Accelerating Matlab DSP Code on the GPU
Seth Benton spent a few days testing Jacket to accelerate MATLAB on NVIDIA GPUs, and found it surprisingly easy to speed up DSP code. He ran 2D FFT and interp2 benchmarks on a MacBook Air with a GeForce 9400M, seeing impressive speedups for large images while hitting GPU memory and precision limits at high sizes. The post shares practical tips on casting to GPU types, minimizing CPU-GPU transfers, and when GPU acceleration is most useful.
GPGPU DSP
Shehrzad Qureshi kicks off his DSP blog by championing GPGPU, focusing on Nvidia's CUDA and real-product experience. He argues that with CPU clock speeds stalled, large-scale parallelism on GPUs is the practical path forward for many signal-processing tasks. The post traces GPGPU history from shader 'hacks' to modern APIs and previews future posts comparing CUDA vs OpenCL, Intel's Larrabee, and Nvidia Fermi.
The Nature of Circles
Averaging angles the usual way can produce nonsense: the mean of 0 and 359 degrees is not 179.5 when working with circular data. Peter Kootsookos shows the correct approach using vectorial or phasor averaging, converting angles to unit complex numbers and taking the argument of their sum. The short post points to directional statistics and a related IEEE paper for deeper details.
Music/Audio Signal Processing
Julius Orion Smith III traces his journey from musician to music/audio DSP researcher, sharing the choices that shaped his career and research focus. He recounts work on violin modeling and waveguide synthesis, then highlights modern prototyping tools like Faust and Octave that accelerate experimentation. Read for practical career advice on coursework, publishing, and why free open-source tools matter for rapid audio research.
Time Machine, Anyone?
Causal filters can look like time machines, but they do not break physics. Andor Bariska reproduces a classic electronic experiment in MATLAB, showing how a minimum-phase peaking filter and its FDLS biquad approximation produce negative group delay bands that make predictable, bandlimited signals appear to emerge early. The post walks through group delay, discretization, pulse and random-signal tests, and why unpredictability restores causality.
Benford's law solved with DSP
Steve Smith shows that standard DSP tools give a clean, intuitive explanation of Benford's law by treating leading-digit counts as signals on the number line and using convolution and Fourier analysis. He publishes the full derivation as an online chapter after traditional journals showed little interest. The result highlights how time- and spatial-domain DSP techniques can be applied to numeric distributions.
Waveforms that are their own Fourier Transform
Steve Smith admits a long-standing mistake and overturns the claim that only Gaussians are their own Fourier transform. He gives trivial and nontrivial examples, explains why infinitely many such waveforms exist, and shows a quick discrete construction using the DFT with a 1/sqrt(N) normalization. Engineers get an intuitive 30-second argument plus a practical recipe to build self-Fourier signals.
Computing Chebyshev Window Sequences
Rick Lyons gives a compact, practical recipe for building M-sample Chebyshev (Dolph) windows with user-set sidelobe levels, not just theory. The post walks through computing α and A(m), evaluating the Nth-degree Chebyshev polynomial, doing an inverse DFT, and the simple postprocessing needed to form a symmetric time-domain window. A worked 9-sample example and an implementation caveat for even-length windows make this immediately usable.
An Interesting Fourier Transform - 1/f Noise
Power-law signals have a neat Fourier trick: their transforms are power laws too, but with important caveats. Steve Smith walks through the t^α ↔ ω^{-(α+1)} relation, shows how the unit step, the Gamma scaling and a nontrivial phase change the picture, and highlights the special α = -0.5 case that links to 1/f noise. The post frames why phase and physical interpretation keep 1/f noise mysterious.
Components in Audio recognition - Part 1
This post introduces the core components of an audio recognition system, framed against how the human auditory system naturally familiarizes and retrieves tunes. Prabindh Sundareson outlines the three building blocks: an archive store, an analysis and fingerprinting engine that groups tracks, and a front-end that accepts queries and places samples into groups. He previews upcoming posts that will dig into implementations and tradeoffs.
