James F. Kaiser

Energy separation in signal modulations with application to speech analysis

An efficient solution to the fundamental problem of estimating the time-varying amplitude envelope and instantaneous frequency of a real-valued signal that has both an AM and FM structure is provided. Nonlinear combinations of instantaneous signal outputs from the energy operator are used to separate its output energy product into its AM and FM components. The theoretical analysis is done first for continuous-time signals. Then several efficient algorithms are developed and compared for estimating the amplitude envelope and instantaneous frequency of discrete-time AM-FM signals. These energy separation algorithms are used to search for modulations in speech resonances, which are modeled using AM-FM signals to account for time-varying amplitude envelopes and instantaneous frequencies. The experimental results provide evidence that bandpass-filtered speech signals around speech formants contain amplitude and frequency modulations within a pitch period. >

On amplitude and frequency demodulation using energy operators

It is shown that the nonlinear energy-tracking signal operator Psi (x)=(dx/dt)/sup 2/-xd/sup 2/x/dt/sup 2/ and its discrete-time counterpart can estimate the AM and FM modulating signals. Specifically, Psi can approximately estimate the amplitude envelope of AM signals and the instantaneous frequency of FM signals. Bounds are derived for the approximation errors, which are negligible under general realistic conditions. These results, coupled with the simplicity of Psi , establish the usefulness of the energy operator for AM and FM signal demodulation. These ideas are then extended to a more general class of signals that are sine waves with a time-varying amplitude and frequency and thus contain both an AM and an FM component; for such signals it is shown that Psi can approximately track the product of their amplitude envelope and their instantaneous frequency. The theoretical analysis is done for both continuous- and discrete-time signals. >

The use of a masking signal to improve empirical mode decomposition

Empirical mode decomposition (EMD) provides a new method for analyzing signals from a nonlinear viewpoint. EMD is defined by an algorithm requiring experimental investigation instead of rigorous mathematical analysis. We show that EMD yields its own interpretation of combinations of pure tones. We present the problem of mode mixing and give a solution involving a masking signal. The masking signal method also allows EMD to be used to separate components that are similar in frequency that would be inseparable with standard EMD techniques.

METEOR: a constraint-based FIR filter design program

It is proposed to specify a filter only in terms of upper and lower limits on the response, find the shortest filter length which allows these constraints to be met, and then find a filter of that order which is farthest from the upper and lower constraint boundaries in a minimax sense. The simplex algorithm for linear programming is used to find a best linear-phase FIR filter of minimum length, as well as to find the minimum feasible length itself. The simplex algorithm, while much slower than exchange algorithms, also allows the incorporation of more general kinds of constraints, such as concavity constraints (which can be used to achieve very flat magnitude characteristics). Examples are given to illustrate how the proposed and common approaches differ, and how the proposed approach can be used to design filters with flat passbands, filters which meet point constraints, minimum phase filters, and bandpass filters with controlled transition band behavior. >

Design of FIR filters with flatness constraints

We treat the problem of designing lowpass FIR digital filters that are very flat at zero frequency, smooth in the passband, and minimax in the stopband. The cases of lowpass and lowpass-differentiator are studied in particular. An effective design algorithm is described based on linear programming with a single equality constraint on the first derivative at the origin plus a concavity constraint in the passband. An empirical design equation for estimating model order in the odd-length lowpass-differentiator case is given.

Methods for stress classification: nonlinear TEO and linear speech based features

Speech production variations due to perceptually induced stress contribute significantly to reduced speech processing performance. One approach that can improve the robustness of speech processing (e.g., recognition) algorithms against stress is to formulate an objective classification of speaker stress based upon the acoustic speech signal. An overview of methods for stress classification is presented. First, we review traditional pitch-based methods for stress detection and classification. Second, neural network based stress classifiers with cepstral-based features, as well as wavelet-based classification algorithms are considered. The effect of stress on linear speech features is discussed, followed by the application of linear features and the Teager (1990) energy operator (TEO) based nonlinear features for effective stress classification. A new evaluation for stress classification and assessment is presented using a critical band frequency partition based the TEO feature and the combination of several linear features. Results using NATO databases of actual speech under stress are presented. Finally, we discuss issues relating to stress classification across known and unknown speakers and suggest areas for further research.

