Back to results
US 7,091,409 B2Granted
Music feature extraction using wavelet coefficient histograms
AssigneeUniversity of Rochester
InventorsTao Li, Qi Li, Mitsunori Ogihara
Life of this patent
Estimating the term…
Abstract
A music classification technique computes histograms of Daubechies wavelet coefficients at various frequency subbands with various resolutions. The coefficients are then used as an input to a machine learning technique to identify the genre and emotional content of music.
Claims26
A method for automatically forming a feature set describing an electronic signal representing a piece of music, the method comprising:
- (a) receiving the electronic signal into a computing device;
- (b) performing a wavelet decomposition of the electronic signal to obtain a plurality of wavelet coefficients in a plurality of subbands;
- (c) forming a histogram of the wavelet coefficients in each of the subbands;
- (d) calculating an average, variance and skewness of each of the histograms;
- (e) calculating a subband energy of each of the histograms;
- and
- (f) forming the feature set such that the feature set comprises the average, variance, skewness, and subband energy of at least some of the subbands, wherein step (f) comprises forming the feature set such that the feature set comprises the average, variance, skewness, and subband energy of fewer than all of the subbands.
4 more elements
A method for automatically forming a feature set describing an electronic signal representing a piece of music, the method comprising:
- (a) receiving the electronic signal into a computing device;
- (b) performing a wavelet decomposition of the electronic signal to obtain a plurality of wavelet coefficients in a plurality of subbands;
- (c) forming a histogram of the wavelet coefficients in each of the subbands;
- (d) calculating an average, variance and skewness of each of the histograms;
- (e) calculating a subband energy of each of the histograms;
- and
- (f) forming the feature set such that the feature set comprises the average, variance, skewness, and subband energy of at least some of the subbands;
- wherein step (b) comprises convolving at least part of the electronic signal with a Daubechies wavelet filter, and wherein step (b) is performed with less than all of the electronic signal.
5 more elements
A method for automatically forming a feature set describing an electronic signal representing a piece of music, the method comprising:
- (a) receiving the electronic signal into a computing device;
- (b) performing a wavelet decomposition of the electronic signal to obtain a plurality of wavelet coefficients in a plurality of subbands;
- (c) forming a histogram of the wavelet coefficients in each of the subbands;
- (d) calculating an average, variance and skewness of each of the histograms;
- (e) calculating a subband energy of each of the histograms;
- (f) forming the feature set such that the feature set comprises the average, variance, skewness, and subband energy of at least some of the subbands;
- and
- (g) using the feature set to classify the piece of music into at least one of a plurality of categories of music, wherein step (g) is performed using a multi-class classification algorithm, and wherein step (g) is performed using a plurality of binary classification algorithms.
5 more elements
A method for automatically forming a classifier algorithm for classifying a piece of music represented by an electronic signal into one or more of a plurality of musical genres, the method comprising:
- (a) receiving into a computing device a plurality of classified electronic signals, each of the classified electronic signals representing a known piece of music which has already been classified into one or more of the plurality of musical genres;
- (b) for each of the classified electronic signals:
- (i) performing a wavelet decomposition of the classified electronic signal to obtain a plurality of wavelet coefficients in a plurality of subbands;
- (ii) forming a histogram of the wavelet coefficients in each of the subbands;
- (iii) calculating an average, variance and skewness of each of the histograms;
- (iv) calculating a subband energy of each of the histograms;
- and
- (v) forming a feature set such that the feature set comprises the average, variance, skewness, and subband energy of at least some of the subbands;
- and
- (c) automatically forming the classifier algorithm from the feature sets such that the classifier algorithm properly classifies the known pieces of music.
7 more elements
A device for automatically forming a feature set describing an electronic signal representing a piece of music, the device comprising:
- an input for receiving the electronic signal;
- and
- a computing device, in communication with the input, for:
- performing a wavelet decomposition of the electronic signal to obtain a plurality of wavelet coefficients in a plurality of subbands;
- forming a histogram of the wavelet coefficients in each of the subbands;
- calculating an average, variance and skewness of each of the histograms;
- calculating a subband energy of each of the histograms;
- and
- forming the feature set such that the feature set comprises the average, variance, skewness, and subband energy of at least some of the subbands, wherein the computing device forms the feature set such that the feature set comprises the average, variance, skewness, and subband energy of fewer than all of the subbands.
6 more elements
A device for automatically forming a feature set describing an electronic signal representing a piece of music, the device comprising:
- an input for receiving the electronic signal;
- and
- a computing device, in communication with the input, for:
- performing a wavelet decomposition of the electronic signal to obtain a plurality of wavelet coefficients in a plurality of subbands;
- forming a histogram of the wavelet coefficients in each of the subbands;
- calculating an average, variance and skewness of each of the histograms;
- calculating a subband energy of each of the histograms;
- and
- forming the feature set such that the feature set comprises the average, variance, skewness, and subband energy of at least some of the subbands;
- wherein the computing device performs the wavelet decomposition by convolving at least part of the electronic signal with a Daubechies wavelet filter, and wherein the wavelet decomposition is performed with less than all of the electronic signal.
7 more elements
A device for automatically forming a feature set describing an electronic signal representing a piece of music, the device comprising:
- an input for receiving the electronic signal;
- and
- a computing device, in communication with the input, for:
- performing a wavelet decomposition of the electronic signal to obtain a plurality of wavelet coefficients in a plurality of subbands;
- forming a histogram of the wavelet coefficients in each of the subbands;
- calculating an average, variance and skewness of each of the histograms;
- calculating a subband energy of each of the histograms;
- forming the feature set such that the feature set comprises the average, variance, skewness, and subband energy of at least some of the subbands;
- and
- using the feature set to classify the piece of music into at least one of a plurality of categories of music, wherein the computing device classifies the piece of music using a multi-class classification algorithm, and wherein the computing device classifies the piece of music using a plurality of binary classification algorithms.
7 more elements
A device for automatically forming a classifier algorithm for classifying a piece of music represented by an electronic signal into one or more of a plurality of musical genres, the device comprising;
- an input for receiving a plurality of classified electronic signals, each of the classified electronic signals representing a known piece of music which has already been classified into one or more of the plurality of musical genres;
- and
- a computing device, in communication with the input for forming the classifier algorithm by:
- for each of the classified electronic signals;
- (i) performing a wavelet decomposition of the classified electronic signal to obtain a plurality of wavelet coefficients in a plurality of subbands;
- (ii) forming a histogram of the wavelet coefficients in each of the subbands;
- (iii) calculating an average, variance and skewness of each of the histograms;
- (iv) calculating a subband energy of each of the histograms;
- and
- (v) forming a feature set such that the feature set comprises the average, variance, skewness, and subband energy of at least some of the subbands;
- and
- automatically forming the classifier algorithm from the feature sets such that the classifier algorithm properly classifies the known pieces of music.
9 more elements
Dependencies are inferred from the claim text.