, for the selection process of the desired content.

  A process is known in which an information processing apparatus stores content and reproduces the stored content. Here, the content is a music, a voice spoken by a person, a mechanical sound, a sound that occurs in the natural world, for example, a sound that is used as a background sound, a moving image, a signal represented as a function of time, or a still image, Hereinafter, a description will be given using music as an example.

  When the user of the apparatus selects a desired song, for example, a favorite song, it depends on bibliographic items such as the name of the song and the artist name. However, an increase in the storage capacity of the storage unit in which the music is stored causes a large number of music to be stored, and the user of the apparatus feels difficulty in selecting a desired music. In particular, in a portable device, the display screen of the display unit is small, and only a list of bibliographic items of a small number of songs can be displayed, and the difficulty is remarkable.

  Furthermore, the user needs a long time in order to know whether or not the unknown music is the desired music, other than just listening to the music. Therefore, there is a problem that it is difficult to select the desired music piece without knowing that the desired music piece is stored while the desired music piece is stored in the apparatus.

Therefore, for example, the feature value of the music stored in the apparatus is extracted, the impression value is calculated from the extracted feature value, and stored in association with the music. On the other hand, the user inputs the degree of impression of desired music. Then, the apparatus calculates an impression value from the degree of the input impression, searches for a stored impression value having a small distance from the calculated impression value, and selects a song associated with the searched impression value. A process of searching for a desired music piece is known (for example, see Patent Document 1).
JP 2002-278547 A (pages 18-19, FIG. 1)

  However, in the method disclosed in Patent Document 1 described above, the degree of impression is a predetermined decimal property, and it is impossible to accurately represent the user's desired degree. As a result, there has been a problem that a desired music cannot be searched.

The present invention has been made to solve the above problems, depending on the example of the examples and / or selected undesired content was the desired content selected, the information processing equipment to search for desired content The purpose is to provide.

In order to achieve the above object, an information processing apparatus according to the present invention searches a content storage unit for storing content and a first type of the content from the content storage unit, and obtains the first type obtained by the search. A feature amount of the first type of content obtained by calculating the feature amount of the first group including the first type of content from the feature amount vector of the content, and obtained by the search with a representative vector of the feature amount vector of the content as a center. A second type of content having the smallest distance from a vector symmetrical to the vector and / or a feature vector whose distance is smaller than a predetermined value is searched from the content storage means, and the searched second type A group feature quantity calculating means for calculating a feature quantity of the second group consisting of the second type of content from a feature quantity vector of the content, and the content A first distance between the feature vector of the content stored in the storage means and the feature quantity of the first group calculated by the collective feature quantity calculation means; the feature vector of the content; and the collective feature quantity A second distance between the second group and the feature amount of the second group calculated by the calculating means is calculated, and the content whose first distance is smaller than the second distance is searched as the first type of content. And / or content search means for searching for the content having the first distance larger than the second distance as the second type of content .

According to the present invention, it can be provided depending on the example of the examples and / or selected undesired content was the desired content selected, the information processing equipment to find the desired content.

  Embodiments of an information processing apparatus and a content search program according to the present invention will be described below with reference to the drawings.

(First embodiment)
FIG. 1 is a block diagram showing a configuration of an information processing apparatus according to the first embodiment of the present invention to which the content search program according to the first embodiment of the present invention is applied. This information processing apparatus is a computer that operates using a program, and includes a control unit 11 that controls the entire apparatus, a display unit 12, an input device 13, a music registration unit 21, a feature extraction unit 22, The content storage unit 23, the collective feature amount calculation unit 24, the dictionary storage unit 25, the music search unit 26, the music playback unit 27, and the music playback speaker 27 a are included.

  The content storage unit 23 stores content 23a. As will be described in detail later, the content 23a includes digitized content and information indicating the characteristics of the content. The dictionary storage unit 25 stores a dictionary 25a. As will be described in detail later, the dictionary 25a includes information indicating characteristics of a desired content group.

  The operation of each unit of the information processing apparatus according to the embodiment of the present invention configured as described above will be described.

  The display unit 12 is controlled by the control unit 11 to display characters / numbers and image data including a cursor, and the displayed data is controlled in response to an input operation from the input device 13 or the like. Switching is performed upon receiving an instruction from the unit 11.

  The input device 13 includes keys including numeric keys for inputting numerals, hiragana characters, alphabetic characters, and symbol characters, character keys, and a plurality of function keys including cursor movement keys and scroll keys. When a key of the input device 13 is pressed, the identifier of the key is notified to the control unit 11 and displayed as a character on the display unit 12 by the control unit 11 and notified to each unit or control is performed. . The input device 13 may include input elements other than keys, such as a mouse and a touch panel.

  The music registration unit 21 receives the music data, and causes the feature extraction unit 22 to extract the feature quantity vector of the received music data. Then, the received music data and the feature vector extracted by the feature extraction unit 22 are stored in the content storage unit 23 as the content 23a.

  The feature extraction unit 22 receives music data, extracts a feature vector of the received music data, and transmits the extracted feature vector.

  The collective feature amount calculation unit 24 reads out the feature amount vector of the content 23a specified as an example of the desired content 23a and / or the content 23a specified as an example of the undesired content 23a from the content storage unit 23, and the features thereof From the quantity vector, a vector and a matrix indicating a group of desired contents 23a and a vector and a matrix indicating a group of undesired contents 23a are calculated and stored in the dictionary storage unit 25 as a dictionary 25a.

  The music search unit 26 refers to the dictionary 25a stored in the dictionary storage unit 25, and calculates a desired degree of the content 23a from the feature amount vector of the content 23a. Then, the content storage unit 23 is searched for the content 23a having the designated desired degree.

  The music playback unit 27 reads the music data of the content 23a searched by the music search unit 26 from the content storage unit 23, plays back the read music data, and outputs it from the speaker 27a.

  Hereinafter, an operation related to the search for the content 23a having a specified desired degree in the information processing apparatus according to the present embodiment will be described.

  FIG. 2 shows an example of the configuration of the content 23a. The content 23a includes a content identifier 23b, content data 23c, a name 23d, an artist name 23e, an album name 23f, and a feature vector 23g.

  The content identifier 23b is an identifier that uniquely identifies the content 23a. The content data 23c is data of the content 23a. For example, the content data 23c is data obtained by encoding digital data indicating music. However, it may not be encoded.

  The name 23d is the name of the content 23a. The artist name 23e is the name of the performer or singer of the music when the content 23a is a music. Further, when the content 23a is an utterance, it is a speaker of the utterance. The album name 23f is the name of the album including the content 23a.

  The feature quantity vector 23g is a vector indicating the feature of the content data 23c. The feature quantity vector 23g includes M first feature quantity vectors 23g1, second feature quantity vectors 23g2,..., Mth feature quantity vectors 23gM. Consists of vectors. The i-th feature vector 23gi (1 ≦ i ≦ M) is an N-dimensional vector and includes elements Xi1, Xi2,.

