The author presents a self-adapting noise reduction system which is based on a four-microphone array combined with an adaptive postfiltering scheme. Noise reduction is achieved by utilizing the directivity gain of the array and by reducing the residual noise through postfiltering of the received microphone signals. The postfiltering scheme depends on a Wiener filter estimating the desired speech signal and is computed from short-term measurements of the autocorrelation and cross-correlation functions of the microphone signals. The noise reduction system has been tested experimentally in a typical office room. The system produces an enhanced speech signal with barely noticeable residual noise if the input SNR is greater than 0 dB. The received noise power-measured in the absence of the speech signal-can be reduced by 28 dB.>

This paper Presents a learning procedure for speaker-dependent word recognition systems which are based on the principle of dynamic time warping. The reference templates are created by averaging word tokens for each class. The averaging, procedure, which is based on a purely sequential processing of the tokens, contains additional weighting operations for word boundaries and scaling of the time axis. These operations improve the robustness of the learning procedure. The new learning procedure has been tested with different speech examples, some of which have been recorded in extremely noisy conditions with casual speakers. In all cases the learning procedure yields very reliable reference templates.

Recognition of connected word strings can be performed by segmenting the word string automatically into single-word components which are then classified by a single-word recognition system. We propose and investigate a segmentation procedure which is based completely on statistical principles. An estimation algorithm, adapted to the statistical data of the signal parameters, determines the word boundaries. This procedure, which offers several advantages over other methods, has been tested with connected digits. The results show that an estimation algorithm based on quadratic polynomials yields sufficiently accurate segmentation. Recognition results for 2-to 4-digit strings are presented in this paper.

This paper investigates the optimum entropy versus distortion performance of quantizers optimized for uniform, Gaussian, Laplacian, and gamma-distributed memoryless sources which are useful models of the quantizer input signals in speech or picture coding schemes. We list the maximally obtainable signal-to-quantization noise ratios for one-dimensional optimum (i.e. entropy-coded) quantizers in the important low bit-rate region. These results have been obtained by an iterative solution of a set of nonlinear equations. Additionally we have also computed the corresponding rate-distortion functions by employing the Blahut-algorithm. These latter results upperbound the performances of multi-dimensional quantization schemes, and a comparison with the former results indicates the penalty to be paid for restricting a coder to perform a one-dimensionai quantization. It will be shown that the differences can be significant in the low bit-rate region.