Speech and Music Discrimination using Gaussian Mixture Model _________________________ Seminar Program Project Team Dr. Deep Sen CHOI Arthur, Tsz Kin Derek, Ka Chun (Supervisor) (3015809) CHENG (3015631) _________________________ Speech and Music Discrimination using GMM _________________________ Speech and Music Discrimination using GMM Motivations • Many researches on HMM, not too many using GMM • GMM reduce complexity compared to HMM • Our feature extraction methods will reduce complexity • Multimedia files search/storage still under develop • Fit University requirement _________________________ Speech and Music Discrimination using GMM _________________________ Speech and Music Discrimination using GMM Applications • Audio Database Indexing • Automatic Bandwidth Allocation • Broadcast Browsing • Intelligent Signal Processing • Intelligent Audio Coding • Audio file Compression • Audio Clip Editing _________________________ Speech and Music Discrimination using GMM Approaches Deterministic Signals can be analysis as completely specified functions of time Un-deterministic Signals must analysis probilistically [Tele3013 notes] _________________________ Speech and Music Discrimination using GMM Procedures 1. Read a signal 2. Segmented it into small frames 3. Extract features of each frames 4. Classify each frames _________________________ Speech and Music Discrimination using GMM Feature Extractions _________________________ Speech and Music Discrimination using GMM Classification _________________________ Speech and Music Discrimination using GMM music speech silence speech _________________________ Speech and Music Discrimination by using GMM Segmentation Reasons • Get a better estimation result • Achieve a Real-Time behavior Music Signal Problems and solutions • Frames too big -- Classification accuracy decrease • Frames too small -- Feature extraction accuracy decrease • Chose frame size ~20ms _________________________ Speech and Music Discrimination using GMM _________________________ Speech and Music Discrimination using GMM 4 Hz modulation energy Speech energy has a characteristic energy modulation peak around the 4Hz syllabic rate. [Houtgast & Steeneken 1985] Reasons • Accurately separate speech signals and music signals (~94%) • Easy to implement in Matlab • Novel and Robust _________________________ Speech and Music Discrimination using GMM Music Signal Speech Signal _________________________ Speech and Music Discrimination using GMM _________________________ Speech and Music Discrimination using GMM Music Signal Speech Signal Energy vs. Time _________________________ Speech and Music Discrimination using GMM Zero-Crossing Count (ZCC) The zero-crossing count is the total number of times that a signal goes through the x-axis over a certain time. Speech signals High ZCC Music signals Low ZCC Reasons • ZCC of a speech signal is significantly high • Very easy to implement in Matlab • Mature and Robust _________________________ Speech and Music Discrimination using GMM _________________________ Speech and Music Discrimination using GMM _________________________ Speech and Music Discrimination using GMM Spectral Roll-off Point The spectral roll-off point measures the “skewness” of the spectrum. Reasons • Music usually has more energy in the high frequency range • Useful for separate different kind of speech later _________________________ Speech and Music Discrimination using GMM Spectral Roll-off Point Spectral Roll-off Point = SR where, _________________________ Speech and Music Discrimination using GMM power Music Signal frequency power Speech Signal frequency _________________________ Speech and Music Discrimination using GMM Entropy Modulation Music appears to be “ordered” compared with a speech signal [J.Pinquier, J.L. Rouas, R. Andre-Obercht 2002] Higher Entropy means higher “ordered” Higher Dynamism means higher rate of changes Reasons • Accurately separate speech signals and music signals(~90%) • Novel and Robust _________________________ Speech and Music Discrimination using GMM Music Signal Speech Signal _________________________ Speech and Music Discrimination using GMM [J. Ajmera, I.A. McCowan, H.Bourlard 2002] _________________________ Speech and Music Discrimination using GMM Instantaneous entropy Average entropy Average Instantaneous entropy _________________________ Speech and Music Discrimination using GMM Pulse Metric The beat of a piece of music is one of the clearest features of the music. [K.D. Martin, E.D.Scheirer, B.L. Vercoe 1988] _________________________ Speech and Music Discrimination using GMM Other Features • Spectral Centroid • Spectral Flux • Silence Ratio • Short-Time Energy Ratio • Volume Dynamic Change • Number of Segments • Segment Duration • …etc _________________________ Introduction to Gaussian Mixture Model (GMM) • Differentiation of speech and music from a sound source • Use for speech processing, mostly for speech recognition, speaker identification and voice conversion • Model densities and to represent general spectral features Why we choose GMM? Low complexity Rate independence Bit scalability Short computation time What is Gaussian Mixture Model? Gaussian Mixture Model consist of a set of local Gaussian modes, and an integrating network. Different Gaussian distributions represent different domain of feature space, and have different output characteristics GMM try to describe a complex system using combination of all the Gaussian clusters, instead of using a single model