  The content 23a includes the name 23d, the artist name 23e, and the album name 23f as bibliographic items, but is not limited thereto. For example, other bibliographic items such as a composer name, a songwriter name, and a seller name may be included. In addition, these bibliographic items are not essential information, and information may not be stored. As will be described later, the content identifier 23b, the content data 23c, and the feature vector 23g are indispensable information in order to search and reproduce the content 23a according to a desired degree.

  FIG. 3 shows an example of the configuration of the dictionary 25a. The dictionary 25a includes a dictionary identifier 25b, a dictionary name 25c, a first centroid vector 25d, an inverse matrix 25e of the first variance-covariance matrix, a second centroid vector 25f, and a second variance-covariance. And an inverse matrix 25g of the matrix.

  The dictionary identifier 25b is an identifier that uniquely identifies the dictionary 25a. The dictionary name 25c is the name of the dictionary 25a. The first centroid vector 25d and the second centroid vector 25f are both N-dimensional vectors, and are vectors having the same number of dimensions as the feature amount vector 23g included in the content 23a. The inverse matrix 25e of the first variance-covariance matrix and the inverse matrix 25g of the second variance-covariance matrix are both N × N-dimensional matrices.

  The first centroid vector 25d and the inverse matrix 25e of the first variance-covariance matrix indicate the feature amount of the group including the content 23a designated as desired, and the content designated as desired. The barycenter and distribution of each vector included in the feature quantity vector 23g of 23a are shown.

  The second center-of-gravity vector 25f and the inverse matrix 25g of the second variance-covariance matrix indicate the feature amount of the group including the content 23a designated as undesired, and the feature amount of the content 23a designated as undesired The barycenter and distribution of each vector included in the vector 23g are shown.

  The music registration unit 21 receives the content data, the name, the artist name, and the album name, and the received content data, the name, the artist name, and the album name are the content data 23c and the name 23d, respectively. And artist name 23e and album name 23f. Here, the content data is music data encoded by the MP3 system or the AAC system, but is not limited thereto. When the content data is analog data, the music registration unit 21 converts the analog data into digital data and sets the encoded data in the content data 23c.

  Here, as a means for receiving, for example, the apparatus includes a communication unit (not shown) that communicates with the cellular communication network and receives via the cellular network, but is not limited thereto. The apparatus may have an Internet communication unit (not shown) and receive via the Internet. Further, the apparatus may have an e-mail transmission / reception unit (not shown), and may be received as a body of an e-mail received by the e-mail transmission / reception unit or a file attached to the e-mail. Furthermore, it may be received via a removable storage medium.

  Further, the name 23d, the artist name 23e, and the album name 23f may be information input by a predetermined operation of the input device 13.

  Next, the music registration unit 21 activates the feature extraction unit 22 using the content data 23c as a parameter. Then, the M N-dimensional vectors transmitted from the feature extraction unit 22 are received. Then, the content storage unit stores the content 23a obtained by setting the M N-dimensional vectors transmitted from the feature extraction unit 22 in the feature quantity vector 23g and further setting the content identifier 23b that uniquely identifies the content 23a. 23.

  FIG. 4 shows a flowchart of an operation in which the feature extraction unit 22 receives the content data 23c and creates M N-dimensional vectors that are the feature amount vectors 23g.

  The feature extraction unit 22 is activated by the music registration unit 21 using the content data 23c as a parameter and starts operating (step S22a). Then, the content data 23c given as a parameter is divided at a predetermined time interval, and it is determined which M frames are selected assuming that a plurality of frames are created (step S22b). Here, the predetermined time interval is suitably about several seconds. Further, as will be described later, a feature vector is calculated from the selected frame.

  Further, when M frames are selected, the M feature amount vectors calculated from each of the frames are vectors that accurately indicate the characteristics of the entire given content data 23c. Therefore, M is large. Moderately good. However, if it is large, the processing amount of the operation described below increases. Therefore, although M depends on the reproduction time of the content data 23c, etc., it is appropriate that M is about several tens.

  In addition, since the feature vector is a vector indicating the characteristics of the entire content data 23c, it is appropriate to select frames located at approximately equal intervals in time among all the frames of the content data 23c. . However, since the prelude at the beginning and the end after the end of the music may have a tone different from that of the entire content data 23c, the frames related to the predetermined time at the beginning and end may not be selected.

  Next, the feature extraction unit 22 calculates decoded data corresponding to the M frames selected in step S22b in the content data 23c given as parameters, and for the decoded frames, Pre-processing for monauralization and low-pass filtering is performed and stored in the feature extraction unit 22 (step S22c).

  The content data 23c is encoded, and in order to decode the data, the data is subjected to, for example, Huffman decoding to obtain frequency information. Then, the frequency information is subjected to frequency time conversion by inverse quantization processing to obtain waveform data. Here, the feature extraction unit 22 may store the obtained frequency information and use it for processing described later. According to the method using the frequency information, frequency time conversion is unnecessary, and the amount of calculation can be reduced.

  Then, the frames in the M frames stored in the feature extraction unit 22 are sequentially read one by one (step S22d). It is checked whether or not the frame is exhausted and could not be read (step S22e). When the frame is not exhausted and read, the feature extraction unit 22 calculates the average value and standard deviation of the number of zero crossings for the read frame. (Step S22f).

That is, the feature extraction unit 22 divides the frame into NN pieces at a predetermined time interval, and creates NN subframes. The amplitude of the discrete waveform signal in the nth subframe (0 ≦ n ≦ NN−1) is S (n, t). Here, t is a variable indicating the time in the subframe, and 0 ≦ t ≦ T−1. Then, the zero crossing number Zc (n) of the n-th subframe is expressed by the amplitude at a certain time t (1 ≦ t ≦ T−1) and the time before that time, as shown in the following (Formula 1). This is calculated as the number when the sign of the amplitude at t-1 is different. Next, the average value and standard deviation of the calculated NN zero crossing numbers Zc (n) are calculated (0 ≦ n ≦ NN−1).

  Next, the feature extraction unit 22 calculates the average value and the standard deviation of the mel frequency cepstrum coefficients related to the frame read in step S22d (step S22g). The calculation of the mel frequency cepstrum coefficient will be described later.

  Next, the feature extraction unit 22 normalizes a vector whose elements are the average value and standard deviation of the number of zero crossings calculated in step S22f and the average value and standard deviation of the mel frequency cepstrum coefficients calculated in step S22g. The feature vector x before normalization is used, and this feature vector is normalized (step S22h). Note that the dimension of the feature vector is determined by the number of mel frequency cepstrum coefficients described later, and the dimension is assumed to be N-dimensional.

  In normalization, the feature extraction unit 22 includes a representative vector μ of M feature vector vectors before normalization regarding a plurality of frames of the content 23a as much as possible, and between the feature vector and the representative vector μ. The variance matrix Σ is stored. Here, the representative vector μ is an average vector.

