Armein Z. R. Langi

Multifractal processing of speech signals

This paper describes a novel framework in processing speech signals using multifractality concepts, and shows that multifractality could be a new fundamental tool for speech processing. New approaches using the self-similarity model of complexity have been developed to simplify processing of complicated speech signals. The approaches associate signals with fractal sets and then characterize the sets using real numbers called fractal dimensions. This paper extends the approaches by associating speech signals with measures and then characterizing the measures with multifractal curves such as Mandelbrot dimensions (denoted as f(/spl alpha/)) or Renyi dimensions D/sub q/. Such curves (also known as singularity curves) are the characterization of speech signals. This paper shows that multifractal (or singularity) processing of speech is capable of providing important processing aspects: decomposition, representation, and spectrum characterization-a capability that makes Fourier processing of signals is fundamental. The multifractal approach can decompose speech into various segments based on variations of the segments Holder exponents. The results can be used for new speech segmentation schemes. The multifractal approach is also used in characterizing speech through singularity spectrum. This can be used to develop better accuracy speech recognition schemes. Finally, the paper describes a process to reconstruct speech signals from their singularity representation, by the means of the so-called wavelet maxima.

A DSP implementation of a voice transcoder for VoIP gateways

The objective of this research is to design and implement high-quality speech compression in real-time on a single-chip transcoder system. Based on voice over Internet protocol (VoIP) requirements, we have decided to implement multipulse maximum, likelihood quantization (MP-MLQ) and algebraic code excited linear prediction (ACELP) (used in an ITU-T G.723.1 standard) on a DSP processor TMS320C5402. The goal is to implement a high quality speech coding (with signal-to-noise ratio, SNR of more than 10 dB), at a low bit rate of 8 kbit/s or less. The coder must have a delay not more than 100 ms. Furthermore the resulting system must be shown to fit within a single chip. Using an ITU-T reference code, we use a series of iterative code optimization efforts. This rapid development approach is successful in achieving the requirements in a three man-month effort. The program requires 58-73 MIPS computation and 39 Kword memory, thus fits within the single chip of a 100 MIPS TMS320C5409. Total delay time is 59.4 ms. The bit rates are as low as 5.3 kbit/s and 6.3 kbit/s, with SNRs of 11.52 dB and 12.72 dB, respectively.

Compression of aerial ortho images based on image denoising

Abstract only given. Discusses the compression of an important class of computer images, called aerial ortho images, that result from geodetic transformation computations [Kinsner, 1994]. The computations introduce numerical noise, making the images nearly incompressible losslessly because of their high entropy. The use of classical lossy compression schemes is also not desirable because their effects on the original image are unknown. We then propose the use of image denoising coupled with lossless image compression, that preserves selected image characteristics. Two denoising schemes for a compression ratio of 2:1 are compared. The first scheme is based on a Donohos (1992) wavelet shrinking scheme which preserves image smoothness. We study the effect of various shrinking parameter values on the compression ratio and image quality, where 35.5 dB peak signal-to-noise ratio (PSNR) is obtained for a compression ratio of 2.03:1. This approach preserves high-frequency information, so that sharp edges do not become blurred as in classical filtering methods. This is critically important, because the main feature of ortho images is in its flatness and its precision of edge position. The second scheme is based on preserving pixel predictability [Kostelich and Schreiber, 1993), leading to a variant of planar predictive coding. This approach adds, to the edge preserving capability, the limitation in pixel deviation between the original and denoised images to be within one grayscale level. As a result, two different predictive coding schemes achieve a compression ratio of 2:1 at 49.9 dB and 51.2 dB PSNR.

Rapid development of a real-time speech coder on a TMS320C54x DSP

This paper describes a rapid implementation of a real-time voice compression (codec) on a DSP processor for voice over Internet Protocol (VoIP) gateways and applications. Voice codec algorithms such as G.723.1 are complex, hence they take many manhours to implement in real-time. By identifying control flow and dataflow, and applying iterative refinement of a C reference source code, we are able to obtain real-time implementation on a 100 MIPS TMS320C5402 digital signal processing (DSP) starter kit (DSK) in three man-months only. At current implementation, the encoder and decoder use 70 MIPS and 6 MIPS, respectively. They occupy 26 Kword program memory and 17 Kword data memory.

