TrueSound

My first slide of my presentation

Inspiration:

Last year, I created a prototype of a Portable Warning System, which is a portable device that detects if cars are coming toward pedestrians to stop pedestrian-car crashes and saves their lives. The program used TensorFlow and mainly focused on video detection and neglecting the audio detection. I became interested in the audio side and that was the reason I created the deep learning program. Additionally, my mom was visiting India this week and she told me constantly about the noise pollution there. This prompted the noise reduction part of the program.

What It Does:

The program TrueSound consists of two parts: one part is a deep-learning based program that takes in the UrbanSound8K dataset and produces a machine learning model that accurately predicts which one of the ten categories test sounds fall into (96% accuracy on train data, 73% accuracy on test data, 8730 data points). The second part is an algorithm that takes in a .wav noisy file and takes out the noise from the file to produce a clearer voice.

How I Built It:

I used TensorFlow (tf.keras) and Google Colaboratory with additional modules such as librosa, IPython, pandas, numpy, and scipy as the main backend. I used the public dataset UrbanSound8K to train the model (80% train, 20% test)

Challenges:

Originally, I wanted to expand the algorithm for background noise reduction and ambiance removal into one that almost eliminates background noise, not just white noise. However, after tinkering with it for a couple of hours and building a new program for scratch, I wasn't going anywhere with the program, and thus I scratched that idea. Additionally, I did a lot of research into modules I have never used before such as librosa, IPython, and scipy. Librosa especially demanded an understanding of linear algebra and matrices that I did not have at the start. I quickly picked up what was necessary for the program that I was building and persevered through the setbacks.

Accomplishments that I'm proud of:

Learning what was necessary for the programs as waveforms were complicated to me at first
Comprehending what I learned and turning that knowledge into a cohesive program
Building a successful white-noise-removal program and a deep-learning program

What I learned:

Don't overspeculate what you can accomplish in 24 hours: this is only my third hackathon :)
Do a project that you like (like I did this time) and even though you don't know the subject at first, you can persevere and create the best program that you can