  Further, the representative vector μ and the variance matrix Σ may be obtained by obtaining a predetermined number of code vectors from the feature quantity vector by vector quantization, and obtaining the average vector of the obtained code vectors and the variance matrix Σ. Here, the code vector is a predetermined number of vectors for which the distance between the feature vectors and the code vector closest to each feature vector is obtained and the sum of the distances is minimized. However, the present invention is not limited to this.

  The distance between the two vectors is obtained as, for example, the Euclidean distance, that is, the square root of the sum of the squared differences obtained by squaring the difference between the corresponding elements of the vector, but is not limited thereto. The number of code vectors is suitably about 1/50 to 1/20 (1/50 to 1/20) of the number of feature vectors, but is not limited to this and is determined in advance. May be.

  According to this method of obtaining the representative vector μ and the variance matrix Σ through vector quantization, the representative vector μ and the variance matrix Σ that are appropriate from the biased content 23a, that is, calculated from an infinitely large number of content 23a. You can ask for something close to Therefore, when the number of feature quantity vectors is small, the effect of this method is remarkable.

  The representative vector μ and the variance matrix Σ may be calculated using feature amount vectors calculated from all the contents 23 a stored in the content storage unit 23. Further, it may be calculated using a feature amount vector calculated from the content 23a by the feature extraction unit 22 in the past.

  Note that the variance matrix Σ is a matrix composed only of diagonal components, that is, components other than the diagonal components may be zero. According to this processing that the variance matrix Σ is a matrix composed only of diagonal components, the variance matrix Σ can be expressed as a vector, and it is possible to reduce the storage capacity used and the calculation amount. is there.

  Further, the representative vector μ and the dispersion matrix Σ may be received from the outside of the apparatus, for example. Further, it may be stored at the time of shipment of the apparatus. Note that the average vector of a plurality of vectors is a vector having an average value for each element of the plurality of vectors as an element.

The feature extraction unit 22 represents the normalized feature vector y from the inverse matrix of the variance matrix Σ and the feature vector x before normalization obtained in step S22h, as in (Equation 2) below. It is obtained by multiplication with a vector obtained by subtracting the vector μ and stored in the feature extraction unit 22.

  Then, the feature extraction unit 22 returns to the operation of sequentially reading the frames in step S22d. On the other hand, when the frame is exhausted in step S22e, the feature extraction unit 22 activates the feature extraction unit 22 with M N-dimensional feature amount vectors obtained in step S22h and stored in the feature extraction unit 22. The data is transmitted to the processing unit (step S22i), and the operation ends (step S22j).

  FIG. 5 shows a flowchart of the operation for calculating the average value and standard deviation of the mel frequency cepstrum coefficients in step S22g. The feature extraction unit 22 starts this operation (step S22m), emphasizes the high frequency range (step S22n), and multiplies the Hanning window function to perform discrete Fourier transform, thereby obtaining the discrete power spectrum P (n, f ) Is calculated. Here, f is a variable indicating the frequency, and 0 ≦ f ≦ F (step S22o).

  Next, the feature extraction unit 22 divides the discrete power spectrum P (n, f) into I sections so that the adjacent sections are overlapped by half at the mel frequency interval, and the discrete power spectrum of each I section is divided. Calculate the logarithm of the sum. Then, the I values are subjected to discrete cosine transform, and I-th order I-1 order mel frequency cepstrum coefficients are calculated for each subframe (step S22p). Then, an average value and a standard deviation of NN values for each subframe of each coefficient are calculated as a feature amount of the frame (step S22q), and the operation is ended (step S22r).

  Note that the feature extraction unit 22 performs processing 2 of calculating the average value and standard deviation of the number of zero crossings described in step S22f, and processing of calculating the average value and standard deviation of the mel frequency cepstrum coefficients described in step S22g. Obviously, any one of the two processes may be performed first, and which process may be performed later.

With the operation described above, the feature extraction unit 22 transmits a feature quantity vector calculated from the following four kinds of feature quantities. Here, the number of zero crossings indicates the center frequency of the content data 23c, the mel frequency cepstrum coefficient indicates the slope of the spectrum, and the standard deviation of these numbers indicates changes in these numbers over time. The feature amount vector appropriately indicates the feature of the content data 23c, but the feature amount is not limited thereto.

  For example, the feature extraction unit 22 uses the degree of temporal change of the entire discrete power spectrum P (n, f) or for each band divided into the I section as the feature amount in addition to or instead of the feature amount. May be.

  In addition, the feature extraction unit 22 transmits the normalized feature vector. However, the present invention is not limited to this, and the feature vector before normalization may be transmitted. As will be described later, for example, the collective feature amount calculation unit 24 calculates a vector that is symmetric about the feature amount vector and the representative vector μ by calculating an inverse vector of the normalized feature amount vector. To do. This process can be replaced by a process of subtracting the feature vector before normalization from a vector obtained by multiplying the representative vector μ by 2.

  Also, according to the multiplication of the variance matrix Σ with the inverse matrix, a feature vector whose size is normalized is used, but by not performing the multiplication, the unnormalized size of the vector is used. May be used.

  Next, the operation of the collective feature quantity calculation unit 24 will be described. FIG. 6 shows a flowchart of the operation of the collective feature quantity calculation unit 24. The collective feature amount calculation unit 24 is activated by the music playback unit 27 and starts operating (step S24a). And the information which identifies the desired content 23a is received from the music reproduction part 27 (step S24b), and the information which identifies the content 23a which is not desired is received (step S24c).

  Here, the information for identifying the content 23a is the content identifier 23b, but is not limited thereto. Any of the name 23d, the artist name 23e, the album name 23f, or a combination thereof may be used. In this case, if a plurality of content identifiers 23b are obtained by searching the content 23a with one piece of identifying information, it is assumed that the plurality of content identifiers 23b are received.

  In addition, a desired degree or an undesired degree may be added to the information for identifying the content 23a. The degree is a positive number, and the larger the number, the greater the degree.

  Further, as described below, information for identifying the desired content 23a is not received in step S24b (empty information is received), or information for identifying the undesired content 23a is received in step S24c. It is not necessary, but it is essential that either one is received.

  Next, the collective feature quantity calculation unit 24 obtains a vector indicating the feature of the desired content 23a by reading the feature quantity vector 23g and further creating a new feature quantity vector (step S24d). That is, when information for identifying the desired content 23a is received in step S24b, the content 23a is searched using that information as a search key, and the searched feature vector 23g is obtained.

  Further, when information for identifying the undesired content 23a is received in step S24c, the collective feature amount calculation unit 24 searches the content 23a using the information as a search key, and calculates an inverse vector of the searched feature amount vector 23g. A feature vector may be obtained by creating the feature vector.

