David Burshtein

Signal enhancement using beamforming and nonstationarity with applications to speech

We consider a sensor array located in an enclosure, where arbitrary transfer functions (TFs) relate the source signal and the sensors. The array is used for enhancing a signal contaminated by interference. Constrained minimum power adaptive beamforming, which has been suggested by Frost (1972) and, in particular, the generalized sidelobe canceler (GSC) version, which has been developed by Griffiths and Jim (1982), are the most widely used beamforming techniques. These methods rely on the assumption that the received signals are simple delayed versions of the source signal. The good interference suppression attained under this assumption is severely impaired in complicated acoustic environments, where arbitrary TFs may be encountered. In this paper, we consider the arbitrary TF case. We propose a GSC solution, which is adapted to the general TF case. We derive a suboptimal algorithm that can be implemented by estimating the TFs ratios, instead of estimating the TFs. The TF ratios are estimated by exploiting the nonstationarity characteristics of the desired signal. The algorithm is applied to the problem of speech enhancement in a reverberating room. The discussion is supported by an experimental study using speech and noise signals recorded in an actual room acoustics environment.

Design and analysis of nonbinary LDPC codes for arbitrary discrete-memoryless channels

We present an analysis under the iterative decoding of coset low-density parity-check (LDPC) codes over GF(q), designed for use over arbitrary discrete-memoryless channels (particularly nonbinary and asymmetric channels). We use a random- coset analysis to produce an effect that is similar to output symmetry with binary channels. We show that the random selection of the nonzero elements of the GF(q) parity-check matrix induces a permutation-invariance property on the densities of the decoder messages, which simplifies their analysis and approximation. We generalize several properties, including symmetry and stability from the analysis of binary LDPC codes. We show that under a Gaussian approximation, the entire q-1-dimensional distribution of the vector messages is described by a single scalar parameter (like the distributions of binary LDPC messages). We apply this property to develop extrinsic information transfer (EXIT) charts for our codes. We use appropriately designed signal constellations to obtain substantial shaping gains. Simulation results indicate that our codes outperform multilevel codes at short block lengths. We also present simulation results for the additive white Gaussian noise (AWGN) channel, including results within 0.56 dB of the unrestricted Shannon limit (i.e., not restricted to any signal constellation) at a spectral efficiency of 6 bits/s/Hz.

Iterative and sequential Kalman filter-based speech enhancement algorithms

Speech quality and intelligibility might significantly deteriorate in the presence of background noise, especially when the speech signal is subject to subsequent processing. In particular, speech coders and automatic speech recognition (ASR) systems that were designed or trained to act on clean speech signals might be rendered useless in the presence of background noise. Speech enhancement algorithms have therefore attracted a great deal of interest. In this paper, we present a class of Kalman filter-based algorithms with some extensions, modifications, and improvements of previous work. The first algorithm employs the estimate-maximize (EM) method to iteratively estimate the spectral parameters of the speech and noise parameters. The enhanced speech signal is obtained as a byproduct of the parameter estimation algorithm. The second algorithm is a sequential, computationally efficient, gradient descent algorithm. We discuss various topics concerning the practical implementation of these algorithms. Extensive experimental study using real speech and noise signals is provided to compare these algorithms with alternative speech enhancement algorithms, and to compare the performance of the iterative and sequential algorithms.

Bounds on the maximum likelihood decoding error probability of low density parity check codes

We derive both upper and lower bounds on the decoding error probability of maximum-likelihood (ML) decoded low-density parity-check (LDPC) codes. The results hold for any binary-input symmetric-output channel. Our results indicate that for various appropriately chosen ensembles of LDPC codes, reliable communication is possible up to channel capacity. However, the ensemble averaged decoding error probability decreases polynomially, and not exponentially. The lower and upper bounds coincide asymptotically, thus showing the tightness of the bounds. However, for ensembles with suitably chosen parameters, the error probability of almost all codes is exponentially decreasing, with an error exponent that can be set arbitrarily close to the standard random coding exponent.

