Inspiration
We were inspired by 2010 etsch era websites that functioned off of one neat ability that it had. In this case it is a website that allows you to see whether or not your voice actually matches your DNA (physical appearance) which is pretty neat.
What it does
This website is a website that allows you to see whether or not your voice actually matches how you look. It comes with a live transcription and an audio file to play back how you sound, and obviously it also comes with the prediction.
How we built it
We trained python code off of a previously existing database: https://www.kaggle.com/datasets/rtatman/speech-accent-archive and had it determine whether the voices the frontend sent it matched a particular accent or age range.
Challenges we ran into
The main challenge we ran into was training the code off the database, when we first did it the accuracy we were getting was .09% which is obviously bad and the it took us hours to train it once. Once we optimized it we were able to get upwards of 44% accuracy and it was able to train in about 30 minutes.
Accomplishments that we're proud of
The best accomplishment that we are proud of us actually being able to train off of a dataset. We've never actually taken any machine learning or data science class which forced us to learn a lot about how the process actually worked.
What's next for VocalDNA
The obvious next step would be finding a proper dataset to train with. The current one we are using has some issues with its formatting, affecting the confidence of our age prediction the most.
Built With
- css3
- html5
- javascript
- librosa
- numpy
- python
- sklearn
- xgboost
Log in or sign up for Devpost to join the conversation.