  Here, the feature quantity vector 23g is composed of M N-dimensional vectors. The inverse vector of the feature quantity vector 23g is M N-dimensional feature quantity vectors composed of the inverse vectors of the M vectors. In addition, when the information which identifies the desired content 23a is not received by step S24b, it is essential to obtain the feature-value vector by producing the said reverse vector.

  In addition, when the information which identifies the content 23a which is not desired is received by step S24c, although the collective feature-value calculation part 24 created the inverse vector, it is not restricted to an inverse vector. You may round off suitably each element of the vector which comprises an inverse vector. When each element is expressed in binary, a predetermined lower bit may be set to 0.

  Further, each element of the vector constituting the inverse vector may be multiplied by a positive number. Multiplying a positive number less than 1 reduces the distance between the feature vector 23g of the desired content 23a received in step S24b and the multiplied vector, which is desired by the music search unit 26 described later. The range of the content 23a determined as can be narrowed. Further, if the number exceeding 1 is multiplied, the distance increases, so that the range of the content 23a determined to be desired can be widened.

  Further, it may be a vector that is closer to the inverse vector than the searched feature vector 23g, that is, a vector that has a distance from the searched feature vector 23g larger than the distance from the inverse vector.

  Note that the calculation of the distance is a calculation of the distance between the M vectors and the M vectors. Therefore, for example, the distance between each of the M vectors and the other M vectors, for example, the Euclidean distance may be calculated, and the sum of the M × M Euclidean distances may be set as the distance.

  Alternatively, one is selected and combined from one M vector and one from the other M vectors, and the distance between these two vectors, for example, the Euclidean distance is calculated. The total sum of the calculated M Euclidean distances may be the minimum sum obtained by changing the combination. Moreover, it is good also as the distance between the average vectors of each M vector. Here, the average vector is a vector that calculates an average value for each element of the vector and uses the calculated average value as an element.

  Then, the collective feature value calculation unit 24 calculates the centroid vector and the inverse matrix of the variance-covariance matrix from the feature value vector obtained in step S24d (step S24e). The center-of-gravity vector is an N-dimensional vector, and is an average vector of vectors obtained by multiplying each feature amount vector by the degree received in step S24b or step S24c, but is not limited thereto. For example, the degree may be multiplied only by the read feature value vector 23g. Further, the center-of-gravity vector may be an average vector without multiplying the degree. The inverse matrix of the variance-covariance matrix is an N × N-dimensional matrix.

  The variance-covariance matrix is a matrix composed only of diagonal components, that is, components other than the diagonal components may be zero. In this case, the inverse matrix of the variance-covariance matrix is a matrix composed of only diagonal components, that is, the components other than the diagonal components are zero. According to the processing assuming that the variance-covariance matrix and the inverse matrix of the variance-covariance matrix are matrices composed only of diagonal components, these matrices can be expressed as vectors, and the storage capacity used can be reduced. The amount of calculation can be reduced.

  Next, the collective feature quantity calculation unit 24 obtains the feature quantity vector of the content 23a that is not desired by reading the feature quantity vector 23g or by newly creating it (step S24f). This operation is the same as the operation for obtaining the feature vector of the desired content 23a in step S24d already described.

  That is, the information for identifying the undesired content 23a received in step S24c is used by using the information for identifying the undesired content 23a received in step S24c in place of the information for identifying the desired content 23a received in step S24b. Instead of this, the description is omitted because it is the same except that information for identifying the desired content 23a received in step S24b is used.

  Then, the collective feature quantity calculation unit 24 calculates the centroid vector and the inverse matrix of the variance-covariance matrix from the feature quantity vector obtained in step S24f (step S24g). The calculation of the centroid vector and the inverse matrix of the variance-covariance matrix is as described in step S24e. For example, the degree may be multiplied only by the read feature vector 23g.

  Next, the collective feature amount calculation unit 24 sets the centroid vector obtained in step S24e to the first centroid vector 25d, and the inverse matrix of the variance-covariance matrix obtained in step S24e to the first variance-covariance matrix. In the inverse matrix 25e, the centroid vector obtained in step S24g is set to the second centroid vector 25f, and the inverse matrix of the variance-covariance matrix obtained in step S24g is set to the inverse matrix 25g of the second variance-covariance matrix. .

  Furthermore, the collective feature amount calculation unit 24 inputs information for identifying the dictionary 25a by a predetermined operation of the input device 13, sets the information for identifying the input dictionary 25a to the dictionary name 25c, and further stores the dictionary 25a. The dictionary 25a obtained by setting the uniquely identifying dictionary identifier 25b is stored in the dictionary storage unit 25 (step S24h), and the operation is terminated (step S24i).

  Note that the collective feature amount calculation unit 24 needs to perform the three-stage process regarding the desired content 23a described in step S24b, step S24d, and step S24e in this order. Further, the three-stage process relating to the undesired content 23a described in step S24c, step S24f, and step S24g needs to be performed in this order. However, it is obvious that any of these two processes can be performed first and which can be performed later. Also, it is obvious that the execution of one stage of one process and the execution destination of the other stage of the other process are arbitrary.

  Here, in the above description, it is described whether it is desired or not, but it is not limited to this. For example, the user designates whether or not each of the music pieces classified into a certain type according to the impression at the time of listening or the music that the user prefers to listen to in a certain situation is desired independently, and the dictionary differs depending on each. 25a may be created. In this case, the dictionary name 25c may be, for example, “feeling reminiscent of the sea”, “feeling perfect when commuting”, or the like. Further, what is desired corresponds to dislike, and what is not desired may correspond to favor, and does not cause any trouble.

  The above explanation is an operation when the collective feature quantity calculation unit 24 creates the dictionary 25a when the dictionary 25a is not created. However, the operation of the collective feature quantity calculation unit 24 is limited to this. It is not a thing. That is, when the dictionary 25a is created and stored in the dictionary storage unit 25, information for identifying the content 23a that has not been used to create the stored dictionary 25a is received in step S24b and step S24c, and received. Learning processing may be performed in which the stored dictionary 25a is corrected based on the information that has been stored.

  In this learning process, the collective feature quantity calculation unit 24 stores information for identifying the content 23 a used for creating the stored dictionary 25 a in the collective feature quantity calculation unit 24 or stores it in the dictionary storage unit 25. The dictionary 25a may be newly created from the stored information and the received information.

  Further, the total weight value used for calculating the centroid vector and the variance-covariance matrix when the stored dictionary 25 a is created are stored in the collective feature value calculation unit 24 or stored in the dictionary storage unit 25. The stored total value and matrix may be updated with the received information to create the dictionary 25a. According to this update process, the amount of calculation can be reduced as compared with a newly created process.

  The collective feature amount calculation unit 24 is activated by the music reproduction unit 27, but is not limited thereto. It may be activated by another processing unit. In this case, the collective feature quantity calculation unit 24 receives information for identifying the content 23a received in step S24b and step S24c from the activated processing unit. Further, at the time of activation, the processing may be performed only when it is determined that the activation is obtained by a predetermined key operation of the input device 13. Further, the collective feature amount calculation unit 24 may be activated by a predetermined operation of the input device 13.

