SpeechPolice

Inspiration

As gamers, we love the voice- and text-chat platform Discord, but it cannot be denied that toxicity and hate speech are prevalent. Automatic moderation for text channels is commonly seen in an attempt to keep the channels civilized, but to our knowledge, no bot is able to moderate channels with voice or speech. In the modern era of online gaming and streaming, it is more important than ever to be able to create a space for all to come together without needing to fear about constantly being attacked or harassed. We want to create an environment where everyone, regardless of their race, age, gender identity, and sexual orientation to celebrate the joys gaming can bring.

What it does

SpeechPolice can be summoned into a voice channel with !join and dismissed with !leave. When active, it listens to all users and detects toxicity and hate speech using machine learning, among other techniques. Upon detecting an infraction, it issues a warning to the offender and issues a ban after repeated offense.

How we built it

Coffee and chocolate and Yerba Mate.

Challenges we ran into

We could not read, we are starving, sleep deprived, inexperienced, hysterical, cold, and stressed about midterms. And we could not find a great dataset for hate speech.

Accomplishments that I'm proud of

At the start of the hackathon, a majority of our team had never used any of the programs that we worked with today. From JavaScript and MongoDB Atlas to TensorFlows and machine learning to Git and HTML, we had to learn all of it from scratch. In fact, 2 of our members are attending their first hackathon and one of us literally started coding last month, with his only experience being self taught C++. To do so, all 3 of us pulled all-nighters, and even though it was rough at times, and we really wanted to go to sleep, we persevered and finished our projects with additional specs! The joy and relief of finally finishing the project made everything worth it.

What we learned

We learned stuff and its kinda cool, and that one should not give Frank more than two Yerba Mates to drink in one night.

What's next for SpeechPolice

We could try to make the machine learning more accurate by finding more appropriate data. We also note that some information such as tone of voice is lost when converting to text. If we had the data and resources to train our system from scratch, we could potentially improve our system.

Built With

• discord
• google-cloud
• html
• mongodb
• node.js
• tensorflow