Performance comparison of denoising methods for heart sound signal

This paper presents the performance analysis and comparison of three denosing methods for heart sound signal based on wavelet transform (WT), total variation (TV), and empirical mode decomposition (EMD). Extensive simulations are performed using normal and abnormal heart sound data and the performance is evaluated in terms of signal-to-noise ratio (SNR), root mean square error (RMSE), and percent root mean square difference (PRD). The simulation results show that EMD based denosing method outperforms two other methods.

Segmented fractal dimension measurement of 1-D signals: a wavelet based method

Fractal dimension is an important characteristic of signals that contain information about their structure complexity. Although, fractal dimension estimation of geometric objects has shown very good results, it is a big problem to estimate the dimension of a 1-D signal. As the dynamics of the signal may vary over time, then its fractal dimension also varies over time. This paper shows our estimations of fractal dimensions of 1-D signal based on wavelet coefficients which are obtained from the maxima lines of a continuous wavelet transform. Instead of their exact estimated values, we are more interested in analyzing their dimension variation over time. Our measurement shows that a stationary signal has relatively constant dimension, on the other hand, a non-stationary signal has varying dimension. These measurements lead us to acquire a different method of signal characterization.

A wavelet-based measurement of signal fractal dimensions

This paper describes a study of measuring signal fractal dimensions (especially in a form of Lipschitz exponents /spl gamma/) using wavelets. The procedure is as follows. Given a one-dimensional signal f(t) and its corresponding wavelet transform (at a scale a and a position b) as W/sub f/(a, b), we find wavelet maxima lines l(a, b) and their corresponding wavelet maxima |W/sub f/(a, b)|. Suppose the signal f(t) has a Lipschitz exponent /spl gamma/ at t=b/sub 0/, and there is a maxima line l(a,b) reaching b/sub 0/ as a/spl rarr/0. The corresponding wavelet maxima in the line satisfy an inequality |W/sub f/(a, b)|/spl les/Ca/sup /spl gamma/+0.5/ for some constant C and a/spl rarr/0. A log-log plot on the inequality estimates the Lipschitz exponent /spl gamma/. We have performed an experiment of the procedure for f(t)=1-|0.5-t|/sup /spl gamma//, where the Lipschitz component /spl gamma/ varies from 0.1 to 0.9 at a 0.1 interval. The procedure provides relatively good estimates for 0.5/spl les//spl gamma//spl les/0.9, with relative errors less them 10%.

Automatic segmentation and detection of heart sound components S1, S2, S3 and S4

In this paper, we propose an automatic segmentation and detection of heart sound components (S1, S2, S3 and S4) which incorporates Empirical Mode Decomposition (EMD) denoising, autocorrelation-based cardiac cycle calculation, Shannon energy envelope extraction, first derivative peak and boundary detection, and real peak selection using Herons formula. The proposed method is evaluated on synthetic data corrupted by white Gaussian noise. The simulation results show that the proposed method is able to segment and identify the heart sound component correctly from normal and abnormal heart sound data.

Optimizing Voice over Internet Protocol (VoIP) networks based-on Extended E-model

In this paper, we optimize low speed networks access for designing an efficient Voice over Internet Protocol network, especially in 64 kbps, 128 kbps, and 256 kbps bandwidth capacities. We draw some analytical approach based on Extended E-model to quantify some levels of quality of service. Using the model, those levels are determined by network parameters such as delay, jitter, packet loss level, network utilization, and bandwidth capacity and by implementation configuration such as voice coder, packetization scheme, and size of jitter buffer. Our objective is to find maximum number of calls in some given bandwidth capacities while maintaining a certain level of the quality of service. Using numeric estimation, we found that the optimum solution for some given bandwidth capacities can be achieved by applying G.723.1 5.3 kbps voice coder, packet loss level less than 1%, jitter less than 80 ms, and network utilization less than 80%.

A procedure for singularity measurement using wavelet

Two important key factors in signal processing are singularity analysis and dynamical behaviour, as singularities and dynamics carry most of the signal information. Wavelet analysis is very good in localization of singularities. This paper describes a method in measuring singularity of a simple well-known one-dimensional signal using a wavelet approach. The singularity, by mean of a Lipschitz exponent of a function, is measured by taking the slope of a log-log plot of scales and wavelet coefficients along modulus maxima lines of a wavelet transform. Using this method, we measure the dimension of a particular function f(t)=1-|c-t|/sup /spl lambda// where c is a constant and /spl lambda/ varies from 0.1 to 0.9 with a 0.1 interval. This procedure yields good estimation of the Lipschitz exponent when 0.5/spl les//spl lambda//spl les/0.9.