Asymptotic enumeration methods for analyzing LDPC codes

We show how asymptotic estimates of powers of polynomials with nonnegative coefficients can be used in the analysis of low-density parity-check (LDPC) codes. In particular, we show how these estimates can be used to derive the asymptotic distance spectrum of both regular and irregular LDPC code ensembles. We then consider the binary erasure channel (BEC). Using these estimates we derive lower bounds on the error exponent, under iterative decoding, of LDPC codes used over the BEC. Both regular and irregular code structures are considered. These bounds are compared to the corresponding bounds when optimal (maximum-likelihood (ML)) decoding is applied.

On the application of a ldpc codes to arbitrary discrete-memoryless channels

We discuss three structures of modified low-density parity-check (LDPC) code ensembles designed for transmission over arbitrary discrete memoryless channels. The first structure is based on the well-known binary LDPC codes following constructions proposed by Gallager and McEliece, the second is based on LDPC codes of arbitrary (q-ary) alphabets employing modulo-q addition, as presented by Gallager, and the third is based on LDPC codes defined over the field GF(q). All structures are obtained by applying a quantization mapping on a coset LDPC ensemble. We present tools for the analysis of nonbinary codes and show that all configurations, under maximum-likelihood (ML) decoding, are capable of reliable communication at rates arbitrarily close to the capacity of any discrete memoryless channel. We discuss practical iterative decoding of our structures and present simulation results for the additive white Gaussian noise (AWGN) channel confirming the effectiveness of the codes.

Speech enhancement using a mixture-maximum model

We present a spectral domain, speech enhancement algorithm. The new algorithm is based on a mixture model for the short time spectrum of the clean speech signal, and on a maximum assumption in the production of the noisy speech spectrum. In the past this model was used in the context of noise robust speech recognition. In this paper we show that this model is also effective for improving the quality of speech signals corrupted by additive noise. The computational requirements of the algorithm can be significantly reduced, essentially without paying performance penalties, by incorporating a dual codebook scheme with tied variances. Experiments, using recorded speech signals and actual noise sources, show that in spite of its low computational requirements, the algorithm shows improved performance compared to alternative speech enhancement algorithms.

Support Vector Machine Training for Improved Hidden Markov Modeling

We present a discriminative training algorithm, that uses support vector machines (SVMs), to improve the classification of discrete and continuous output probability hidden Markov models (HMMs). The algorithm uses a set of maximum-likelihood (ML) trained HMM models as a baseline system, and an SVM training scheme to rescore the results of the baseline HMMs. It turns out that the rescoring model can be represented as an unnormalized HMM. We describe two algorithms for training the unnormalized HMM models for both the discrete and continuous cases. One of the algorithms results in a single set of unnormalized HMMs that can be used in the standard recognition procedure (the Viterbi recognizer), as if they were plain HMMs. We use a toy problem and an isolated noisy digit recognition task to compare our new method to standard ML training. Our experiments show that SVM rescoring of hidden Markov models typically reduces the error rate significantly compared to standard ML training.

Expander graph arguments for message-passing algorithms

We show how expander-based arguments may be used to prove that message-passing algorithms can correct a linear number of erroneous messages. The implication of this result is that when the block length is sufficiently large, once a message-passing algorithm has corrected a sufficiently large fraction of the errors, it will eventually correct all errors. This result is then combined with known results on the ability of message-passing algorithms to reduce the number of errors to an arbitrarily small fraction for relatively high transmission rates. The results hold for various message-passing algorithms, including Gallagers hard-decision and soft-decision (with clipping) decoding algorithms. Our results assume low-density parity-check (LDPC) codes based on an irregular bipartite graph.

Efficient maximum-likelihood decoding of LDPC codes over the binary erasure channel

We propose an efficient maximum-likelihood (ML) decoding algorithm for decoding low-density parity-check (LDPC) codes over the binary-erasure channel (BEC). We also analyze the computational complexity of the proposed algorithm.