Multiple Pitch Tracking and Harmonic Segregation Algorithm for Auditory Scene Analysis
Kazuki Nishi, Mototsugu Abe, Shigeru Ando · Transactions of the Society of Instrument and Control Engineers · 1998
This paper describes a new algorithm of the auditory scene analysis, by which we can automatically choose a single sound stream from many background ones and reconstruct its original waveform. The features of our algorithm are as follows: 1) Efficient use of harmonics structure, i.e., stream is characterized by pitch and its dynamics; therefore it will be segregated by a time-varying comb filter. 2) Analysis and synthesis in the wavelet domain for extracting a stream so that its spectrum is invariant with the pitch estimation error. 3) The use of Parzen's density estimate and non-parametric Kalman filter for the optimum and multimodal tracking of pitch candidates. This algorithm is examined by several simulations and experiments using some artificial and real world data.