Inspiration Switzerland is a very diverse country which is also noticeable regarding speech. Every canton is proud of his own dialect. However, collecting data for speech to text recognition is very difficult. This is where our application kicks in.
What it does The application is a matching system for flatmates. The already existing shared flats can sign up and publish relevant data about them. Everyone looking for a flat on the other hand can swipe through the list of available flats in a tinder like style. To match with a shared flat, the searching person has to speak out loud a text which is provided by the system. The flat owner can listen to these texts to decide if this is their next flatmate or not. Using this approach, we can collect annotated data throughout Switzerland.
How we built it As a backend system, a mongo-db hosted on Swisscom Cloud is used to store all relevant information. Using self setup a REST-API (implemented using Flask), the data can be accessed by our Android app by exchanging JSON data. User data is collected from Facebook and automatically built into the flat advertisements.
Challenges we ran into At the end, the application was much more complex than we thought. So for example, the amount of REST interfaces is quite high compared to what we did on other hackathons. Furthermore, using the Facebook API and Swisscom Cloud was not straight forward and therefore time consuming
Accomplishments that we're proud of We are very proud to present a fully working application allowing to match users in a Tinder like style. Despite the time constraints and the complexity of the system, we were able to finish the project on time and even style the Android app such it looks nice. The collected speech annotated data can be used to allow Swiss natives to talk to their electronic devices.
What we learned Using unknown APIs always takes more time than one thinks. Using them is often not straight forward and not as one wishes. There is often outdated information available and not suitable to our use-case. Furthermore, a small system can become complex quickly.
What's next for Zimmer Zwitscher Currently we are not sure if we will refine the idea. Therefore the code is not publicly available. However, the idea of collecting data for machine learning in a playful way is something we think is worth to dig deeper.