Optimal and Adaptive Subband Beamforming Principles and Applications

This part discusses signal processing methods for speech extraction in use with voice communication applications such as personal digital assistants (PDA:s), mobile telephone terminals and personal computers. The speaker will be distant from the device and thus the speech signal entering the device will be subject to reverberation as well as disturbed by background noise. Further more, the comprehension is reduced due to the acoustic coupling between the loudspeaker and the microphone/microphones. This acoustic coupling will result in an echo effect at the far end of the communication link. The proposed structure consists of a multi sensor array which allows for techniques of directional processing. The channels in the array are sampled by analog-to-digital converters and followed by digital finite impulse response filters. The construction of these filters in order to achieve speech extraction by means of directional processing in the near-field are considered in this study. The background noise and the acoustic coupling are aimed to be simultaneously reduced while the speech originating from the location of interest is left essentially undistorted by the filtering operation. Optimal broadband array optimization techniques are examined in the array near-field and a new modified adaptive frequency domain RLS approach is proposed in order to track variations in the surrounding noise environment. An efficient polyphase filter-bank is used for frequency transformation and reconstruction. Aliasing and magnitude distortion effects in the filter-bank transformations are theoretically evaluated and an efficient over-sampled filter-bank is proposed in order to circumvent such problems. Different optimal fixed beamforming methods are considered; an optimal array gain optimization procedure, a theoretical diffuse noise field model for a point source and a least squares solution. The proposed adaptive frequency domain optimization algorithm is compared to the optimal beamformers and simulations on real speech signal recordings in a car hands-free mode show that the adaptive algorithm performance is close to the optimal optimization methods, while the computational complexity is substantially reduced.