Inspiration: We wanted to tackle a very challenging artificial intelligence concept.
How: The whole project is divided into three parts: 1) Signal processing - Extracting various features of sound like frequency domain, time domain, spectral domain, etc and segmenting them individually. 2) Machine learning - Feeding the segmented data to neural networks and creating different variations of the same original input. 3) Sound synthesis - Combining the segmented data to create new unique music.
Challenges: Signal processing of Wav files is a hard nut to crack. There have been projects made previously that compose new music, but with midi files. Wav/mp3 files are more commonly used, harder to read and there are a lot of features to extract and process. For the time being, we stuck to processing only the frequency and time domain of Wav files. Another issue was to understand what key features will manipulate to the right kind of music.
Accomplishments: Creating music that actually sounds melodious, without the team having any musical background.
What we learned: We learned how to learn and use multiple packages efficiently. In spite of the project being very complex, we made sure not to give up.
What's next: Getting deep into computer-music and creating much more complex music.