A NEW ALGORITHM FOR TWO-SPEAKER PITCH TRACKING IN SINGLE CHANNEL PARADIGM

We describe a new method for extracting the pitch contours of two speakers from their mixture. The technique can be effectively applied to the single channel speech separation problem where common blind source separation approaches fail to handle the problem. The algorithm consists of two stages: a sinusoidal spectrogram based track extractor and a joint minimum mean square harmonic modeling pitch estimator. The pitch contours nominees are first determined using the sinusoidal spectrogram. The harmonic pitch estimator is then used to extract the best two pitch contours. We compare the proposed technique with the two state-of-the-art approaches. Experimental results, conducted on a database consisting of mixed signals with the target to jammer ratios (TJR) ranging from -9 to +9 dB, show that our technique significantly outperforms the two other techniques