According to the AAA Foundation for Traffic Safety, ONE in FIVE accidents are caused due to drowsy and drunk driving. This is a preventable problem, and we wanted to look for a better way to tackle drowsy driving.
What it does
A driver, upon entering her/his car, is asked to speak out a few sentences to the Raspberry Pi mounted on the dashboard. The device (with the support from an Arduino) analyzes the input voice, and determines if the driver is fit to drive. If she/he is not fit to drive, the tool dispatches text messages to close family members/friends alerting them of the driver's location, following which they may take action.
How we built it
Our tool uses the Speech-To-Text IBM Watson Service to generate a JSON representation of the input speaker's voice. We then parse the JSON in Python and analyze the following things:
a) Time differences between words spoken, and calculate the average of these time differences. If the average is larger than a threshold value, we trigger a flag.
b) Analyze the time taken to enunciate long words (we use the example word 'encyclopedia') and compare it to the average expected enunciation time.
c) Analyze the time taken to enunciate certain phonetics (we use the phonetic sound 'ch' in 'chose') based on scientific evidence that certain phonetic sounds take longer time to enunciate when intoxicated.
d) Analyze the rate of enunciation of the 26 letters of the English Alphabet. The tool returns a warning flag if average time between characters, and the average time enunciating each character is above a specified threshold value.
All these four tests are based on scientific research conducted on this topic. Based on the collective weight of the four tests, we provide a confidence value of the level of drunkenness on a scale of 0 to 1.
Challenges we ran into
It was difficult to pinpoint the specific nuances in speech we had to look out for. It took us a good deal of reading to come up with a (non exhaustive, as of now) list of aberrations in speech patterns.
Accomplishments that we are proud of
This is our first hackathon, and we're very excited to turn in a project for the expo. We were very happy to see that our tool generates almost-accurate responses for the drunkenness tests!
What we learned
We learnt a lot about designing IoT based technologies in general, and also the ability to easily use the Watson API to our advantage. We also learnt that working with Edison boards suck! (Unless you're lucky and you get a non-faulty board!)
What's next for Drink And D-Rive
We will definitely attempt to complete the entire hardware aspect of the project - as well as add additional features to the detection algorithm, such as reaction-time analysis, tone analysis and possibly some visual detection as well. The reason we did not want to initially incorporate visual cues to the tool was in order to keep the cost of the device relatively affordable, and as non-intrusive as possible.