Neighbor weight distance map of the first channel of the audio signals , with 8x8 SOM
Neighbor weight distance map of the first MFCC of the first channel of the audio signal , with 20x20 SOM
Weight positions map of the first MFCC of the first channel audio signals , with 20x20 SOM
Scheme of a 6x6 SOM
Hits map of the first MFCC of the second channel audio signals , with 6x6 SOM
Neighbor weight distance map of the second channel of the audio signal , with 20x20 SOM
Tired of so superficial apps such as based just on physical appearance, we want to try something that really connects people with each other based on their interests and personalities. While you are driving, cooking, working you can just record a voice message and send it to the person that might be your soul mate!
What it does
Basically, daily you can record a short voice message telling a funny, crazy, interesting story about your day available on the network for 24 hours. If someone is attracted by your voice or your story, he or she will just say a YES. And, if you say a YES back... hey you got a match! Now you can reply with a new voice message and start a chat. Every YES you say will train a SOM neural network that will help you to get messages from people with voices you'll probably like. The most you use the app, the highest is the chance you will get in touch with sexy voices you like! Tired of writing and making typo errors, you will just need to talk to the nice Alexa.
How I built it
The Alexa App is built with the Alexa Skills Kit and the business logic ( built with Node JS, because Node JS s awesome!) is hosted by the awesome guys from Swisscom - Thank You for that .
A Self Organizing Map is trained with pre-processed signals of Swiss German voices messages. The clustering allows to identify the similarities between different voices and asses if a new voice sample could be appreciate by the user.
Challenges I ran into
Dataset was too small to obtain a network with good generalization capabilities. A backend developer might have given us some more hours of sleep.
Accomplishments that I'm proud of
Four hours of sleep in three days and still working brain (maybe!). Using our passion about machine learning to make people fall in love (and not in superficial way!) is great :)
What I learned
The Alexa environment is awesome and a ton of fun to hack with! Hackathons are amazing opportunities to meet lots of smart and crazy people with amazing ideas and no fear to realize them.
What's next for LoveVibes
From a business perspective LoveVibes is a great opportunity to collect trainig data in swiss german in a rather high quality. With the help of manual annotaitions you can train some neural networks to translate spoken swiss german in written high german.