Gaussian mixtures or clusters Use to describe a complex system instead of using a single model Represents a dataset by a set of mean and covariance Gaussian Mixture Model A Gaussian Mixture Model is represented by: M f (x, ) iN (x, i, i ) i 1 is the P-dimensional input vector i is the mixture weights N (x, i, i ) is the component densities Clustering ‘clustering’ is a technique from pattern classification A technique to group samples P-dimensional feature vector is considered as a point in space and all points ‘near’ if are clustered together clustering Grey circle represents the variance of distribution Gaussian component density P-variate Gaussian function of the form: N (x, i i ) i i 1 1 12 T ( i ) exp( ( x ) i (x i )) p i 2 (2 ) 2 is the mean vector is the covariance matrix Covariance matrix Indicates the dispersion of distribution In mathematics, it is defined as the matrix whose ij th element ij is the covariance of and i x ij ji xi i xj j i,j=1…d xj Covariance matrix The diagonal components of the covariance matrix are the variances of individual random variables Off-diagonal components are the covariance of two random variables, j and i Symmetric matrix x x Full covariance matrix The most powerful Gaussian model as it fits the data best drawback! Needs a lot of data to estimate parameters Costly in high-dimensional feature spaces Diagonal covariance matrix Good compromise between quality and model size Gaussian components can act together to model the overall probability density function Capable of modelling the correlations between the feature vector Review the Gaussian mixture density The matrix weight i must satisfy the condition 1 and i 0 M i i Three components compose the Gaussian mixture density: mean vectors, covariance matrices and mixture weights Expectation-maximization (EM) Estimate the mean vector, covariance matrix and mixture weight Recursively updates distribution of each Gaussian model and conditional probability Idea of Expectation-maximization Instead of starting with a random configuration of all components and improve upon this configuration with expectation-maximization. We start with the optimal one-component mixture. Then start repeating two steps until convergence i) Inset a new components and ii) Apply EM until convergence Convergence Theorem The sequence of likelihood is monotonically-increasing and bounded, the likelihood will converge to a local maximum EM algorithm n log f x Assume denote the loglikelihood of the dataset under k-component matrix fk 1. Compute the optimal one-component mixture f 1. Set k=1 2. Find the optimal new component x; * and corresponding matrix weight * ( Xn , f k ) k i i 1 * , arg max log 1 f x x ; n * while keeping , k i 1 fk fixed i i EM algorithm 3. Set fk 1x 1 fk x x; * and k=k+1 4. Update fk until convergence * * Speech/music discrimination by using GMM An interesting feature of GMM, component densities of mixture may represent… Different phonetic events for modelling speech Different portion of the sound when used to model spectra of sound from musical instrument Achievement Identified optimized frame size Obtained robust features Performed a few tests Implemented some Matlab codes Studied the Gaussian Mixture Models (GMMs) and some of their mathematical expressions Next year planning Comprehensive and more in-depth research on GMMs Model the sound source base on GMMs Evaluate noise effect Matlab implementation for speech/music separation Next year planning Investigate a novel classification method – Support Vector Machine (SVM) Differentiate Male and female speech Differentiate Classical and Non-Classical Music Generate a final thesis report _________________________ Speech and Music Discrimination using GMM _________________________ Speech and Music Discrimination using GMM Resources • Internet, Microsoft Sound Recorder, Matlab • Neural Networks for Pattern Recognition (Bishop 1996) • Processing and Perception of Speech and Music (Morgan 2000) • Research Papers _________________________ Speech and Music Discrimination using GMM Management Plan • Dec – Feb 04 Matlab Implementations Investigate noise effect Research on Support Vector Machine Experiments • Jan 05 Separating class., non-class. music • Feb 05 Separating male, female speech • Mar – Jun 05 Separate Chamber music and Orchestra Music. Separate Baby speech. (if have time) Perception of Speech and Music (2000), John Wiley & Sons, Inc., USA. Thank you Joseph F. Hair, JR., Rolph E. Anderson, Ronald L. Tatham, William C. Black, Multivariate Data Analysis 4th Edition (1995), Prentice-Hall International, Inc. USA. Keinosuke Fukunaga, Computer Science and Scientific Computing: Introduction to Statistical Pattern Recognition 2nd Edition (1990), Academic Press, Inc., California, USA., ISBN 0-12-269851-7 Marty J.Schmidts, Understanding and Using Statistic (1975), D.C Health and Company, Canada. ISBN 0-669-94490-4 Norman L.Johnson, Samuel Kotz, Distributions in statistics: Continuous univariate distributions vol.1 (1970), Houghton Mifflin Company, Boston, USA Richard A. Johnson, Dean W. Wichern, Applied Multivariate Statistical Analysis (1992), Prentice-Hall, Inc., New Jersey, USA. ISBN 0-13-041400-X Richard J.Harris, A Primer of Multivariate Statistics (1975), Academic Press Inc., New York, USA. ISBN 0-12-327250-5 Thomas D. Rossing, The Science of Sound (1982), Addison-Wesley Publishing Company Inc., USA., ISBN 0-201-06505-3 Thomas D. Rossing, Neville H. Fletcher, Principles of Vibration and Sound (1995),