A "Cocktail-Party" Processor: Noise Reduction for Multichannel Speech Signals
Hans Werner Strube, Manfred R. Schroeder, Hans Werner Strube, Manfred R. Schroeder, B.F. Rice · 1978
Exploiting the differences between the two ear signals, the human listener can increase the intelligibility for a speaker in a noisy environment. Here we present three signal processing methods that simulate this abiiity for two situations. In the first situation, the noise is uncorrelated between the different channels. The multi- channel signals are treated either by a time-varying Wiener filter (realized by an FFT method) or by a time-vaying Kalman filter in the time do- main. Both filters are based on short-time estimates of the statistical characteristics of the signals. In the second situation, a few speakers, standing in different directions, are recorded by a dummy head. The undesired speakers are suppressed by adaptive noise cancell- ing, again using an FFT technique and short-time power estimates. Initially, the system has to be adapted to the desired direction. In both situations, for SNR's around -3 dB, typical noise reduction values are Abstract-It is possible to coherently cancel the effect of a wideband traveling-wave field at a designated point in space by adaptively filtering numerous spatially-separated,' electrically-transduced samples of the field and applying the filter outputs as a correction signal. The effect of the field may be canceled electronically by subtracting the correc- tion signal from the electrical output of a transducer located at the designated point. This technique may be used to extract a ''message signal which is embedded in the interference and has its source proximate to the designated point. An interference field comprising a finite number of wideband statistically independent plane-wave random processes, each having zero mean, and each incident from a different direction, is treated here. Governing equations for the optimum Wiener filter transfer functions which effect the desired can- cellation are derived in the domains of continuous and discrete time. In the discrete case, the solution is obtained using the method of least squares (after Widrow). Filter design criteria are given. Magnitude and rate of interference reduction at the designated point are shown for specific simulated cases.