Speech separation by means of stationary least-squares harmonic estimation

It is shown that stationary least-squares harmonic analysis and sinusoidal synthesis of speech can provide an efficient modeling tool for the problem of speech separation in voiced regions. For an input signal-to-signal ratio of -9 dB, the procedure yields an output signal-to-noise ratio ranging from 4.7 to 10.8 dB, which represents a gain of 13.7 to 19.8 dB. The interfering speech is not perceptible in the output, the only audible degradation being of tonal character. As in other pitch-based systems, the problem of pitch estimation is still a major limitation of the model. Experimental results confirm the improvement in speech separation over that previously reported in the literature