We have studied the effects of random packet losses in digital speech systems based on 12-bit PCM and 4-bit adaptive DPCM coding. The effects are a function of packet lengthBand probability of packet loss PL. We have also studied tbe benefits of an odd-even sample-interpolation procedure that mitigates these effects (at the cost of increased decoding delay). The procedure is based on arranging a2B-block of codewords into twoB-sample packets, an odd-sample packet and an even-sample packet. If one of these packets is lost, the odd (or even) samples of the2B-block are estimated from the even (or odd) samples by means of adaptive interpolation. Perceptual considerations indicate that packet lengths most robust to losses are in the range 16-32 ms, irrespective of whether interpolation is used or not. With these packet lengths, tolerable PLvalues, which are strictly input-speech-dependent, can be as high as 2 to 5 percent without interpolation and 5 to 10 percent with interpolation. These observations are based on a computer simulation with three sentence-length speech inputs, and on informal listening tests.

Four well-known procedures for analog speech privacy have been compared in terms of residual intelligibility, bandwidth expansion, and encoding delay. Intelligibility scores have been determined from a perceptual experiment where about 70 untrained listeners were given the task of recognizing each of 200 spoken digits that occurred in a balanced set of 50 encrypted four-digit utterances, and by averaging resulting probabilities of correct digit recognition. Bandwidth expansion has been expressed in terms of a new segmental measure that is more sensitive to short-time bandwidth manipulations than a conventional, long-time-averaged power spectrum measurement. Encoding delay is a straightforward function of analog scrambler parameters. The scrambling procedures that have been compared are sample permutation (S), block permutation (B), frequency inversion (F), and a combination of methodsBandF, denoted by [BF]. Sample permutations involved a contiguous set of LS(2 to 128) 8 kHz samples, while block permutations operated on a contiguous set of NB(4 to 128) speech segments each of which was LB(8 to 256) samples long. Frequency inversion is obtained by simply inverting the sign of every other Nyquist (8 kHz) sample. The parameters,L_{s},N_{B}, and LB, determine residual intelligibility as well as transmission properties such as encoding delay and bandwidth. The comparisons in our study provide a quantitative justification for the popular approach [BF]. For example, withN_{B} = 8andL_{B} =128, although the encoding delay is as much as 128 ms, the bandwidth expansion is only about 100 Hz (using the new segmental measure), and the digit intelligibilityIis 20 percent. Note that in the specific problem of recognizing ten digits, purely random (input-independent) listener responses correspond toI = 10percent.

In a simulation of soft decision demodulation-interpolation for digital speech transmission, it has been noted that signal-to-noise ratio gains due to the technique are greater with nonadaptive and periodically adaptive quantizers than with instantaneously adaptive quantizers; further, the gains are greater with PCM than with differential PCM. Our results for nonadaptive PCM reinforce those in a recent analytical paper by Sundberg [1].

The statistical block protection coding scheme [2] for protecting DPCM encoded speech signals through noisy channels has been extended to accommodate DPCM-AQF encoded speech, where AQF stands for adaptive quantization with forward (explicit) transmission of step size. Signal-to-noise ratio (SNR) gains, typically 12 dB, have been achieved over a dynamic range> 20dB for a bit error rate (BER) of 1.4 percent and the SNR improvement is found to increase with BER. Perceptual improvements in decoded speech have a good correspondence with the gains in SNR. The penalty for this substantial enhancement in the performance of the DPCM-AQF system is an increase in transmission bit rate of 3.5 percent and encoding delay of 64 ms.

Consider the delta-modulation (DM) of a first-order Gauss-Markov signalX. Let the adjacent-sample correlation inXbec, and let the (first-order) DM predictor coefficient bea. We express the quantizer input Qrin the form[S_{r} - aE_{r - 1} + (c - a)X_{r - 1}], where Sris an "innovations" term,aE_{r - 1}denotes the effect of quantizationerror(E)feedback and(c - a)X_{r - 1}reflects the effect of using ana \neq c. For the important case ofc \rightarrow 1(which models over-sampled DM inputs), we propose the simplifying assumption [7] of uncorrelatedXandE; with this assumption, our formalization of quantizer input leads very simply to interesting results in linear (LDM) and adaptive delta modulation (ADM). The LDM results are generalizations of known expressions for optimum values ofa, and the step-size Δ, and the value of signal-to-noise ratio SNR. For ADM, we derive optimum multiplier values for step-size adaptations with a one-bit memory, using the case ofc \sim a = 1for simplicity. Our results depend on modeling instantaneous step-size adaptation as a mechanism for tracking the expected magnitude ofQ; existing literature has formalized such adaptation models only for the case of multi-bit quantizers.

A decision theoretic approach to the problem of verifying that a given observation "belongs" to a claimed pattern is examined. The problem presupposes a nonzero probability of impersonations, and verifier actions are ternary in general: a claimed identity is either 1) accepted or 2) rejected, or it is 3) "blanked" when the verifier refuses to take either of the actions 1) and 2). Corresponding costs-of-wrong decision will be typically different in applications like speaker verification. We characterize sample observations by normal distributions and provide explicit results for the analytically simple situation of "efficient" measurements whose interpattern variances are much greater than their intrapattern variances. In our specific model this will imply that the pdf of a measurement under the "impersonator" hypothesis is "diffuse" in relation to the pdf of the same measurement under any of several "true-pattern" hypotheses. We demonstrate that the likelihood ratio function, in a typical verification task, possesses a finite upper bound and that this is related to a certain threshold effect for verifier decisions. We further mention the suitability of a minimax strategy for verification and compare its performance with those of optimal (minimum-cost) and maximum-likelihood procedures for the binary decision (accept-or-reject) problem. We provide numerical examples that illustrate the minimax performance as a function of 1) a parameter which reflects the peculiarity of the claimed pattern, and of 2) the measurement dimensionality.