Components in Audio recognition - Part 1
This post introduces the core components of an audio recognition system, framed against how the human auditory system naturally familiarizes and retrieves tunes. Prabindh Sundareson outlines the three building blocks: an archive store, an analysis and fingerprinting engine that groups tracks, and a front-end that accepts queries and places samples into groups. He previews upcoming posts that will dig into implementations and tradeoffs.
Accelerating Matlab DSP Code on the GPU
Seth Benton spent a few days testing Jacket to accelerate MATLAB on NVIDIA GPUs, and found it surprisingly easy to speed up DSP code. He ran 2D FFT and interp2 benchmarks on a MacBook Air with a GeForce 9400M, seeing impressive speedups for large images while hitting GPU memory and precision limits at high sizes. The post shares practical tips on casting to GPU types, minimizing CPU-GPU transfers, and when GPU acceleration is most useful.
ICASSP 2011 conference lectures online (for free)
For the first time, the oral sessions of ICASSP 2011 were recorded and posted online for free, giving engineers worldwide easy access to the conference. The talks span speech and communication signal processing, plus eclectic topics like bio-inspired methods, where Prof. Sayed uses a distributed LMS model to reproduce group predator and prey behavior. Expect some theoretical material, but many presentations are practical and inspiring for DSP practitioners.
ES Week Emphasis on Component Based Design
Howdy everyone from beautiful Salzburg/Austria,
A week full of presentations on embedded systems at ESWeek was quite a mindful. Similar to most academic conferences, there was only a few papers worth taking back home to think about. Amongst these were:
1. Keynote talk by Hermann Eul from Infineon: He presented Infineon's view on SDR and its evolution. This talk was quite inspirational. However the most interesting slide on complexity of SDR evolution was removed. I wish I could give this...
Engineering the Statistics
Do you remember the probability course you took in undergrad? If you were like me, you would consider it one of those courses that you get out of confused. But maybe a time will come where you regret skipping class because of the lecturer's persisting attempts to scare you with mathematical involved nomenclature.As you might have guessed, I had this moment few months back where I had to go deep into statistical analysis. I learned things the hard way, or maybe it is the right way. I mean...
Why is Fourier transform broken
Many engineers know the Gibbs phenomenon without grasping its root cause. This post shows that the problem comes from using the incomplete metric space of continuous functions, C[a,b], for Fourier series, and explains how switching to Lp spaces resolves convergence in the mean but allows functions to differ on sets of measure zero. It also reminds readers that Fourier analysis gives no time localization, so be mindful of its limits.
State Space Representation and the State of Engineering Thinking
Most, if not all, textbooks in signal processing (SP) thoroughly covers the frequency analysis of signals and systems alike, including the Fourier and the Z-transform that produce the well known Transfer Function. Another way of signal analysis, not as popular in signal processing though, is State Space representation. State space models describes the internal signals of the system or the process and how it affect the output, in contrast to the frequency representation that only describe the...
The correct answer to the quiz of @apolin
A compact Simulink model explains why certain DFT rows behave like negative-frequency bandpass filters, using dftmtx(8) rows as impulse responses. The demo shows that a 2 kHz tone with phase 0 or pi produces identical real parts and opposite imaginary parts, making a negative-frequency interpretation unnecessary. It also illustrates how a 6 kHz tone under 8 kHz sampling aliases to 2 kHz with opposite phase, visible in PSD plots.
The Nature of Circles
Averaging angles the usual way can produce nonsense: the mean of 0 and 359 degrees is not 179.5 when working with circular data. Peter Kootsookos shows the correct approach using vectorial or phasor averaging, converting angles to unit complex numbers and taking the argument of their sum. The short post points to directional statistics and a related IEEE paper for deeper details.
