Multi-feature speech/music discrimination system
AssigneeInterval Research Corporation
InventorsEric D. Scheirer, Malcolm Slaney
Life of this patent
Estimating the term…
Abstract
A speech/music discriminator employs data from multiple features of an audio signal as input to a classifier. Some of the feature data is determined from individual frames of the audio signal, and other input data is based upon variations of a feature over several frames, to distinguish the changes in voiced and unvoiced components of speech from the more constant characteristics of music. Several different types of classifiers for labeling test points on the basis of the feature data are disclosed. A preferred set of classifiers is based upon variations of a nearest-neighbor approach, including a K-d tree spatial partitioning technique.
Claims25
A method for discriminating between speech and music content in an audio signal, comprising the steps of:
- selecting a set of audio signal samples;
- measuring values for a plurality of features in each sample of said set of samples;
- defining a multi-dimensional feature space containing data points which respectively correspond to the measured feature values for each sample, and labelling each data point as relating to speech or music;
- measuring feature values for a test sample of an audio signal and determining a corresponding data point in said feature space;
- determining the label for at least one data point in said feature space which is close to the data point corresponding to said test sample;
- and
- classifying the test sample in accordance with the determined label.
A method for determining whether an audio signal contains music content, comprising the steps of:
- dividing the audio signal into a plurality of frequency bands;
- determining modulation frequencies of the audio signal in each band;
- identifying the amount of correspondence of the modulation frequencies among the frequency bands;
- and
- classifying whether audio signal has musical content in dependence upon the identified amount of correspondence;
- wherein the step of determining the modulation frequencies in a frequency band comprises the steps of:
- determining an energy envelope of the frequency band;
- identifying peaks in the energy envelope;
- and
- calculating a windowed autocorrelation of the peaks.
A method for determining whether an audio signal contains music content, comprising the steps of:
- dividing the audio signal into a plurality of frequency bands;
- determining modulation frequencies of the audio signal in each band;
- identifying the amount of correspondence of the modulation frequencies among the frequency bands;
- and
- classifying whether audio signal has musical content in dependence upon the identified amount of correspondence;
- wherein the step of identifying the amount of correspondence of the modulation frequencies comprises the steps of:
- determining peaks in the modulation frequencies for each band;
- selecting a first pair of frequency bands;
- counting the number of modulation frequency peaks which are common to both bands in the selected pair;
- and
- repeating said counting step for all possible pairs of frequency bands.
A method for discriminating between speech and music content in audio signals that are divided into successive frames, comprising the steps of:
- selecting a set of audio signal samples;
- measuring values of a feature for individual frames in said samples;
- determining the variance of the measured feature values over a series of frames in said samples;
- defining a multi-dimensional feature space having at least one dimension which pertains to the variance of feature values;
- defining a decision boundary between speech and music in said feature space;
- measuring a feature value for a test sample of an audio signal and a variance of a feature value, and determining a corresponding data point in said feature space;
- and
- classifying the test sample in accordance with the location of said corresponding point relative to said decision boundary.
A method for detecting speech content in an audio signal, comprising the steps of:
- selecting a set of audio signal samples;
- measuring values for a plurality of features in samples of said set of samples;
- defining a multi-dimensional feature space containing data points which respectively correspond to the measured feature values for each sample, and labelling whether each data point relates to speech;
- measuring feature values for a test sample of an audio signal and determining a corresponding data point in said feature space;
- determining the label for at least one data point in said feature space which is close to the data point corresponding to said test sample;
- and
- indicating whether the test sample is speech in accordance with the determined label.
A method for detecting music content in an audio signal, comprising the steps of:
- selecting a set of audio signal samples;
- measuring values for a plurality of features in samples of said set of samples;
- defining a multi-dimensional feature space containing data points which respectively correspond to the measured feature values for each sample, and labelling whether each data point relates to music;
- measuring feature values for a test sample of an audio signal and determining a corresponding data point in said feature space;
- determining the label for at least one data point in said feature space which is close to the data point corresponding to said test sample;
- and
- indicating whether the test sample is music in accordance with the determined label.
Dependencies are inferred from the claim text.