Reproducing the Cocktail Party Effect

The Cocktail Party Effect allows the binaural listener to concentrate on speech from one talker, suppressing others in his environment. As contrasted to the monaural listener, the binaural auditor gains between 5 and 15 db signal‐to‐noise ratio necessary for 50% intelligibility. A gain of this magnitude cannot be reconciled with linear array theory which prescribes addition of the two available aural inputs. An alternative accounting can be made by postulating signal processing based upon properties of binaural hearing. In this modal the binaural processor derives a temporary signal which is used to gate the aural input. In effect, the gating signal leaves the major portions of the preferred talkers speech envelope intact while suppressing sound from other talkers or background noise when these do not overlap with the preferred speech. A preliminary circuit version of such a processor which derives the gating signal by cross correlation has been built and tested. Subjective measurements in two‐ and three...

On external properties satisfied by the I o -sinh window

The I o -sinh window has been used extensively in signal processing applications. By the choice of one parameter, one can conveniently trade its desirable main lobe width (resolution) and side lobe rejection properties. The extremal property satisfied by the dissipative wave equation Greens function is derived. The I o -sinh window is shown to be a special case of the Greens function for the 1-D dissipative wave operator. As such, it satisfies the energy equality of the dissipative wave equation. As a natural extension, we also consider 2-D and 3-D Greens function solutions as logical choices for higher order windows.

Design of Wide-Band Sampled-Data Filters

The frequency domain of wide-band linear sampleddata filters is considered. The sampled-data filter is termed ?wideband? when the frequency range of useful approximation to its continuous counterpart approaches half the sampling frequency. Sampled-data filter representations for continuous filters can be obtained using several different design procedures.[1] A particular design method utilizing the bilinear transformation is developed. The method is especially useful in designing wide-band sampleddata filters which exhibit relatively flat frequency-magnitude characteristics in successive pass- and stop-bands. Filters of this type are widely used in network simulation and data processing problems.[2] The design method possesses two chief advantages over the standard z transform.[3] The first is that the transformation used is purely algebraic in form. This means it can be applied easily to a continuous filter having a rational transfer characteristic expressed in either polynomial or factored form. The second advantage is the elimination of aliasing[4] errors inherent in the standard z transform. Thus, the sampled-data filter obtained by this design method exhibits the same frequency response characteristics as the continuous filter, except for a nonlinear warping of the frequency scale. Compenation for this warping can be made by a suitable frequency scale modification. Some of the more common filter networks to which the design method can be applied effectively are the Butterworth, Bessel, Chebyshev, and elliptic-filter structures. The design method consists first of obtaining a rational transfer characteristic for a continuous filter that satisfies the design specifications.

Demodulation of AM–FM resonances in speech using energy separation

Motivated by theoretical and experimental evidences [e.g., in H. Teager and S. Teager, Proc. NATO ASI: Speech Production and Speech Modeling, Bonas, France (1989)] that various nonlinear phenomena during speech production cause modulations of the airflow, AM–FM models for speech resonances and a novel efficient algorithm to estimate their parameters were proposed in [P. Maragos, J. Kaiser, and T. Quatieri, IEEE Trans. Signal Process. 41, 3024–3051 (1993)]. The algorithm uses the differential operation Ψ(x)=(x)2−xẍ to detect modulations in speech signals by tracking the physical energy implicit in the particular ‘‘source’’ producing the observed acoustic resonance signal and by separating this energy into its time‐varying amplitude and frequency components. In this paper experimental results are reported on using refinements of this energy separation algorithm to measure modulations in speech resonances. These results indicate that voiced speech signals, bandpass filtered around speech formants, contain s...