GPGPU DSP
Shehrzad Qureshi kicks off his DSP blog by championing GPGPU, focusing on Nvidia's CUDA and real-product experience. He argues that with CPU clock speeds stalled, large-scale parallelism on GPUs is the practical path forward for many signal-processing tasks. The post traces GPGPU history from shader 'hacks' to modern APIs and previews future posts comparing CUDA vs OpenCL, Intel's Larrabee, and Nvidia Fermi.
Off Topic: The True Gravitational Geodesic
The third of my off topic Physics series resulting in the true gravitational geodesic equation and some surprising results about gravity.
The Nature of Circles
Averaging angles the usual way can produce nonsense: the mean of 0 and 359 degrees is not 179.5 when working with circular data. Peter Kootsookos shows the correct approach using vectorial or phasor averaging, converting angles to unit complex numbers and taking the argument of their sum. The short post points to directional statistics and a related IEEE paper for deeper details.
Engineering the Statistics
Do you remember the probability course you took in undergrad? If you were like me, you would consider it one of those courses that you get out of confused. But maybe a time will come where you regret skipping class because of the lecturer's persisting attempts to scare you with mathematical involved nomenclature.As you might have guessed, I had this moment few months back where I had to go deep into statistical analysis. I learned things the hard way, or maybe it is the right way. I mean...
Why is Fourier transform broken
Many engineers know the Gibbs phenomenon without grasping its root cause. This post shows that the problem comes from using the incomplete metric space of continuous functions, C[a,b], for Fourier series, and explains how switching to Lp spaces resolves convergence in the mean but allows functions to differ on sets of measure zero. It also reminds readers that Fourier analysis gives no time localization, so be mindful of its limits.
State Space Representation and the State of Engineering Thinking
Most, if not all, textbooks in signal processing (SP) thoroughly covers the frequency analysis of signals and systems alike, including the Fourier and the Z-transform that produce the well known Transfer Function. Another way of signal analysis, not as popular in signal processing though, is State Space representation. State space models describes the internal signals of the system or the process and how it affect the output, in contrast to the frequency representation that only describe the...
ICASSP 2011 conference lectures online (for free)
For the first time, the oral sessions of ICASSP 2011 were recorded and posted online for free, giving engineers worldwide easy access to the conference. The talks span speech and communication signal processing, plus eclectic topics like bio-inspired methods, where Prof. Sayed uses a distributed LMS model to reproduce group predator and prey behavior. Expect some theoretical material, but many presentations are practical and inspiring for DSP practitioners.
FREE Peer-reviewed IEEE signal processing courses
The IEEE signal processing society is offereing FREE peer reviewed courses, though not many, they are peer reviewed and span differenet topics like; wavelets, speech analysis, and statistical detection.
Enjoy
http://cnx.org/lenses/ieeesps/endorsements?b_start:int=0&-C=
Components in Audio recognition - Part 1
This post introduces the core components of an audio recognition system, framed against how the human auditory system naturally familiarizes and retrieves tunes. Prabindh Sundareson outlines the three building blocks: an archive store, an analysis and fingerprinting engine that groups tracks, and a front-end that accepts queries and places samples into groups. He previews upcoming posts that will dig into implementations and tradeoffs.
GPGPU DSP
Shehrzad Qureshi kicks off his DSP blog by championing GPGPU, focusing on Nvidia's CUDA and real-product experience. He argues that with CPU clock speeds stalled, large-scale parallelism on GPUs is the practical path forward for many signal-processing tasks. The post traces GPGPU history from shader 'hacks' to modern APIs and previews future posts comparing CUDA vs OpenCL, Intel's Larrabee, and Nvidia Fermi.
Software Defined Radio at SAMOS
Lets start off with so 'hot' SDR track held at SAMOS conference this year. The academic community relatively active in the SDR architecture domain including UMich, WisMad, Linkoping, IMEC and others all presented their views on Software Defined Radio and unveiled a part of their work. We from IMEC 'finally' made our SyncPro architecture public. You can find more about our vector synchronization processor architecture from our
