Reducing cocktail party noise by adaptive array filtering

This paper introduces an adaptive filtering method that rejects interfering speech from nonpreferred directions. Phase delays in interfering signals received by three receivers are used to obtain a low signal-to-noise ratio (SNR) reference input for an adaptive filter. Since adaptive filter output SNR is inversely proportional to its reference input SNR, this method works very well theoretically. To minimize the excess steady-state error in conventional schemes, a master-slave structure is introduced. Simulations with digitized speech data, tested both by comparing waveforms and by listening, showed significant improvements. Due to the computational simplicity of the method, real-time realization is feasible.