  Here, referring to FIG. 7, the collective feature value calculation unit 24 creates a dictionary 25a from the received feature value vector of the desired content 23a and the feature value vector of the undesired content 23a. An outline of the data flow will be described. This description is an outline description and does not necessarily cover all cases.

  From the identifier of the desired content 23a received in step S24b, the feature quantity vector 23g of the content 23a is searched in step S24d. Along with the retrieved feature vector 23g, the feature vector 23g of the content 23a is obtained from the identifier of the undesired content 23a received in step S24c, and an inverse vector of the vector is created in step S24d. It is used to calculate the centroid vector of S24e and the inverse matrix of the variance-covariance matrix.

  Further, in step S24f, the feature vector 23g of the content 23a is retrieved from the identifier of the undesired content 23a received in step S24c. Along with the searched feature vector 23g, a feature vector 23g of the content 23a is obtained from the identifier of the desired content 23a received in step S24b, and an inverse vector of the vector is created in step S24f. It is used to calculate the centroid vector of S24g and the inverse matrix of the variance-covariance matrix.

  The centroid vector calculated in step S24e, the inverse matrix of the variance-covariance matrix, and the centroid vector calculated in step S24g and the inverse matrix of the variance-covariance matrix are stored as the dictionary 25a.

  Next, the operation of the music search unit 26 will be described. The music search unit 26 is activated by the control unit 11 according to a predetermined operation of the input device 13 and starts operating. Then, in accordance with a predetermined operation of the input device 13, the content 23a is searched by using any one of the name 23d, the artist name 23e, and the album name 23f, or a combination thereof as a search key, and the searched content 23a is ordered.

  Alternatively, the contents 23a are ordered according to a desired degree with reference to the dictionary 25a. Furthermore, these two orderings are combined and ordered. Then, one or a plurality of sets of the content 23a name 23d, artist name 23e, and album name 23f ordered in a predetermined order are displayed on the display unit 12.

  When one of one or more of the displayed name 23d, artist name 23e, and album name 23f is selected according to a predetermined operation of the input device 13, the music search unit 26 reproduces the music. The unit 27 is activated, and the content data 23c of the selected content 23a is transmitted to the music playback unit 27 for playback.

  Note that, depending on the size of the display screen of the display unit 12, only a part of the display is displayed, and the remaining part is displayed by scrolling. In addition, the music search unit 26 is based on random numbers from a part to be displayed so that the storage capacity required for the content 23a is within a predetermined reproduction time, within a predetermined reproduction time, or with a predetermined You may select the number of music pieces and display the selected one.

  Further, the display is not limited to the display unit 12. A playlist composed of the concatenation of the content identifiers 23b of the content 23a to be displayed may be created and stored in the content storage unit 23. Further, the communication unit may be controlled to transmit.

  When the content data 23c is being played back by the music playback unit 27, if the content 23a is input as desired by a predetermined operation of the input device 13, the music search unit 26 and the content identifier 23b The fact is stored in the music search unit 26. When the desired degree is input together, the degree is also stored. When the content 23a is input as not desired, the music search unit 26 stores the content identifier 23b and the fact that the content 23a is not desired in the music search unit 26. When an undesired degree is input together, the degree is also stored.

  When the content data 23c is being reproduced by the music reproducing unit 27, when reproduction of the different content data 23c is instructed by a predetermined operation of the input device 13 within a predetermined elapsed time after the reproduction is started, The music search unit 26 may consider that the content 23a is input if it is not desired. This is because it is determined that it is not desirable to stop listening in a short time.

  In FIG. 8, the music search unit 26 refers to the dictionary 25a to order the contents 23a according to a desired degree, and includes the contents 23a, the artist name 23e, and the album name 23f. The flowchart of the operation | movement which displays one or more on the display part 12 is shown.

  The music search unit 26 refers to the dictionary 25a by a predetermined operation of the input device 13 and orders the content 23a according to a desired degree, and the name 23d of the content 23a, the artist name 23e, Then, the operation of displaying one or more of the album names 23f on the display unit 12 is started (step S26a).

  Then, the music search unit 26 searches the dictionary storage unit 25 for a dictionary 25a in which the dictionary name input by the operation of the input device 13 and the dictionary name 25c are the same, and the first centroid vector 25d of the searched dictionary 25a. Then, the inverse matrix 25e of the first variance-covariance matrix, the second centroid vector 25f, and the inverse matrix 25g of the second variance-covariance matrix are read (step S26b).

  Next, the music search unit 26 sequentially reads the feature amount vector 23g of the content 23a from the content storage unit 23 (step S26c). Here, it is checked whether or not the content 23a can be read due to exhaustion (step S26d).

  When the content 23a is read without exhaustion, the music search unit 26 includes each of the first to Mth feature quantity vectors 23g1 to 23gM constituting the feature quantity vector 23g and the first read in step S26b. The distance (referred to as the first distance) to the group (referred to as the first distance) indicated by the centroid vector 25d and the inverse matrix 25e of the first variance-covariance matrix, and the similarity to the weight Degree (referred to as first similarity) is calculated. Here, the first similarity is smaller as the first distance is larger, and is larger as the first distance is smaller.

  Further, the music search unit 26 includes each of the first to Mth feature quantity vectors 23g1 to 23gM, the second centroid vector 25f read in step S26b, and an inverse matrix of the second variance-covariance matrix. The distance (referred to as the second distance) with the group (referred to as the second group) indicated by 25g and the similarity (referred to as the second similarity) with the group are calculated. Here, the second similarity is smaller as the second distance is larger, and is larger as the second distance is smaller (step S26e).

For example, the m-th feature vector 23gm is xm (1 ≦ m ≦ M), the distance to the i-th group is the Mahalanobis distance, and the i-th similarity is a value obtained by adding 1 to the Mahalanobis distance. The reciprocal is set (i = 1, 2). That is, if the first centroid vector 25d is vg1, the inverse matrix 25e of the first variance-covariance matrix is D1, the second centroid vector 25f is vg2, and the inverse matrix 25g of the second variance-covariance matrix is D2. The i-th similarity Si (i = 1, 2) is calculated by the following (Formula 3).

  In the above description, the Mahalanobis distance is used as the i-th distance to the i-th group, but the present invention is not limited to this. For example, the i-th distance may be a distance between the m-th feature vector 23gm and vgi (the first centroid vector 25d or the second centroid vector 25f), for example, the Euclidean distance ( i = 1, 2).

