Process and apparatus for eliminating loudspeaker interference from microphone signals

DETAILED DESCRIPTION It is the object underlying the present invention to effectively and efficiently eliminate loudspeaker interference from the microphone signal used for voice input.
This object is attained by the measures set forth in claim 1, in particular by a computation unit in which a Fourier transform unit determines the spectra X(f,T) and Z(f,T) of the microphone signal and the loudspeaker signal, respectively, and processes these to compute the transfer function H(f,T) in multiplying, smoothing and division layers as well as a convolution layer, with the signal propagation delay effective between the loudspeaker and the microphone--as determined in the delay unit--taken into account.
These measures act to effectively eliminate loudspeaker interference from the microphone signal used for voice input, obviating a push-to-talk key.
If the loudspeaker signal is generated by a system controlled by voice input, such as a voice-controlled television receiver, the inventive system ensures that voice signals the television set itself generates will not be processed as voice control commands.
The range of applications in which the present invention is suitable is much broader, however.
For this reason, reference will generally be had in the explanations that follow to the elimination of loudspeaker interference from a microphone signal.
The present invention presents a novel method in which a network-like computation unit is used to determine the transfer function of the loudspeaker-space system in a direct manner.
The computation effort the method requires is quite modest, and the method is extremely resistant to noise disturbances and duplex situations as, unlike all other methods, it does not assume as a starting point the special case of other non-stationary signals--apart from the loudspeaker signal--being absent from the microphone.
These measures are convergent without problems for very long filters too, and the filter length is limited only by available memory.
The system implements readily on any hardware platform; it is fully adaptive and does not have to be pre-configured for expected signal propagation delays