Pitch Tracking MUMT 611 Philippe Zaborowski February 2005
Pitch Tracking MUMT 611 Philippe Zaborowski February 2005
Pitch Tracking � Goal is to track the fundamental � Vast area of research mostly focused on voice coding � Dozens of different algorithms � All algorithms have limitations � None are ideal
Technical Difficulties: Piano
Technical Difficulties: E. Bass
Algorithm Classification � Time Domain � Spectral Domain � Combined Time/Spectral Domain � Neural Networks
Time Domain � � Common Features: � Analysis performed on sample basis instead of buffered intervals � No transformation needed � Cheap on computation Common Drawbacks: � Not suited for signals where the fundamental is weak and the harmonics are strong � DC offset can be a problem
Time Domain � Threshold Crossing (zero crossing)
Time Domain � Dolansky (1954)
Time Domain � Rabiner and Gold (1969)
Time Domain � Autocorrelation (Rabiner 1977)
Time Domain � Average Magnitude Difference Function (Ross 1974)
Time Domain � Cooper and Ng (1994)
Time/Spectral Domain � Least-Square (Choi 1995) � Combines the reliability of frequency-domain with high resolution of time-domain � Able to analyze shorter signal segments � Suitable for real-time � Uses constant Q tranform
Spectral Domain � � Common Features: � Transformation from time to spectral domain is computationally intensive � Superior control and analysis of formants Common Drawbacks: � Simple study of spectrum not enough � DFT based algorithms use equally spaced bins
Spectral Domain � FFT with different harmonic analysis: � Maximum of FFT (Division Method) � Piszczalski and Galler (1979) � Harmonic Product (Schroeder 1968)
Spectral Domain � Constant Q transform (Brown and Puckette 1992)
Spectral Domain � Cepstrum (Andrews 1990)
Conclusion � � Spectral Domain: � Give good results � Require a demanding analysis of spectrum Time Domain: � Generally inferior to spectral domain � Some have comparable results with less computation
- Slides: 18