Inspiration


We were looking to build out additional features for existing speech recognition and conversational agents. General-purpose chatbots are few and far between, whereas closed-domain bots lack the ability to maintain a conversation any more meaningful than an automated FAQ.

What it does

Bots know what we mean based on the text we type. However, humans get and give a lot of signals from intonation; that would essentially go by the wayside with chatbots. This would aim to improve the user experience by taking that data into account.

How we built it

We used a 1D convolutional neural network with a single dense layer to predict probabilities for a given set of emotions (8 per sex). We used Flask for the Model-as-a-Service REST API. Additionally, electron was used for the frontend.

Challenges we ran into

  • Training resources were not available, leading us to lose a ton of time. GPU quotas were not able to be used from the GCP credits, rendering them useless. To combat that, we had to use a manual solution on colab, and port it by hand to a compute engine.
  • Issues gathering data.
  • Troubles generalizing within the time available.

Accomplishments that we're proud of

  • ML in 24 hours.
  • A nice frontend in a library we didn't have too much experience with.
  • A working demo.

What we learned

What's next for S4T - Someone To Talk To

Built With

Share this project:

Updates