Having previously written a couple of blog posts regarding the use of LLMs to write DSP code, I've spent the last few months working on a project that has shown the landscape has changed dramatically.
The previous blog posts are here:
All things DSP, AI , ML, IoT and more ...
Having previously written a couple of blog posts regarding the use of LLMs to write DSP code, I've spent the last few months working on a project that has shown the landscape has changed dramatically.
The previous blog posts are here:
The Mel-frequency Cepstrum (MFC) and it's associated outputs, the Mel-frequency Cepstral Coefficients (MFCCs), are commonly used for speech applications such as speaker and speech recognition, using neural networks. Unfortunately, the nature of the MFC means that it is not always ideally suited to applications such as vibration analysis and predictive maintenance
The MFC uses logarithmicaly spaced frequency banks to replicate how the human ear hears sound. This approach can lead to very large savings in the number of MIPS required for the recognition part of speaker and speech recognition. Unfortunately, this logarithmic frequency space hides frequencies that are closely spaced meaning that this approach is sub-optimal for applications such as machine vibration analysis, where small variations in vibrational frequency can indicate problems with the machine, particularly the bearings.
The following diagram shows a simple Mel-spaced filterbank, with 12 separate filters:
Copyright © 2024 Delta Numerix
Back in March 2023 I wrote the following blog post about using Generative AI and Large Language Models (LLMs) to write code: Are Chat-GPT and Google Bard The New Frontier For Writing DSP Code?
Since then, I have used these tools in many projects and have made a number of observations. In general, the more complex the task you are setting for the LLM, the more likely the performance of each is going to diverge and also the more likely it is that, as a programmer, you are going to have to test the code extensively to find the bugs.
Using these tools is a bit like an artist generating a preliminary sketch, rather than the final polished painting, with all of the correct detail.
I have found three main uses for Generative AI in coding:
I have tried all of the following: Gemini, Google Code Assist, Chat-GPT, Bing and Co-Pilot. I have had the best coding results with Gemini (Bard) however if I find that this is struggling then I will try them all because they all have strengths and weaknesses.
It is important that you know what you want to do because there is no guarantee you will receive a correct answer!
A useful trick is to try the same request multiple times because, unlike a traditional search engine, an LLM with give you a different responses each time. Handily, Gemini automatically generates 3 draft solutions and you can click on the tabs provide to review each.
I have observed that LLMs are much better at writing Python than lower level languages (C/C++ etc.). In Python, it will almost certainly produce a working solution using the Numpy/Scipy library functions, that may just need some final tuning.
If you are writing code for a lower level language then the best option is often to take a two stage approach:
Generative AI is very good for converting between languages and Gemini will add comments to code that does not contain original comments. This is particularly useful if you work with a colleague who is not very dilligent with their code commenting ;-). It is worth noting, however, that comments are sometimes wrong due to AI misunderstanding the intention of the code.
Sometimes the conversion process will skip complex code sections, in a program, entirely so if this happens then the next step is to copy those sections and convert them separately.
Converting code from Python to C/C++ is generally very easy because they both use 0-based array indexing. Matlab, however, is more complex because it uses 1-based array indexing and this confuses the LLM. When converting Matlab code to Python or C/C+++ then I generally use the following request, which I then follow with the code section:
convert the following matlab code, with 1 based array indexing, to Python and Numpy, with 0 based array indexing
One final example of a gotcha is that Matlab uses FIR filter order whereas Scipy uses the number of coefficients.
As well as documenting code, LLMs are very good at debugging code however it is often important to explicitly specify the language in the request, rather than leaving it to the LLM to decide what language the code is written in.
Finally, Test! Test! Test!
Copyright © 2024 Delta Numerix
Back in the 1980s one of my first tasks, as a junior engineer, was to write a very simple filter design program, in GWBasic!
In the 1990s I updated it to Borland C and added some new functionality.
In the 2000s I added a GUI front end and more functionality.
For the last 20 years this has done myself and my customers well and has been my goto filter design tool but has languished in recent times so I've finally found the time to open-source it and it's now part of the SigLib DSP library.
https://github.com/Numerix-DSP/siglib
The 33rd annual running of the University Of Oxford Digital Signal Processing course will be running live, in Oxford, UK, from Tuesday 4th to Friday 7th June 2024.
The courses are presented by experts from industry for Engineers in industry and over the last 30 years has trained many hundreds of Engineers, from all areas of Science and Engineering.
Here is a summary of the two courses.
https://www.conted.ox.ac.uk/courses/digital-signal-processing-theory-and-application
This course provides a good understanding of DSP principles and their implementation and equips the delegate to put the ideas into practice and/or to tackle more advanced aspects of DSP. 'Hands-on' laboratory sessions are interspersed with the lectures to illustrate the taught material and allow you to pursue your own areas of interest in DSP. The hands-on sessions use specially written software running on PCs.
Subjects include:
A one-day supplement to the Digital Signal Processing course that takes the theory and translates it into practice.
The course will include a mixed lecture and demonstration format and has been written to be independent of target processor architecture.
The course will show how to take common DSP algorithms and map them onto common processor architectures. It will also give a guide line for how to choose a DSP device, in particular how to choose and use the correct data word length for any application.