  When the distance between vectors is used as the i-th distance, the inverse matrix 25e of the first variance-covariance matrix and the inverse matrix 25g of the second variance-covariance matrix are not used for calculating the distance. do not do. Therefore, the amount of calculation can be reduced. On the other hand, if the distance between the vector and the group such as Mahalanobis distance is used, if the inverse matrix 25e of the first variance-covariance matrix and the inverse matrix 25g of the second variance-covariance matrix are different, an appropriate first The similarity Si of i (i = 1, 2) can be calculated.

  The distance may be a distance using vector quantization. The music search unit 26 performs vector quantization on the feature amount vector of the desired content obtained by the operation of step S24d of the collective feature amount calculation unit 24 to calculate a predetermined number of code vectors, and the calculated code vector and the m-th code vector. The minimum distance among the Euclidean distances to the feature quantity vector 23gm is defined as the first distance.

  Similarly, the music search unit 26 calculates a predetermined number of code vectors by vector quantization of the feature amount vector of the undesired content obtained by the operation of step S24f of the collective feature amount calculation unit 24, and the calculated code vector The second distance is the minimum distance among the Euclidean distances between and the m-th feature vector 23gm.

  Here, it is assumed that the number of code vectors for calculating the first distance is equal to the number of code vectors for calculating the second distance, but the present invention is not limited to this. Whether or not the distance using vector quantization is used is unrelated to whether or not a code vector calculation by vector quantization is performed by the feature extraction unit 22. In addition, the number of code vectors for calculating the distance and the number of code vectors calculated by the feature extraction unit 22 are irrelevant.

  Further, it may be a probabilistic distance using a Gaussian mixture model. In other words, the music search unit 26 expresses the feature quantity vector of the desired content obtained by the operation of step S24d of the collective feature quantity calculation unit 24 as a distribution in which a plurality of Gaussian distributions are mixed. The reciprocal of the occurrence probability of the m-th feature vector 23gm in the mixed Gaussian distribution or the reciprocal of the logarithmic value of the occurrence probability is defined as the first distance.

  Similarly, the feature vector of undesired content obtained by the operation of step S24f of the collective feature calculator 24 is expressed as a distribution in which a plurality of Gaussian distributions are mixed. Then, the reciprocal of the occurrence probability of the m-th feature vector 23gm in the mixed Gaussian distribution or the reciprocal of the logarithmic value of the occurrence probability is set as the second distance.

  Further, it may be a probabilistic distance using a histogram. That is, the music search unit 26 creates a histogram for each element of the feature amount vector of the desired content obtained by the operation of step S24d of the collective feature amount calculation unit 24, and the m-th feature amount vector 23gm in the histogram. The reciprocal of the occurrence probability or the reciprocal of the logarithmic value of the occurrence probability is set as the first distance.

  Similarly, a histogram is created for each element of an undesired content feature vector obtained by the operation of step S24f of the collective feature calculator 24, and the reciprocal of the occurrence probability of the m-th feature vector 23gm in the histogram, Alternatively, the reciprocal of the logarithm of the occurrence probability is set as the second distance.

  When using stochastic distances using these Gaussian mixture models and stochastic distances using histograms, the distance is the reciprocal of the occurrence probability or the reciprocal of the logarithm of the occurrence probability. is not. The distance may be a non-increasing function with respect to the occurrence probability or a logarithmic value of the occurrence probability.

  According to the method using the distance using the vector quantization, the stochastic distance using the Gaussian mixture model, and the stochastic distance using the histogram, the centroid explained in steps S24e and S24g of the collective feature quantity calculation unit 24. It is not necessary to calculate the vector and the inverse matrix of the variance-covariance matrix. Instead of the calculation, the collective feature quantity calculation unit 24 may calculate a code vector, a distribution in which a plurality of Gaussian distributions are mixed, or a histogram for each element corresponding to each method.

  During the operation of the collective feature amount calculation unit 24, information for identifying the desired content 23a is input in step S24b of the flowchart shown in FIG. 6, and information for identifying the undesired content 23a is input in step S24c. In this case, there is a possibility that the inverse matrix 25e of the first variance-covariance matrix and the inverse matrix 25g of the second variance-covariance matrix are greatly different, and the distance between the vector such as the Mahalanobis distance and the group is used. Great effect.

  As described above, when the first similarity S1 and the second similarity S2 are calculated with respect to the M vectors xm, the music search unit 26 allows the content 23a to be desired by the user. A certain similarity R is calculated. The similarity R increases as the first similarity S1 increases, and decreases as the second similarity S2 increases.

  For example, the music search unit 26 calculates the number of vectors Cnt in which the first similarity S1 is larger than the value obtained by adding the predetermined constant T1 to the second similarity S2 (0 ≦ Cnt ≦ M). Here, T1 is a predetermined constant, for example, 0. If the Cnt is large, the similarity R of the feature vector 23g of the content 23a read in step S26c is large, and if the Cnt is small, the similarity R is calculated to be small (step S26f), and step S26c. The operation proceeds to the sequential reading of the feature vector 23g of the content 23a.

Here, Cnt and the similarity R are calculated by, for example, the following (Equation 4-1) and (Equation 4-2), respectively.

  The higher the similarity R is, the higher the possibility that the content 23a read in step S26c is the user's desired content 23a according to the dictionary 25a read in step S26b.

  On the other hand, when the feature vector 23g of the content 23a is not completely read in step S26d, the music search unit 26 orders the information for identifying the content 23a according to the similarity R calculated in step S26f, that is, sorts. To do. And the information which identifies the content 23a whose similarity R is the predetermined | prescribed range is displayed on the display part 12 (step S26g), and operation | movement is complete | finished (step S26h).

  In step S26g, the content 23a for displaying the identification information on the display unit 12 has a high similarity R, for example. Further, the similarity R is small. The similarity R is a value within a predetermined range. Here, when a thing with a large similarity R is displayed, the user is suitable for listening to a music piece that is usually desired, that is, a favorite piece. Also, when the one with a low similarity R is displayed, it is suitable for the user to listen to music that is not usually desired, that is, with little reproduction. If the similarity R is displayed within a predetermined value range, the user is suitable for listening to music that is not usually desired but is not greatly different from the desired music.

  Note that the music search unit 26 calculates the distance between each of the first to Mth feature quantity vectors 23g1 to 23gM constituting the feature quantity vector 23g and the first group in step S26e, and further to the first group. When calculating the distance between each of the Mth feature quantity vectors 23g1 to 23gM and the second group, elements less than N of the first to Mth feature quantity vectors 23g1 to 23gM that are N-dimensional are used. It may be calculated.

  For example, when the distance is calculated using a vector obtained by removing the i-th element (1 ≦ i ≦ N), the first centroid vector 25d and the vector obtained by removing the i-th element of the second centroid vector 25f are used. The matrix obtained by removing the i-th row and i-th column elements of the inverse matrix 25e of the variance-covariance matrix and the inverse matrix 25g of the second variance-covariance matrix is used.

