Inspiration
"Je lui parlais de bridge, de golf, de politique et de cravates. Et la grande personne était bien contente de connaître un homme aussi raisonnable." - Le Petit Prince
"I would bring myself down to his level. I would talk to him about bridge, and golf, and politics, and neckties. And the grown-up would be greatly pleased to have met such a sensible man." - The Little Prince
Pitch
Talking to people to network and develop your personal brand is hard. Whether it is at a tech function, forced corporate outing, All adults want to talk about are sports, mortgages and the internal politics of countries they will never set foot in themself, So we made an application that takes the pain out of soPhisticAtIoN, that will listen in on conversations, pick out keywords, and suggest trivia to allow you to speak on the level of adults.
How we built it
We get a "conversation" block of roughly 15 seconds --> ~25 words
E.g.
A: I heard the JWST recently found an exoplanets, that's cool.
B: You're right, it is cool. Very cool :).
A: Imagine if it finds life!
B: I want aliens.
This is encoded into a corpus (via the Google's Speech Recognition) as the following:
I heard the JWST recently found an exoplanets, that's cool. You're right, it is cool. Very cool :). Imagine if it finds life! I want aliens.
We now utilise keyword extraction algorithms (utilising an open-source library known as KeyBert). The algorithm outputs $n$ keywords, with their respective probabilities, $p_i$. For now, under the pre-testing process, we use $n = 5$, as this is a good standard for our project.
Once we have the 5 keywords, we perform a very simple Google Search, and take the top few results. Scraping these websites, we get a good corpus of data that is useful for our codebase. We then use an algorithm to identify texts relating to the topics (keywords). This text is summarised and outputted in the form of a wav audio with Google's TTS.
This audio is played, and that's it.
Challenges we ran into
CS2100 Make-up Exam
Ziming was unable to train the BERT model overnight and was very sad in the morning.
Accomplishments that we're proud of
We actually made a good summarization algorithm.
What we learned
We learnt more about NLP libraries such as Keras NLP and HuggingFace.
Built With
- html
- javascript
- python
- stt
- summarize
- tailwind
- tts
- web-scraping
Log in or sign up for Devpost to join the conversation.