Attendee Feedback From Previous Courses:
It was informative, enjoyable and stimulating
Excellent content, very lively thanks to the 2 excellent presenters - Anonymous
A very good introduction to DSP theory
Excellent lecturers! Really useful information and very understandable
Great mix of theory and practice
The lecturers gave a detailed and excellent explanation of the fundamental topics of DSP with real world engineering practice.
This session closes the gap and clears up much confusion between classroom DSP theories and actual DSP implementation.
Very good session, with in-depth discussion on the math and background.
These courses will be held at the University of Oxford, UK
Copyright © 2024 Delta Numerix
When performing frequency domain (FFT) based processing it is often useful to display a spectrogram of the frequency domain results. While there is a very good SciPy spectrogram function, this takes time domain data and does all of the clever stuff. However if you are processing data in the frequency domain you often just want to build the spectrogram dataset and keep appending FFT results to it.
A spectrogram is a 3D plot, with the following configuration:
This program will use the Matplotlib function imshow() to display the spectrogram.
There are two key tricks to using imshow() for this purpose:
import matplotlib.pyplot as pltimport numpy as npfrom scipy import signal# Global ConfigurationplotXAxisSecondsFlag = True # Set to True to Plot time in seconds, False to plot time in samplesplotYAxisHzFlag = True # Set to True to Plot frequency in Hz, False to plot frequency in binsFs = 10000 # Sampling Frequency (Hz)timePeriod = 10 # Time period in secondssampleLength = Fs*timePeriodsinusoidFrequency = 1000 # Frequency of sine wave (Hz)fftLength = 256 # Length of the FFThalfFftLength = fftLength >> 1window = np.hanning(fftLength)time = np.arange(sampleLength) / float(Fs) # Generate sinusoid + harmonic with half the magnitudex = np.sin(2*np.pi*sinusoidFrequency*time) * ((time[-1] - time)/1000) # Decreases in amplitude over timex += 0.5 * np.sin(2*2*np.pi*sinusoidFrequency*time) * (time/1000) # Increases in amplitude over time# Add FFT frames to the spectrogram list - Note, we use a Python list here becasue it is very easy to append tospectrogramDataset = []i = 0while i < (len(x) - fftLength): # Step through whole datasetx_discrete = x[i:i + fftLength] # Extract time domain framex_discrete = x_discrete * window # Apply window functionx_frequency = np.abs(np.fft.fft(x_discrete)) # Perform FFTx_frequency = x_frequency[:halfFftLength] # Remove the redundant second half of the FFT resultspectrogramDataset.append(x_frequency) # Append frequency response to spectrogram dataseti = i + fftLength# Plot the spectrogramspectrogramDataset = np.asarray(spectrogramDataset) # Convert to Numpy array then rotate and flip the datasetspectrogramDataset = np.rot90(spectrogramDataset)z_min = np.min(spectrogramDataset)z_max = np.max(spectrogramDataset)plt.figure()plt.imshow(spectrogramDataset, cmap='gnuplot2', vmin = z_min, vmax = z_max, interpolation='nearest', aspect='auto')plt.title('Spectrogram')freqbins, timebins = np.shape(spectrogramDataset)xlocs = np.float32(np.linspace(0, timebins-1, 8))if plotXAxisSecondsFlag == True:plt.xticks(xlocs, ["%.02f" % (i*spectrogramDataset.shape[1]*fftLength/(timebins*Fs)) for i in xlocs]) # X axis is time (seconds)plt.xlabel('Time (s)')else:plt.xticks(xlocs, ["%.02f" % (i*spectrogramDataset.shape[1]*fftLength/timebins) for i in xlocs]) # X axis is samplesplt.xlabel('Time (Samples)')ylocs = np.int16(np.round(np.linspace(0, freqbins-1, 11)))if (plotYAxisHzFlag == True):plt.yticks(ylocs, ["%.02f" % (((halfFftLength-i-1)*Fs)/fftLength) for i in ylocs]) # Y axis is Hzplt.ylabel('Frequency (Hz)')else:plt.yticks(ylocs, ["%d" % int(halfFftLength-i-1) for i in ylocs]) # Y axis is Binsplt.ylabel('Frequency (Bins)')plt.show()
Copyright © 2024 Delta Numerix
The course first moved online in 2020 and has received excellent reviews from the attendees.
The course will run from Monday 8th April to Friday 17th May 2024, with live online classes one afternoon per week.
Based on the classroom course, Digital Signal Processing (Theory and Application), this online course consists of weekly live online tutorials and also includes a software lab that can be run remotely. We'll include all the same material, many of the existing labs and all the interaction of the regular course.
Online tutorials are delivered via Microsoft Teams once each week and practical exercises are set to allow you to practice the theory during the week.
You will also have access to the course VLE (virtual learning environment) to communicate with other students, view and download course materials and tutor support is available throughout.
Code examples will be provided although no specific coding experience is required.
The live tutorials will be on Wednesday each week from 13:00 - 14:30 and 15:00 - 16:30 (GMT) with a 30-minute break in between.
You should allow for 10 - 15 hours study time per week in addition to the weekly lessons and tutorials.
After completing the course, you should be able to understand the workings of the algorithms we explore in the course and how they can solve specific signal processing problems.
Full details are available here: https://www.conted.ox.ac.uk/courses/digital-signal-processing-online.
Copyright © 2024 Delta Numerix