Institute for Systems Research

In this dissertation, we propose an invariance principle fornon-parametric speech representations in acoustical environments.We stipulate that the topology of the codevectors in a vector quantization (VQ) codebookas defined in terms of class posterior distributionswill be preserved in a certain information-theoretic sense,and make this invariance principle our basis in deriving normalizationalgorithms that correct for the acoustical mismatch betweenenvironments.

We develop topology preserving algorithms in two frameworks, constrained distortionminimization (VQ with a topology preservation constraint) andinformation geometry (alternating minimization with a topology preservation constraint) and show their equivalence.Finally, we report results on the Wall Street Journal data,the Spoken Speed Dial corpus and the TI Cellular Corpus.

The algorithm is shown to improve performancesignificantly in all three tasks, most notably in the more difficult problemof cellular hands free microphone speech wherethe technique decreases theword error for continuous ten digit recognition from 23.8% to 13.6% and the speaker dependent voice callingsentence error from 16.5% to 10.6%.

We present a computational auditory model that exhibits spectro-temporal MTFs consistent with the salient trends in the data.The model is used to demonstrate the potential relevance of these MTFsto the assessment of speech intelligibility in noise and reverberantconditions.

First, several fundamental issues regarding the statistical properties of fading channels are provided. We demonstrate the constraints, that must be satisfied so that the channel can be regarded as impaired by ﲦlat (i.e., non-frequency-selective) fading with a constant fading factor over each symbol duration. Throughout this thesis these constraints are assumed to be satisfied.

We next investigate the channel capacity and cutoff rates for fading channels with DPSK-modulated input signals and perfect symbol interleaving. The impact of the channel state information (CSI), on these information-theoretic limits is also discussed. We introduce several symbol metrics for soft-decision decoding, and their performance is investigated by analytical derivation as well as by simulation. Furthermore, we define a bit metric for DQPSK modulation, and compare this bit metric to dibit (symbol) metrics.

We then consider the problem of error concealment for mobile radio, communications with a maximum-likelihood soft- decision decoder. A normalized codeword reliability is defined as the decision reliability information when CSI is not available. We employ a given, interpolation algorithm on a particular land mobile radio system and design a rule for selecting unreliable codewords. Simulation results show that error concealment can decrease the minimum operational signal-noise ratio (SNR) by 3 dB or more.

Finally, we address the problem of exploiting the residual redundancy in the source to enhance the channel decoder's performance -i.e., maximum a posteriori (MAP) decoding. We use two simple source models to demonstrate that MAP decoding can achieve significant gain over maximum-likelihood decoding. We employ a practical CELP-based land mobile radio system to show that significant residual redundancy does exist in the output of some source encoders. Simulation results show that a 2 - 3 dB gain can be achieved by MAP decoding (over ML) at low SNR.

We begin by quantifying the amount of ﲲesidual redundancy inherent in the LSP's of Federal Standard 1016 CELP. This is done by modeling the LSP's as first and second-order Markov chains. Two models for LSP generation are proposed; the first model characterizes the intra-frame correlation exhibited by the LSP's, while the second model captures both intra-frame and inter-frame correlation. By comparing the entropy rates of the models thus constructed with the CELP rates, it is shown that as many as one-third of the LSP bits in every frame of speech are redundant.

We next consider methods by which this residual redundancy can be exploited by an appropriately designed channel decoder. Before transmission, the LSP's are encoded with a forward error control (FEC) code; we consider both block (Reed- Solomon) codes and convolutional codes. Soft-decision decoders that exploit the residual redundancy in the LSP's are implemented assuming additive white Gaussian noise (AWGN) and independent Rayleigh fading environments. Simulation results employing binary phaseshift keying (BPSK) indicate coding gains of 2 to 5 dB over soft-decision decoders that do not exploit the residual redundancy.

In the first part, TCM schemes that provide high coding gains over the flat, slow Rayleigh distributed fading channel are presented. It is shown that the use of two encoders in parallel used to specify the in-phase and quadrature components of the transmitted signal results in large performance improvements in bit error rates when compared with conventional TCM schemes in which a single encoder is used. Using this approach which we label ﲉ-Q TCM codes with bandwidth efficiencies of 1, 2, and 3 bits/sec/Hz are described for various constraint lengths. The performance of these codes is evaluated using tight upper bounds and simulation.

In the second part, the use of space diversity with three different combining schemes is investigated. Expressions for the cutoff rate parameter are shown for the three combining schemes over the fully interleaved Rayleigh-distributed flat fading channel. Also, tight upper bounds on the pairwise error probability are derived for the three combining schemes. Examples of I-Q TCM schemes with diversity combining are shown. The cutoff rate and a tight upper bound on the pairwise error probability are also derived for maximal ration combining with correlated branches.

In the last part the problem of reliably transmitting trellis coded signals over very noisy channels is considered. Sequence maximum a posteriori (MAP) decoding of correlated signals transmitted over very noisy AWGN and Rayleigh channels is presented. A variety of different systems with different sources, modulation schemes, encoder rates and complexities are simulated. Sequence MAP decoding proves to substantially improve the performance at very noisy channel conditions especially for systems with moderate redundancies and encoder rates. A practical example for coding the CELP line spectral parameters (LSPs) is also considered. Two source models are used. Coding gains of as much as 4 dB are achieved.

The ripples parameters that characterize cortical cells are distributed somewhat evenly, with the characteristic ripple frequencies ranging from 0.2 to over 2 cycles/octave and the characteristic angular frequency typically ranging from 2 to 20 Hz. Many responses exhibit periodicities not found in the spectral envelope of the stimulus. These periodicities are of two types. Slow rebounds with a period of about 150 ms appear with various strengths in about 30 % of the cells. Fast regular firings, with interspike intervals of the order of 10 ms are much less common and may reflect the ability of certain cells to follow the fine structure of the stimulus.

We begin by quantifying the amount of ﲲesidual redundancy inherent in the LSP's of Federal Standard 1016 CELP. This is done by modeling the LSP's as first-and second-order Markov chains. Two models for LSP generation are proposed; the first model characterizes the intra-frame correlation exhibited by the LSP's, while the second model captures both intra-frame and inter-frame correlation. By comparing the entropy rates of the models thus constructed with the CELP rates, it is shown that as many as one-third of the LSP bits in every frame of speech are redundant.

We next consider methods by which this residual redundancy can be exploited by an appropriately designed channel decoder. Before transmission, the LSP's are encoded with a forward error control (FEC) code; we consider both (Reed-Solomon) codes and convolutional codes. Soft-decision decoders that exploit the residual redundancy in the LSP's are implemented assuming additive white Gaussian noise (AWGN) and independent Rayleigh fading environments. Simulation results employing binary phase-shifting keying (BPSK) indicate coding gains of 2 to 5 dB over soft-decision decoders that do not exploit the residual redundancy.