What it does
Our machine learning model is built into a simple website. Anyone can go to the website to record or upload an audio file. The audio file is transcribed to text and it is run through a model trained on thousands of Wikipedia articles to the topic of conversation. The text is evaluated to determine the topics and label areas of interest. We will soon expand to use a more complete dataset with more topics that have been robustly trained.
How we built it
The core of our model is built in Python using Flask. We used the Google Speech to Text API to transcribe the data and the NLTK library for the natural language processing functionality to analyze the text. We trained a Latent Dirichlet Allocation model on the English Wikipedia Articles 2017-08-20 Models dataset for classification. The model was trained using gensim. Our website is built in React and have integrated the machine learning model into an API to connect it to the website.
Challenges we ran into
We ran into some difficulties in training our model initially, as our dataset was over 17GB and therefore would take days to process. By using a smaller dataset and the accuracy was lower than we expected and with the time constraints we weren't able to take as many passes to train the model as we had hoped.
Accomplishments that we're proud of
We are proud to have figured out how to successfully train a model despite no team members having prior experience in machine learning.
What we learned
We greatly expanded our knowledge of machine learning, natural language processing, and web development.
What's next for discussion
Our number one goal is to provide more accuracy in determining topics. We want to develop a more robust model with the larger dataset to improve our accuracy.
More about our team:
Matthew Chiang http://matthew-chiang.com/