Auditory source separation and its application to automatic music transcription

Abstract

<P>HUMAN BEINGS ARE CAPABLE OF LISTENING TO SIMULTANEOUS SOUND SOURCES WITH ONE EAR. AUDITORY SOURCE SEPARATION IS A COMPUTATIONAL APPROACH TO SIMULATE THIS ABILITY BY EXTRACTING MULTIPLE SOURCES FROM A SINGLE CHANNEL MIXTURE. THIS THESIS PRESENTS NOVEL TECHNIQUES FOR MODELING THE AUDITORY SOURCE SEPARATION IN A STATISTICAL FRAMEWORK. THE SOURCE SIGNALS ARE MODELED AS MARKOV PROCESS, WHICH IS CHARACTERIZED BY MEANS OF LOCAL CONSTRAINTS IN BOTH TIME DOMAIN AND FREQUENCY DOMAIN. TYPICAL AUDIO SIGNALS LIKE SPEECH AND HARMONIC SOUNDS ARE DISCUSSED. SPEECH SIGNALS ARE MODELED BY LINEAR PREDICTIVE CODING WITH SMOOTH CONSTRAINT ON THE COEFFICIENTS. HARMONIC SOUNDS ARE MODELED BY THE NORMALIZED AVERAGE SUMMATION FUNCTION. HIGHER-ORDER STATISTICS ARE MODELED BY INDEPENDENT COMPONENT ANALYSIS. TO REDUCE THE NUMBER OF PARAMETERS TO BE OPTIMIZED, SINUSOIDAL MODELING, AUDITORY FRONT-END AND INDEPENDENT COMPONENT ANALYSIS ARE EMPLOYED FOR DIMENSION REDUCTION. </P><P>THE HARMONIC SOURCE SEPARATION AL