Feature extraction apparatus, feature extraction method, and program thereof

A song search system including:

• a song data selection receiving unit which receives the song data selected by the user from among multiple AAC-format song data stored in a song data storage unit;
• a feature vector creating unit which calculates a feature vector for a feature value stored in a feature storage unit;
• a specific feature vector search unit which searches the feature vectors obtained by the feature vector creating unit for a specific feature vector in which the distance between the feature vector thereof and the feature vector of the song data which has been selected and received by the song data selection receiving unit is equal to or smaller than a predetermined value;
• a song search unit which searches multiple AAC-format song data stored in the song data storage unit for the song data in which the feature vector thereof matches the specific feature vector;
• a feature database creating apparatus provided with:
• the song data storage unit which stores multiple song data in the AAC format, and
• the feature storage unit which obtains the feature value of each of multiple AAC-format song data stored in the song data storage unit, and stores the feature values thus obtained;
• a feature database creating apparatus comprising a feature extraction apparatus provided with:
• an MDCT coefficient extraction unit which extracts MDCT coefficients from a AAC-format song data,
• a classification unit which locates the MDCT coefficients thus extracted by the MDCT coefficient extraction unit on Mel frequency regions so as to uniformly classify the MDCT coefficients into classes, the number of which is the same as that of a predetermined number of Mel filter banks,
• an integrating unit which extracts the MDCT coefficients classified by the classification unit by applying a predetermined window function, and integrates the MDCT coefficients thus extracted, in increments of the Mel filter banks, and
• a feature calculation unit which calculates the feature value by performing logarithmic cosine conversion of the integrated results obtained by the integrating unit,
• wherein the classification unit includes:
• a Mel frequency calculation unit which converts the sampling frequency for the AAC-format song data into a Mel scale, thereby obtaining a Mel frequency,
• a Mel frequency dividing unit which divides the Mel frequency into the same number of classes as that of the Mel filter banks, and calculates the minimum value and the maximum value for each of the Mel frequencies thus divided,
• an original frequency calculation unit which converts the minimum value and the maximum value of each Mel frequency calculated by the Mel frequency dividing unit into a linear scale, thereby obtaining the minimum value and the maximum value for each original frequency;
• an upper-limit frequency calculation unit which calculates the upper-limit frequency based upon the sampling frequency for the AAC-format song data,
• an upper-limit frequency dividing unit which uniformly divides the upper-limit frequency calculated by the upper-limit frequency calculation unit into frequency bands, the number of which is the same as a predetermined number of filter banks,
• a center frequency classification unit which classifies the center frequencies of the frequency bands into regions each of which is defined by the minimum value and the maximum value of the original frequency, thereby classifying the center frequencies into the same number of classes as that of the Mel filter banks, and
• an MDCT coefficient classification unit which classifies the MDCT coefficients, which have been extracted by the MDCT coefficient extraction unit and the number of which is the same as that of the filter banks, into the same number of classes as that of the Mel filter banks, by associating each MDCT coefficient with a corresponding one of the center frequencies which have been classified by the center frequency classification unit into the same number of classes as that of the Mel filter banks;
• and
• wherein the integrating unit extracts the MDCT coefficients classified by the MDCT coefficient classification unit by applying the predetermined window function, and integrates the MDCT coefficients thus extracted, in increments of the Mel filter banks.

A feature extraction method for acquiring a feature value for AAC-format song data, including:

• a first step for extracting MDCT coefficients from the AAC-format song data;
• a second step for locating the MDCT coefficients thus extracted in the first step on Mel frequency regions so as to uniformly classify the MDCT coefficients into classes, the number of which is the same as that of a predetermined number of Mel filter banks;
• a third step for extracting the MDCT coefficients classified in the second step by applying a predetermined window function, and integrating the MDCT coefficients thus extracted, in increments of the Mel filter banks;
• and
• a fourth step for calculating the feature value by performing logarithmic cosine conversion of the integrated results obtained in the third step;
• wherein the second step includes:
• a fifth step for converting the sampling frequency for the AAC-format song data into a Mel scale, thereby obtaining a Mel frequency,
• a sixth step for dividing the Mel frequency into the same number of classes as that of the Mel filter banks, and calculating the minimum value and the maximum value for each of the Mel frequencies thus divided,
• a seventh step for converting the minimum value and the maximum value of each Mel frequency calculated in the sixth step into a linear scale, thereby obtaining the minimum value and the maximum value for each original frequency,
• an eighth step for calculating the upper-limit frequency based upon the sampling frequency for the AAC-format song data,
• a ninth step for uniformly dividing the upper-limit frequency calculated in the eighth step into frequency bands, the number of which is the same as a predetermined number of filter banks,
• a tenth step for classifying the center frequencies of the frequency bands into regions each of which is defined by the minimum value and the maximum value of the original frequency, thereby classifying the center frequencies into the same number of classes as that of the Mel filter banks, and
• an eleventh step for classifying the MDCT coefficients, which have been extracted in the first step and the number of which is the same as that of the filter banks, into the same number of classes as that of the Mel filter banks, by associating each MDCT coefficient with a corresponding one of the center frequencies which have been classified in the tenth step into the same number of classes as that of the Mel filter banks;
• and
• wherein, in the third step, the MDCT coefficients classified in the eleventh step are extracted by applying the predetermined window function, and the MDCT coefficients thus extracted are integrated, in increments of the Mel filter banks.

A computer program product which includes a non-transitory computer readable medium having a set of instructions stored thereon which instructs a computer to execute a feature extraction method for acquiring a feature value for AAC-format song data, including:

• a first step for extracting MDCT coefficients from the AAC-format song data;
• a second step for locating the MDCT coefficients thus extracted in the first step on Mel frequency regions so as to uniformly classify the MDCT coefficients into classes, the number of which is the same as that of a predetermined number of Mel filter banks;
• a third step for extracting the MDCT coefficients classified in the second step by applying a predetermined window function, and integrating the MDCT coefficients thus extracted, in increments of the Mel filter banks;
• and
• a fourth step for calculating the feature value by performing logarithmic cosine conversion of the integrated results obtained in the third step;
• wherein the second step includes:
• a fifth step for converting the sampling frequency for the AAC-format song data into a Mel scale, thereby obtaining a Mel frequency,
• a sixth step for dividing the Mel frequency into the same number of classes as that of the Mel filter banks, and calculating the minimum value and the maximum value for each of the Mel frequencies thus divided,
• a seventh step for converting the minimum value and the maximum value of each Mel frequency calculated in the sixth step into a linear scale, thereby obtaining the minimum value and the maximum value for each original frequency,
• an eighth step for calculating the upper-limit frequency based upon the sampling frequency for the AAC-format song data,
• a ninth step for uniformly dividing the upper-limit frequency calculated in the eighth step into frequency bands, the number of which is the same as a predetermined number of filter banks,
• a tenth step for classifying the center frequencies of the frequency bands into regions each of which is defined by the minimum value and the maximum value of the original frequency, thereby classifying the center frequencies into the same number of classes as that of the Mel filter banks, and
• an eleventh step for classifying the MDCT coefficients, which have been extracted in the first step and the number of which is the same as that of the filter banks, into the same number of classes as that of the Mel filter banks, by associating each MDCT coefficient with a corresponding one of the center frequencies which have been classified in the tenth step into the same number of classes as that of the Mel filter banks;
• and
• wherein, in the third step, the MDCT coefficients classified in the eleventh step are extracted by applying the predetermined window function, and the MDCT coefficients thus extracted are integrated, in increments of the Mel filter banks.