  The element to be removed is to remove the i-th element when the values of the i-th row and the i-th column of the inverse matrix 25e of the first variance-covariance matrix and / or the inverse matrix 25g of the second variance-covariance matrix are small. . Thus, by removing elements that have a large variance, that is, a small contribution to the distance to the first group and / or the distance to the second group, the influence on the similarity R is reduced and the calculation is performed. The amount can be reduced. The effect of this reduction is significant when a large amount of content 23a is stored in the content storage unit 23.

  Here, with reference to FIG. 9, the concept that the music search unit 26 searches the desired music by ordering the contents 23a according to a desired degree with reference to the dictionary 25a will be described. Note that this description is a conceptual description and does not necessarily correspond to a calculated value.

  As shown in FIG. 9, the characteristics of the music are “very dark music”, “dark music”, “slightly dark music”, “light and dark ambiguous music”, “slightly bright music”, “bright music”, and “very bright music”. Suppose that it is represented by an ordered 7-stage consisting of "song" and the middle of two adjacent stages. The music is divided into “relatively dark music” and “relatively bright music” depending on which of the average music whose characteristics are “bright and dark ambiguous music”.

  The user's desired music is a “slightly bright music”, and the desired music that has been listened to is given to the collective feature quantity calculation unit 24, whereby the dictionary 25a shows a first group of desired music. The center-of-gravity vector 25d indicates “slightly bright music”, and the second center-of-gravity vector 25f indicating an undesired group of music indicates “slightly dark music”.

  Therefore, when the music search unit 26 searches for the desired music, the music search unit 26 is close to “slightly bright music” indicated by the first centroid vector 25d and “slightly dark” indicated by the second centroid vector 25f. The music having the characteristics far from the “music” is searched, and the first search result shown in FIG. 9 is obtained. In other words, the first search result is not the second search result indicating a song in a symmetric range in both light and dark directions from a desired “slightly bright song”. Therefore, a search result intended by the user can be obtained without including a “relatively dark song” that is not desired by the user.

  Note that a second search result including a “relatively dark song” that shows a song in a symmetric range in both light and dark directions from a desired “slightly bright song” and includes a “relatively dark song” that deviates from the user's search intention is the second center of gravity. The search result when the concept of the vector 25f, that is, the concept of the music not desired by the user is not used is shown.

  Next, with reference to FIG. 10, in the information processing apparatus according to the present embodiment, the flow of content 23a (indicated by a double line arrow in the figure) and the flow of feature amount (indicated by a solid line arrow in the figure). ) And the flow of specifying the content 23a (indicated by broken arrows in the figure) will be outlined. This description is an outline of the flow of information and does not show all the detailed operations.

  The content is received by the music registration unit 21, and the music registration unit 21 stores the received content in the content 23 a as the content data 23 c and further transmits it to the feature extraction unit 22. The feature extraction unit 22 extracts a feature vector from the received content data 23c and stores it in the content 23a as a feature vector 23g.

  When the designation of the desired content 23a input from the input device 13 is received, the collective feature amount calculation unit 24 refers to the feature amount vector 23g of the content 23a and features necessary for the search of the desired content 23a. A quantity vector is searched and / or created, and a feature quantity such as a centroid vector is stored in the dictionary 25a. The music search unit 26 refers to the feature amount stored in the dictionary 25 a, searches for the desired content 23 a from the content 23 a, and transmits the searched content data 23 c to the music playback unit 27. The music reproducing unit 27 reproduces the received content data 23c and generates sound from the speaker 27a.

  In the above description, the content data 23c included in the content 23a is encoded data. However, the present invention is not limited to this. It may be decoded data.

  In the above description, the concept that the music search unit 26 searches for the desired music with reference to the dictionary 25a has been described as an example. However, the search is not limited to music search. For example, when the content storage unit 23 stores content 23a that is a song and content 23a that is composed of a human voice, either one is a desired content 23a and the other is an undesired content 23a. The amount calculation unit 24 may create the dictionary 25a. When the dictionary 25a is created in this way, the music search unit 26 can search for either the content 23a that is a music or the content 23a that is composed of a human voice.

  Here, it is known that the distribution of frequency components included in the content 23a that is a song and the content 23a that is composed of a human voice are different, and the characteristics (centroid vector and variance covariance) of the groups of the respective content 23a are known. Since the inverse matrix of the matrix is different, the above search is possible.

  In the above description, the music search unit 26 orders the content 23a stored in the content storage unit 23 according to a desired level. However, the present invention is not limited to this. The music search unit 26 may calculate the similarity R of the content data received by the communication unit with reference to the dictionary 25a. When the similarity R is equal to or higher than the first predetermined value and / or equal to or lower than the second predetermined value, the music reproducing unit 27 is controlled to output a predetermined sound from the speaker 27a. You may notify.

  In this case, the communication unit may receive the broadcast, and the music search unit 26 may calculate the similarity R of the content data that is the broadcasted music. Further, the similarity R may be calculated from data at the beginning of the received music, for example, for several seconds. When the calculation is performed based on the beginning data, M feature quantity vectors may be calculated from the content data, or the similarity R may be calculated by calculating less than M feature quantity vectors.

  Further, the notification is not limited to outputting predetermined sound from the speaker 27a. For example, the apparatus may have a vibrator (not shown) and notify by vibration of the vibrator. According to this notification operation, for example, when a broadcast of a favorite song of the user of the apparatus is started when a radio broadcast is received, the user is notified that the favorite song is being broadcast. Can do.

  In the above description, the dictionary 25a is created by the collective feature amount calculation unit 24, but is not limited thereto. For example, it may be received from outside the apparatus. It may also be included when the device is shipped.

  In the above description, the dictionary 25a is stored in the dictionary storage unit 25, but is not limited thereto. It may be created by the collective feature amount calculation unit 24 and transmitted to the music search unit 26.

(Second Embodiment)
The second embodiment is different from the first embodiment in the operation of the collective feature quantity calculation unit 24. Therefore, the operation of the collective feature amount calculation unit 24 of the second embodiment will be described with reference to the drawings. FIG. 11 shows a flowchart of the operation of the collective feature quantity calculation unit 24 according to the second embodiment. Note that the same operation steps as those of the collective feature quantity calculation unit 24 according to the first embodiment are denoted by the same reference numerals, and description thereof is omitted.

  In the operation of step S24d of the collective feature value calculating unit 24 according to the first embodiment, the collective feature value calculating unit 24 obtains the feature value vector 23g of the desired content 23a, and in addition, obtains the feature value of the undesired content 23a. An inverse vector of the vector 23g is used.

  The collective feature quantity calculation unit 24 according to the second embodiment temporarily creates a dictionary 25a by using the inverse vector of the feature quantity vector 23g of the content 23a that is not desired by the collective feature quantity calculation unit 24 according to the first embodiment. Then, the music search unit 26 is controlled so that the first distance between the first centroid vector 25d included in the temporary dictionary 25a and the group represented by the inverse matrix 25e of the first variance-covariance matrix is the largest. The content 23a having the smallest first degree of similarity S1 is searched.

