CSI Adelaide: Was Trayvon Martin murdered?
On February 26th, 2012, Trayvon Martin was shot dead in Florida, USA. The shooter, George Zimmerman, said it was in self-defence. However, the case has taken the USA by storm as the public is divided. Many believe Martin was innocent and unfairly shot by Zimmerman. But also many believe the opposite that Zimmerman was innocent and acted in self-defence. You can find the full story here.
A member of the public who heard the two struggle, before the shooting, called the police and we have the recording of this call. Interestingly in the background you can hear screaming.
The gunman claims that was his voice on the recording screaming, demonstrating that Martin attacked him and he shot Martin in self-defence. Those who believe Martin was innocent say the screams were actually Martin's, thus demonstrating that Martin was the victim and was shot in cold blood.
The screams are very high pitched and do appear to be that of a younger person, which would suggest Martin. However, how can we be certain? As electronic engineers we can use signal processing and forensic engineering techniques to be more certain.
The problem is no one has a benchmark recording of what Trayvon's voice is like. We only have recordings of Zimmerman's voice and the recording of the screaming of an unknown person.
Can you use audio processing and signal classification techniques to shed light on this case? Go ahead and record on your iPhone the normal talking voices of 10 different male friends. Then get them to scream and record those individually. You can also try to record about 10 different male screams from actors in horror movies and also their talking voices. Then you can sample the sound files, extract features, and put those features in a classifier. Can your classifier match the 20 screams correctly to your 20 known talking voices, to better than 90% accuracy? If you can achieve that then you can try out your classifier on Zimmerman's voice and the unknown screamer. Can you come to a conclusion on whether Zimmerman or Martin screamed? To what level of certainty can you assign to your conclusion?
You are potentially helping to shed light on a real case, so your results need to be presented with utmost care and integrity.
Approach and methodology
The sound recordings are .oog files and you may need to convert them to .wav files. You'll then need to read the .wav files into GNU Octave. You'll need to do a bit of reading on voice recognition and decide on some features to extract from the signals. You may then want to experiment with one or two types of classifiers: say, a Support Vector Machine (SVM) and Multiple Discriminant Analysis (MDA). You should be able to find pre-existing MATLAB code for these types of classifiers.
In case you are geniuses and get the job done too quickly, there are a number of extra aspects you can look at: (a) In reality, the recording of the scream is done on a different phone to the recording of Zimmerman's voice. So the next step would be to see how your classification accuracy changes if you use two different recording devices: one for screaming and one for talking. (b) If you use the MDA classifier you can look at a technique called fuzzy c-means as a way of assigning a likelihood measure as to whose voice the scream belongs to. (c) You could finesse the study further and use that fact that there was a large age difference between Martin (17yrs) and Zimmerman (29yrs). So you could record screams of two groups of males of about those ages to simply see if your classifier can distinguish between the two age groups. As a 17yr old voice hasn't fully matured, you might expect that your classifier will pick up an age-dependent difference.
- We don't really expect you determine if Martin was murdered or not. That is for a jury to decide. The idea of this project is for you to simply place a quantitative probability or likelihood measure on whether the screams came from Zimmerman or not.
- To get good marks we expect you to show a logical approach. Explore the problem and see if you can get reasonable classification accuracies with your own test recordings of people screaming.
- We expect all the written work to be place on this wiki. No paper reports are to be handed up. Just hand up a CD with your complete project directory at the end. The entire project directories of each group member can go in separate folders on the same CD or memory stick. You won't be getting the CD or stick back.
- It is expected that you fill out a short progress report on the the wiki each week, every Friday evening, to briefly state what you did that week and what the goals are for the following week.
- It is important to regularly see your main supervisors. Don't let more than 2 week go by without them seeing your face briefly.
- You should be making at least one formal progress meeting with supervisors per month. It does not strictly have to be exactly a month, but roughly each month you should be in a position to show some progress and have some problems and difficulties to discuss.
- The onus is on you to drive the meetings, make the appointments and set them up.
- You are expected to make a YouTube presentation of your whole project.
Relationship to possible career path
This project will familiarize you with techniques in audio, signal classification, and digital forensics. It will also improve your software skills. You will learn about different types of classifiers, which are powerful tools for engineers. Just about every type of electronic engineering job out there these days will have a use for signal classification.
References and useful resources
If you find any useful external links, list them here: