Davor Petrinovic

Causal Cubic Splines: Formulations, Interpolation Properties and Implementations

The paper presents two formulations of causal cubic splines with equidistant knots. Both are based on a causal direct B-spline filter with parallel or cascade implementation. In either implementation, the causal part of the impulse response is realized with an efficient infinite-impulse-response (IIR) structure, while only the anticausal part is approximated with a finite-order finite-impulse-response (FIR) filter. Resulting cubic coefficients are computed from the causal B-spline coefficients by using a third-order output FIR filter with either single-input multiple-output (SIMO) or multiple-input multiple-output (MIMO) structure, depending on the chosen formulation of the cubic spline. The paper demonstrates and proves that the properties of the resulting causal splines are quite different, whether they are based on a more popular B-spline formulation, or a bit neglected tridiagonal matrix formulation. It is shown that the proposed low-complexity but accurate causal interpolators can be realized for many practical applications with the delay of only a few samples.

High efficiency multiplexing scheme for multi-channel A/D conversion

Multiplexing technique is frequently used in multichannel analog to digital (A/D) conversion, thus reducing the system complexity and cost. However, if the sampling frequencies of the input channels are not equal, obtaining high A/D converter utilization is rather complicated. A new rule-based synchronous multiplexing scheme for multi-channel A/D conversion with different channel sampling frequencies is proposed. It is optimal in the sense that the required A/D conversion frequency is a minimum. The channel sequencing is based on the earliest deadline principle, similar to the one used in real-time dynamic scheduling algorithms.

Convolution on the

In this paper, we derive an explicit form of the convolution theorem for functions on an n -sphere. Our motivation comes from the design of a probability density estimator for n -dimensional random vectors. We propose a probability density function (pdf) estimation method that uses the derived convolution result on Sn. Random samples are mapped onto the n -sphere and estimation is performed in the new domain by convolving the samples with the smoothing kernel density. The convolution is carried out in the spectral domain. Samples are mapped between the n-sphere and the n-dimensional Euclidean space by the generalized stereographic projection. We apply the proposed model to several synthetic and real-world data sets and discuss the results.

n

Paper presents development of the acoustic model for Croatian language for automatic speech recognition (ASR). Continuous speech recognition is performed by means of the Hidden Markov Models (HMM) implemented in the HMM Toolkit (HTK). In order to adjust the HTK to the native language a novel algorithm for Croatian language transcription (CLT) has been developed. It is based on phonetic assimilation rules that are applied within uttered words. Phonetic questions for state tying of different triphone models have also been developed. The automated system for training and evaluation of acoustic models has been developed and integrated with the new graphical user interface (GUI). Targeted applications of this ASR system are stress inoculation training (SIT) and virtual reality exposure therapy (VRET). Adaptability of the model to a closed set of speakers is important for such applications and this paper investigates the applicability of the HTK tool for typical scenarios. Robustness of the tool to a new language was tested in matched conditions by a parallel training of an English model that was used as a baseline. Ten native Croatian speakers participated in experiments. Encouraging results were achieved and reported with the developed model for Croatian language.

-Sphere With Application to PDF Modeling

The paper presents a new method for all-pole model estimation based on minimization of the weighted mean square error in the sampled spectral domain. Due to discrete nature of the proposed distance measure, emphasis can be put on an arbitrary set of spectral samples what can greatly improve the model accuracy for periodic signals. Weighting can also be applied to improve the fitting in certain spectral regions according to any desired fidelity criterion. Iterative algorithm for determination of the optimal model is proposed and an exceptionally fast convergence rate is demonstrated. Accuracy of the estimation algorithm is verified on an example of a synthetic vowel for a broad range of pitch frequencies.

Development of Acoustic Model for Croatian Language Using HTK

A new, computationally efficient technique for calculation of the line spectrum frequencies (LSF) that can be applied to any order of the LPC analysis is proposed. It is based on the quotient-difference (Q-D) root-finding algorithm that enables simultaneous solution for all the LSFs. It is an iterative procedure that offers the tradeoff between accuracy and complexity, what is especially important for the real-lime applications. To improve the convergence, a nonlinear mapping of the LSFs is also proposed for low accuracy applications, the method is even more effective then the fast converging Newton-Rapshon method, but is at the same time exceptionally simple, has a very regular structure and requires only basic mathematical operations.