  In addition to the above conditions, a condition that the first distance is not more than a predetermined value, that is, the first similarity S1 is not less than a predetermined value may be added or used instead. good. According to this, since there is a possibility that the content 23a cannot be obtained by the search, it is possible to prevent an inappropriate desired content 23a from being searched.

  Then, the collective feature quantity calculating unit 24 according to the first embodiment uses the inverse vector of the feature quantity vector 23g of the content 23a that is not desired, and the collective feature quantity calculating unit 24 according to the second embodiment Then, the dictionary 25a is created using the searched feature amount vector 23g of the content 23a as the feature amount vector 23g of the desired content 23a (step S24m).

  Similarly, instead of the operation of step S24f of the collective feature value calculating unit 24 according to the first embodiment, the collective feature value calculating unit 24 according to the second embodiment may include the feature vector 23g of the desired content 23a. A dictionary 25a is temporarily created using the inverse vector. Then, the music search unit 26 is controlled so that the second similarity S2 between the group indicated by the second centroid vector 25f included in the temporary dictionary 25a and the inverse matrix 25g of the second variance-covariance matrix. The content 23a having the largest value is searched, and the dictionary 25a is created using the feature vector 23g of the searched content 23a as the feature vector 23g of the undesired content 23a (step S24n).

  According to these processes, all of the feature quantity vectors 23g used in the subsequent processes are stored in the content 23a, and fictitious ones are not included. Therefore, the first centroid vector 25d, the inverse matrix 25e of the first variance-covariance matrix, the second centroid vector 25f, and the inverse matrix 25g of the second variance-covariance matrix calculated thereafter are stored. The content 23a depends more appropriately.

  Note that in step S24m instead of the above processing, a feature vector 23g having a small distance and a reverse vector of the feature vector 23g of the undesired content 23a is searched from the content 23a, and the searched feature vector 23g is used as the desired content. You may use as the feature-value vector 23g of 23a.

  In step S24n, the feature vector 23g having a small distance from the inverse vector of the feature vector 23g of the desired content 23a is searched from the content 23a, and the searched feature vector 23g is the feature vector 23g of the undesired content 23a. It may be used as

  According to these processes, the load for creating the temporary dictionary 25a is reduced. Note that the calculation of the distance is the calculation of the distance between the M vectors and the M vectors, and this distance is as described above. In these cases, a condition that the distance is not more than a predetermined value may be added or used instead. According to this, since there is a possibility that the content 23a cannot be obtained by the search, it is possible to prevent an inappropriate and undesired content 23a from being searched.

  Note that the collective feature amount calculation unit 24 according to the second embodiment needs to perform the processes related to the desired content 23a described in step S24b, step S24m, and step S24e in this order. Further, the processing related to the undesired content 23a described in step S24c, step S24n, and step S24g needs to be performed in this order. However, as described in the explanation of the operation of the collective feature quantity calculation unit 24 according to the first embodiment, which of these processes may be performed first and which may be performed later.

(Other embodiments)
In each of the above embodiments, the content data 23c of the content 23a is music data. However, the present invention is not limited to this. It may be data of all kinds of voices such as voices spoken by humans, mechanical sounds, and voices generated in nature, such as voices used as background sounds, and naturally operates in the same manner.

  Further, when the content data 23c is a moving image, the same processing can be performed by treating the temporal change in the luminance and color difference of a specific pixel in the same manner as the temporal change in audio. Further, the same processing can be performed by taking out a frequency component by performing discrete cosine transform on a macroblock having an arbitrary size. The macroblock may be the entire image. Further, when the content data 23c is a still image, the same processing can be performed by treating the luminance of a specific linear pixel and the change of the color difference on the line in the same manner as the temporal change.

  It does not matter whether the information processing apparatus of the present invention is a fixed apparatus or a portable apparatus. Furthermore, the present invention can naturally be applied to devices that store various contents, such as a moving image playback device mounted on a hard disk, a video camera, a video playback device, a personal computer, a portable music playback device, and a mobile communication terminal device. Moreover, you may combine suitably the element demonstrated in said each embodiment. The present invention is not limited to the above configuration, and various modifications are possible.

The block diagram which shows the structure of the information processing apparatus which concerns on embodiment of this invention. The figure which shows an example of a structure of the content which concerns on embodiment of this invention. The figure which shows an example of a structure of the dictionary which concerns on embodiment of this invention. The flowchart which shows operation | movement of the feature extraction part which concerns on embodiment of this invention. The flowchart which shows the feature-value calculation operation regarding the cepstrum coefficient of the feature extraction part which concerns on embodiment of this invention. 5 is a flowchart showing the operation of a collective feature amount calculation unit according to the first embodiment of the present invention. The flowchart which shows the flow of the feature-value by operation | movement of the collective feature-value calculation part which concerns on embodiment of this invention. The flowchart which shows operation | movement of the music search part which concerns on embodiment of this invention (Operation | movement which searches the content of a predetermined desired level with reference to a dictionary). The figure which shows the concept of the range of the music searched by the music search part which concerns on embodiment of this invention. The flowchart which shows the flow of the content etc. by operation | movement of the information processing apparatus which concerns on embodiment of this invention. The flowchart which shows operation | movement of the collective feature-value calculation part which concerns on the 2nd Embodiment of this invention.

Explanation of symbols

21 music registration unit 22 feature extraction unit 23 content storage unit 23a content 23b content identifier 23c content data 23g feature quantity vector 23g1 first feature quantity vector 23g2 second feature quantity vector 23gi i-th feature quantity vector 23gm m-th feature Quantity vector 23gM Mth feature quantity vector 24 collective feature quantity calculation section 25 dictionary storage section 25a dictionary 25b dictionary identifier 25c dictionary name 25d first centroid vector 25e first inverse covariance matrix inverse matrix 25f second centroid vector 25g Inverse matrix of second variance-covariance matrix 26 Music search unit

Claims (1)

Content storage means for storing content;
Searching for the first type of content from the content storage means, calculating a feature quantity of the first group of the first type of content from a feature quantity vector of the first type of content obtained by the search; A feature having the shortest distance from a vector symmetrical to the feature vector of the first type of content obtained by the search centered on a representative vector of content feature vectors and / or a feature whose distance is smaller than a predetermined value A second type of content having a quantity vector is searched from the content storage means, and a feature quantity of the second group consisting of the second type of content is calculated from the searched feature quantity vector of the second type content. A group feature quantity calculating means;
A first distance between the feature quantity vector of the content stored in the content storage means and the feature quantity of the first group calculated by the group feature quantity calculation means, the feature quantity vector of the content, and the group A second distance between the second group and the feature quantity of the second group calculated by the feature quantity calculation means, and the content of the first type is the content of which the first distance is smaller than the second distance. And / or content search means for searching for the content having the first distance larger than the second distance as the second type of content.

Images