Discrete weighted mean square all-pole modeling

We propose a method for direct digital frequency synthesis (DDS) using a cubic spline piecewise-polynomial model for a phase-to-sinusoid amplitude converter (PSAC). This method offers maximum smoothness of the output signal. Closed-form expressions for the cubic polynomial coefficients are derived in the spectral domain and the performance analysis of the model is given in the time and frequency domains. We derive the closed-form performance bounds of such DDS using conventional metrics: rms and maximum absolute errors (MAE) and maximum spurious free dynamic range (SFDR) measured in the discrete time domain. The main advantages of the proposed PSAC are its simplicity, analytical tractability, and inherent numerical stability for high table resolutions. Detailed guidelines for a fixed-point implementation are given, based on the algebraic analysis of all quantization effects. The results are verified on 81 PSAC configurations with the output resolutions from 5 to 41 bits by using a bit-exact simulation. The VHDL implementation of a high-accuracy DDS based on the proposed PSAC with 28-bit input phase word and 32-bit output value achieves SFDR of its digital output signal between 180 and 207 dB, with a signal-to-noise ratio of 192 dB. Its implementation requires only one 18 kB block RAM and three 18-bit embedded multipliers in a typical field-programmable gate array (FPGA) device.

Calculation of the line spectrum frequencies using the quotient-difference scheme

An approach for reducing the complexity of the switched-adaptive interframe vector prediction (SIVP) that is used for coding speech spectrum envelopes is proposed in this paper. To facilitate the search through the set of switched predictors used for prediction of the input LSF (line spectral frequency) vector, the predictors are organized in a binary tree structure. For a conventional full-searched SIVP coder with N=2/sup b/ predictors, predictions must be performed by all of them in order to determine the best one, while only 2b predictions are sufficient for the proposed binary tree-searched coder. A design procedure for obtaining optimal binary tree-structured predictors is given. The effectiveness of the proposed coder is evaluated and the results are compared to the baseline full-searched coders as a function of the number of predictors and the resolution of the vector quantizers used for quantization of the prediction residual. A discussion of possible applications of the proposed coder is also given.

Spline-based high-accuracy piecewise-polynomial phase-to-sinusoid amplitude converters

This paper presents an extensive statistical analysis of the acoustic startle response of two vocal parameters: fundamental frequency (F0) and root-mean-square energy (E), as well as of the orbicularis oculi (eyeblink) surface electromyography (sEMG). An experiment was conducted in which fourteen participants were exposed to acoustic startle stimuli of varying parameters, i.e., intensity level, duration, rise time, and spectral type, during periods of sustained phonation. Voice recordings of the phonations were taken alongside several physiological signals, of which only the sEMG was analyzed in this paper. Response features (peak value, peak time, latency, rise time, fall time, and duration) were extracted on F0, E and sEMG data, and statistical analysis was conducted using linear mixed effects models to show the response behavior with the varying stimuli. The results for vocal F0 and E data were congruent with sEMG data and earlier work in the field. The results demonstrated that vocal analysis can be used as a feasible alternative to the sEMG eyeblink analysis of acoustic startle responses.

Switched-adaptive interframe vector prediction with binary-tree searched predictors

This paper presents a generic algorithm for the automatic selection of sound synthesis parameters based on input verbal descriptions. The user can specify desired sound characteristics using adjectives and the system will choose parameters which produce a corresponding sound. The unique feature of the algorithm is that it does not rely on prior knowledge of the underlying sound synthesizer so it can be applied to any sound synthesis technique. The first step of the algorithm is the quantification of input adjectives achieved by an expert system based on fuzzy logic. The second step is the selection of a synthesis parameter using a genetic algorithm with the intent of reaching the quantified goal generated by the first step. The results of a subjective evaluation carried out among a group of musicians suggest that the survey participants were generally satisfied with the generated sounds and that the algorithm significantly outperforms random selection of sounds.